Distance measurement calibration method and system based on dTOF lidar

Abstract

The embodiment of the present disclosure provides a ranging calibration method and system based on dTOF laser radar, which comprises the following steps: using an optical beam splitter to divide laser pulses into a reference light path and a measurement light path, receiving signals by two detectors respectively and recording time points, calculating a time difference, and counting zero-meter benchmark data; combining time-to-digital converter quantization characteristics to obtain a system benchmark error calibration parameter set; collecting multi-condition calibration data based on the parameter set, constructing a multi-dimensional error compensation mathematical model, constructing a base function combination through discretization and solving through an optimization algorithm to obtain an error compensation mapping function; obtaining laser flight time and environmental parameters in real time, calculating a compensation value and performing filtering processing to obtain a calibrated ranging result; converting the result into a three-dimensional point cloud, extracting a feature point, registering and optimizing system external parameters, establishing a drift monitoring index, and automatically updating calibration parameters when the index exceeds a standard. The embodiment improves the ranging accuracy and reliability of the dTOF laser radar in a complex environment.

Description

Technical Field

[0001] This disclosure relates to the field of laser ranging technology, and in particular to a ranging calibration method and system based on dTOF lidar. Background Technology

[0002] LiDAR ranging technology has been widely used in fields such as autonomous driving, robot navigation, and industrial automation, serving as a core sensing technology for environmental perception. Among LiDAR technologies, direct time-of-flight (dTOF) technology has attracted much attention due to its long-range measurement capabilities, strong resistance to ambient light interference, and fast response speed.

[0003] Traditional lidar ranging technologies mainly include triangulation and time-of-flight methods. Triangulation involves illuminating the target with light and calculating the distance using geometric trigonometric relationships, but its measurement range is usually limited. Indirect time-of-flight (iTOF) technology determines the distance by modulating continuous light waves and measuring the phase difference, but its accuracy decreases significantly in long-distance applications.

[0004] Current dTOF lidar systems calculate target distance by measuring the time difference between laser pulse emission and reception. This technology uses a single-photon avalanche diode (SPAD) as the receiver, coupled with a time-to-digital converter (TDC) for high-precision time measurement, theoretically achieving centimeter-level ranging accuracy. However, in practical applications, the system suffers from various error sources, including response delays in the laser drive circuit, SPAD response time deviations, TDC quantization errors, temperature drift, and optical system distortion. These errors accumulate and lead to significant ranging deviations. Existing dTOF lidar calibration methods primarily rely on external calibration equipment and fixed calibration procedures, which cannot cope with system drift in dynamic environments. Furthermore, they lack compensation mechanisms for special scenarios such as multipath effects and differences in reflectivity of different materials, resulting in unstable ranging results in complex environments. In addition, the calibration process typically requires manual intervention, preventing adaptive calibration and limiting the performance of dTOF lidar in high-dynamic, high-reliability applications. Summary of the Invention

[0005] The purpose of this disclosure is to provide a ranging calibration method and system based on dTOF lidar.

[0006] To achieve the above objectives, embodiments of this disclosure provide a ranging calibration method based on dTOF lidar, comprising: The laser pulse signal emitted by the laser emitter is acquired, and the laser pulse signal is divided into a reference optical path and a measurement optical path by an optical beam splitter. The laser pulse signal of the reference optical path is received by a reference detector and a reference time point is recorded. The laser pulse signal reflected by the target in the measurement optical path is received by a main detector and the reception time point is recorded. The time difference between the reference time point and the reception time point is calculated, and the zero-meter reference measurement data is statistically analyzed. Combined with the quantization characteristics analysis of the time-to-digital converter, the system reference error calibration parameter set is obtained. Based on the system's reference error calibration parameter set, calibration data is collected under different measurement conditions to establish a multi-dimensional error compensation mathematical model that includes a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is discretized to construct a combination of basis functions and solved by an optimization algorithm to obtain the error compensation mapping function. The time-of-flight data and environmental parameters of the real-time laser pulse are acquired. The environmental parameters are input into the error compensation mapping function to calculate the error compensation value. The error compensation value is applied to the time-of-flight data and processed by a filter to obtain the calibrated measurement result. The calibrated measurement results are converted into three-dimensional point cloud data. Geometric feature points are extracted from the three-dimensional point cloud data for point cloud registration and system extrinsic parameters are optimized. System drift monitoring indicators are established. When the system drift monitoring indicators exceed a preset threshold, the calibration parameters are automatically updated to obtain a self-optimized system parameter set.

[0007] Further, the laser pulse signal is divided into a reference optical path and a measurement optical path by an optical beam splitter. A reference detector receives the laser pulse signal from the reference optical path and records the reference time point. A main detector receives the laser pulse signal reflected from the target by the measurement optical path and records the reception time point. The time difference between the reference time point and the reception time point is calculated, and zero-meter reference measurement data is statistically analyzed. Combined with the quantization characteristics analysis of the time-to-digital converter, a system reference error calibration parameter set is obtained, including: The laser pulse signal emitted by the laser emitter is divided into a reference optical path and a measurement optical path by an optical beam splitter, and a zero-meter reference optical path system is established to obtain the zero-meter reference optical path configuration. Based on the zero-meter reference optical path configuration, after receiving the laser pulse signal of the reference optical path through the reference detector, a time counter is triggered to record the reference time point. After receiving the laser pulse signal returned by the measurement optical path through the target through the main detector, a time counter is triggered to record the receiving time point. The time difference between the reference time point and the receiving time point is calculated to eliminate the influence of the inherent delay of the laser driving circuit and obtain the time difference after eliminating the driving circuit delay. Based on the time difference after eliminating the driving circuit delay, the zero-meter reference measurement data is statistically analyzed to construct a reference detector response time histogram. The mean and standard deviation of the reference detector response time histogram are analyzed to obtain the response time error model, quantify the receiver circuit delay, and obtain the receiver circuit delay parameters. Based on the delay parameters of the receiving end circuit, the quantization characteristics of the time-to-digital converter are analyzed. The quantization error is controlled within a preset range through internal oscillator frequency calibration and interpolation algorithm to obtain quantization error compensation parameters. Based on the zero-meter reference measurement data, the receiver circuit delay parameters, and the quantization error compensation parameters, a comprehensive system error calibration parameter set is calculated to obtain the system reference error calibration parameter set.

[0008] Furthermore, based on the system's reference error calibration parameter set, a multi-dimensional error compensation mathematical model is established by collecting calibration data under different measurement conditions, including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is then discretized to construct a combination of basis functions, which are solved using an optimization algorithm to obtain the error compensation mapping function, including: Based on the system reference error calibration parameter set, zero-meter reference calibration data and standard reflector ranging data were collected under different temperature conditions and different reflectivity conditions. Combining the quantization error compensation parameters and receiver circuit delay parameters in the system reference error calibration parameter set, a functional relationship model between temperature and time delay and a mapping relationship between reflection intensity and time offset were established to obtain temperature-time delay model parameters and reflectivity-time offset model parameters. Based on the temperature-time delay model parameters and the reflectivity-time offset model parameters, the time histogram characteristics of the received signal are analyzed and the multi-peak distribution is identified. A peak separation algorithm is designed to distinguish between direct reflection signals and secondary reflection signals. The background count rate of the detector is monitored and an adaptive threshold algorithm is designed to maintain the signal-to-noise ratio within a preset signal-to-noise ratio threshold range. Thus, the multipath effect detection model parameters and the ambient light compensation model parameters are obtained. Based on the temperature-time delay model parameters, the reflectivity-time offset model parameters, the multipath effect detection model parameters, and the ambient light compensation model parameters, the temperature-time delay model, reflectivity-time offset model, multipath effect detection model, and ambient light compensation model are integrated into a multi-dimensional error compensation mathematical model. A ranging error compensation function including temperature, reflection intensity, target distance, and background light intensity is constructed to obtain a multi-dimensional calibration mathematical model. Based on the multidimensional calibration mathematical model, the function space of the multidimensional calibration mathematical model is adaptively partitioned. Different types of basis functions are constructed to form a heterogeneous basis function set for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters. An adaptive mesh refinement algorithm is designed based on the heterogeneous basis function set, and the optimized mesh distribution is obtained through iterative optimization. Based on the optimized mesh distribution, the conjugate gradient method and multigrid technique are used to solve the linear equation system generated by the discretization of the segmented polynomial to obtain the error compensation mapping function.

[0009] Furthermore, the construction of different types of basis functions to form a heterogeneous basis function set for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters includes: Based on the temperature-time delay model parameters, the physical characteristics of the temperature error are analyzed, and second- to third-order polynomial basis functions are used to characterize the smooth change relationship between temperature and time delay, thus obtaining the low-order polynomial basis function of the temperature error. Based on the reflectivity-time offset model parameters and the low-order polynomial basis function of the temperature error, the piecewise characteristics of the reflectivity error are analyzed. A piecewise continuous smooth function is constructed in different reflectivity intervals using cubic spline basis functions, while ensuring the continuity of the first and second derivatives, to obtain the piecewise spline basis function of the reflectivity error. Based on the parameters of the multipath effect detection model, the low-order polynomial basis function of the temperature error, and the piecewise spline basis function of the reflectivity error, the local characteristics of the distance error are analyzed, and the Gaussian radial basis function is used to characterize the local response characteristics of the distance-related error, thus obtaining the Gaussian radial basis function of the distance error. Based on the low-order polynomial basis function of temperature error, the piecewise spline basis function of reflectivity error, and the Gaussian radial basis function of distance error, a composite basis function is constructed and the weighting coefficients are calculated to obtain the heterogeneous basis function set.

[0010] Further, the process of acquiring real-time laser pulse time-of-flight data and environmental parameters, inputting the environmental parameters into the error compensation mapping function to calculate the error compensation value, applying the error compensation value to the time-of-flight data and processing it through a filter to obtain the calibrated measurement result includes: The emission and reception times of each laser pulse are obtained, the original flight time is calculated, and the received light intensity, current ambient temperature and background light intensity are recorded to obtain the original measurement data set. Based on the original measurement data set, a real-time time histogram is constructed from the flight time data of one hundred to one thousand consecutive measurements in the original measurement data set. The main peak position and distribution width of the real-time time histogram are analyzed to obtain the statistical characteristics of the time histogram. Based on the time histogram statistical characteristics, the received light intensity, current ambient temperature, and background light intensity from the original measurement data set, along with the time histogram statistical characteristics, are input into the error compensation mapping function to calculate the corresponding error compensation value, thereby obtaining the real-time error compensation parameters. Based on the real-time error compensation parameters, the original flight time in the original measurement data set is corrected by applying the real-time error compensation parameters. At the same time, an adaptive Kalman filter is designed based on the statistical characteristics of the time histogram to smooth the measurement results after error correction, reducing random fluctuations while retaining rapidly changing effective information, and obtaining filtered measurement data. Based on the filtered measurement data, the calibrated flight time and distance values ​​are calculated, and the confidence score is calculated based on the statistical characteristics of the time histogram to obtain the calibrated measurement results and the confidence score.

[0011] Furthermore, a real-time time histogram is constructed from the flight time data of one hundred to one thousand consecutive measurements in the original measurement dataset. The main peak position and distribution width of the real-time time histogram are analyzed to obtain the statistical characteristics of the time histogram, including: Obtain flight time data from one hundred to one thousand consecutive laser pulse measurements in the original measurement data set, set the time resolution as the time interval width of the histogram, and obtain the time histogram construction parameters. Based on the time histogram construction parameters, the flight time data is divided into corresponding time intervals, the count value of each time interval is counted, a real-time time histogram is constructed, and an initial time histogram is obtained. Based on the initial time histogram, the position of the main peak in the initial time histogram is identified by a peak detection algorithm, the count value corresponding to the main peak position is calculated, the main peak position is determined, and the main peak position parameter is obtained. Based on the main peak position parameter, locate the data distribution area around the main peak position in the initial time histogram, analyze the peak shape characteristics around the main peak position by the full width at half maximum (FWHM) method or standard deviation calculation, calculate the distribution width around the main peak, and obtain the distribution width parameter. Based on the main peak position parameter and the distribution width parameter, the presence of a multi-peak distribution is identified in the initial time histogram. The signal-to-noise ratio is evaluated based on the count value corresponding to the main peak position parameter and the background noise count value of the initial time histogram to determine the measurement quality and obtain the statistical characteristics of the time histogram.

[0012] Further, the calibrated measurement results are converted into three-dimensional point cloud data, geometric feature points are extracted from the three-dimensional point cloud data for point cloud registration, and system extrinsic parameters are optimized. A system drift monitoring index is established, and the calibration parameters are automatically updated when the system drift monitoring index exceeds a preset threshold, resulting in a self-optimized system parameter set, including: Based on the calibrated measurement results and the confidence score, the data is converted into three-dimensional point cloud data. Geometric feature points of corners, edges and planes are extracted from the three-dimensional point cloud data. A feature descriptor based on local geometric descriptors is established to obtain a set of feature points and the feature descriptor. Based on the feature point set and the feature descriptor, the iterative nearest point algorithm is used to register and match the three-dimensional point cloud data collected at different times, calculate the rotation matrix and translation vector, and obtain the point cloud registration result. Based on the point cloud registration results, the system extrinsic parameters are optimized by maximizing the mutual information between the three-dimensional point cloud data using continuously acquired three-dimensional point cloud data. This enables system self-calibration without the need for a dedicated calibration target, resulting in optimized system extrinsic parameters. Based on the optimized system extrinsic parameters and historical measurement data, a system drift monitoring index system is established, including feature point position offset, point cloud registration accuracy, and measurement consistency. The system status is evaluated periodically, and a recalibration process is triggered when key indicators exceed preset thresholds to obtain system drift monitoring results. The key indicators include feature point position offset, point cloud registration accuracy, and measurement consistency. Based on the system drift monitoring results, a hierarchical recalibration strategy is designed, from parameter adjustment to comprehensive system calibration. The appropriate calibration level is automatically selected according to the degree of drift, the system calibration parameter library is updated, a parameter self-optimization mechanism is established to predict the parameter change trend, and the self-optimized system parameter set is obtained.

[0013] Furthermore, before acquiring the real-time laser pulse time-of-flight data and environmental parameters, the method further includes near-range blind zone elimination and photon accumulation effect suppression steps, including: A preliminary estimate of the target distance is obtained. Based on detector gating technology, the opening time and duration of the detector's receiving window are dynamically adjusted according to the preliminary estimate. The laser pulse width is simultaneously controlled to be variably adjustable within the range of 50 ps to 5 ns. When the preliminary estimate is less than a preset near-range threshold, the laser pulse width is adjusted to 50 ps to 500 ps and the receiving window delay is set to 0.5 nanoseconds to 2 nanoseconds. When the preliminary estimate is greater than a preset far-range threshold, the laser pulse width is adjusted to 1 ns to 5 ns and the receiving window delay is set to 10 nanoseconds to 50 nanoseconds. The adaptively adjusted laser pulse parameters and receiving window parameters are obtained. Based on the adaptively adjusted laser pulse parameters and the receiving window parameters, the flight time data and received light intensity information corresponding to different pulse widths are fused, and the light intensity distribution characteristics within the receiving window are combined with the flight time data through an intensity-assisted distance judgment algorithm to obtain full-range distance measurement capability. Based on the full-range distance measurement capability, for high reflectivity targets identified in the full-range distance measurement capability whose received light intensity is greater than a preset high reflectivity threshold, a multi-zone detector array is used to achieve multi-channel parallel sampling, dispersing the photon arrival time received by a single detector to multiple detection channels, thereby obtaining a dispersed photon arrival time distribution. Based on the dispersed photon arrival time distribution, a corrected histogram is constructed by statistically counting the photons in each channel. The statistical histogram correction algorithm is applied to identify the characteristic peak of the accumulation effect in the dispersed photon arrival time distribution where the peak position is earlier than the actual flight time, and the time offset is calculated for compensation and correction. The compensated and corrected flight time data is used as the flight time data of the real-time laser pulse to obtain the ranging result after suppressing the photon accumulation effect.

[0014] Furthermore, the laser emitter is a VCSEL laser emitter, the reference detector and the main detector are both SPAD detectors, and the time-to-digital converter is a TDC time-to-digital converter.

[0015] This disclosure also provides a ranging calibration system based on dTOF lidar, including: The zero-meter reference calibration module is used to acquire the laser pulse signal emitted by the laser emitter, split the laser pulse signal into a reference optical path and a measurement optical path through an optical beam splitter, receive the laser pulse signal of the reference optical path through a reference detector and record the reference time point, receive the laser pulse signal reflected by the target through the measurement optical path through a main detector and record the reception time point, calculate the time difference between the reference time point and the reception time point and statistically analyze the zero-meter reference measurement data, and obtain the system reference error calibration parameter set by combining the quantization characteristics analysis of the time-to-digital converter. The error modeling module is used to establish a multi-dimensional error compensation mathematical model based on the system's reference error calibration parameter set, by collecting calibration data under different measurement conditions, including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is discretized to construct a combination of basis functions and solved by an optimization algorithm to obtain the error compensation mapping function. The real-time calibration module is used to acquire real-time laser pulse time-of-flight data and environmental parameters, input the environmental parameters into the error compensation mapping function to calculate the error compensation value, apply the error compensation value to the time-of-flight data and process it through a filter to obtain the calibrated measurement result; The self-optimization module is used to convert the calibrated measurement results into three-dimensional point cloud data, extract geometric feature points from the three-dimensional point cloud data for point cloud registration and optimize system extrinsic parameters, establish system drift monitoring indicators, and automatically update calibration parameters when the system drift monitoring indicators exceed a preset threshold, thereby obtaining a self-optimized system parameter set.

[0016] Compared with the prior art, the beneficial effects of the embodiments of this disclosure are as follows: 1. This embodiment integrates a reference detector at the laser emitter to establish a zero-meter reference optical path system, effectively eliminating the influence of the inherent delay of the laser driving circuit and achieving accurate calibration of the system reference error. At the same time, a multi-dimensional error compensation mathematical model including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model is constructed. Suitable basis functions are designed for different error sources to form a heterogeneous set of basis functions, which improves the model's expressive power and calibration accuracy.

[0017] 2. This embodiment designs an adaptive Kalman filter to smooth the measurement results, which can effectively reduce random fluctuations while retaining rapidly changing effective information, thus improving the stability and reliability of the measurement results. At the same time, by maximizing the mutual information between point clouds to optimize the system extrinsic parameters, the system self-calibration without the need for a dedicated calibration target is realized. A system drift monitoring index system is established, and the recalibration process is automatically triggered when key indicators exceed preset thresholds, thus realizing the self-optimization of system parameters and ensuring the long-term stable operation of the system.

[0018] 3. This embodiment employs near-range blind zone elimination and photon accumulation effect suppression techniques. By combining detector gating technology and variable pulse width control with a multi-zone detector array, multi-channel parallel sampling is achieved, effectively solving the problems of dTOF systems in near-range measurement and high-reflectivity target detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the ranging calibration method based on dTOF lidar in this embodiment. Figure 2 This is a schematic diagram of the ranging calibration system based on dTOF lidar in this embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the implementation methods of the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings.

[0022] Example 1: Figure 1 The flowchart of the ranging calibration method based on dTOF lidar in this embodiment is shown as follows. Figure 1 As shown, this embodiment provides a ranging calibration method based on dTOF lidar, which can be executed by a computer system; specifically, the method includes: Step S1: Acquire the laser pulse signal emitted by the laser emitter, divide the laser pulse signal into a reference optical path and a measurement optical path through an optical beam splitter, receive the laser pulse signal of the reference optical path through a reference detector and record the reference time point, receive the laser pulse signal reflected by the target through the main detector through the measurement optical path and record the reception time point, calculate the time difference between the reference time point and the reception time point and statistically analyze the zero-meter reference measurement data, and obtain the system reference error calibration parameter set by combining the quantization characteristics analysis of the time-to-digital converter. Step S2: Based on the system reference error calibration parameter set, collect calibration data under different measurement conditions to establish a multi-dimensional error compensation mathematical model including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. Use a discretization strategy to construct a combination of basis functions for the multi-dimensional error compensation mathematical model and solve it through an optimization algorithm to obtain the error compensation mapping function. Step S3: Obtain the real-time laser pulse time-of-flight data and environmental parameters, input the environmental parameters into the error compensation mapping function to calculate the error compensation value, apply the error compensation value to the time-of-flight data and process it through a filter to obtain the calibrated measurement result; Step S4: Convert the calibrated measurement results into three-dimensional point cloud data, extract geometric feature points from the three-dimensional point cloud data for point cloud registration and optimize system extrinsic parameters, establish system drift monitoring indicators, and automatically update calibration parameters when the system drift monitoring indicators exceed a preset threshold to obtain a self-optimized system parameter set.

[0023] The above steps will be explained in further detail below.

[0024] In this embodiment, in step S1, the laser pulse signal is divided into a reference optical path and a measurement optical path using an optical beam splitter. A reference detector receives the laser pulse signal from the reference optical path and records the reference time point. A main detector receives the laser pulse signal reflected from the target in the measurement optical path and records the reception time point. The time difference between the reference time point and the reception time point is calculated, and zero-meter reference measurement data is statistically analyzed. Combined with the quantization characteristics analysis of the time-to-digital converter, a system reference error calibration parameter set is obtained, including: Step S11: The laser pulse signal emitted by the laser emitter is divided into a reference optical path and a measurement optical path by an optical beam splitter to establish a zero-meter reference optical path system and obtain the zero-meter reference optical path configuration.

[0025] In this step, an optical beamsplitter is first placed in the optical output path of the VCSEL laser emitter. This beamsplitter employs a semi-transparent, semi-reflective structure, dividing the laser pulse signal emitted by the VCSEL into two parts: approximately 5-10% of the light energy is guided to the reference optical path, while the remaining 90-95% continues to propagate along the main optical path. The reference optical path is directly connected to the integrated reference SPAD detector within the system, allowing it to be detected without passing through an external environment; while the main optical path, after passing through the transmitting optical system, is directed towards an external target, reflected, and then converged by the receiving optical system to the main SPAD detector.

[0026] Understandably, this configuration forms a zero-meter reference optical path system, where the reference optical path provides an ideal zero-distance reference signal that can be used to calibrate the system's internal delay. The key to the zero-meter reference optical path system is ensuring that the optical signal received by the reference SPAD is strictly synchronized with the signal emitted by the main optical path, and that the optical path is stable and controllable.

[0027] Typically, to ensure the stability of the zero-meter reference optical path system, the aforementioned beam splitter is manufactured using a highly stable fused silica / ULE ultra-low expansion glass material and is precision-mechanically fixed near the VCSEL transmitter. Simultaneously, the length of the reference optical path is precisely measured and recorded in the system parameters to eliminate fixed deviations caused by the physical length of the optical path. In practical applications, the length of the reference optical path is usually controlled within the range of a few millimeters to a few centimeters and is precisely measured during assembly, with a measurement accuracy better than 0.1 millimeters. This allows for accurate modeling and compensation of changes in the length of the reference optical path even with variations in ambient temperature.

[0028] Typically, when designing a zero-meter reference optical path system, the choice of beam splitting ratio also needs to be considered. For example, a beam splitting ratio that is too small (e.g., less than 3%) will result in a weak reference signal, reducing the accuracy of time measurement; while a beam splitting ratio that is too large (e.g., greater than 15%) will excessively weaken the main optical path energy, reducing the system's measurement distance. This embodiment uses a beam splitting ratio of 5-10%, minimizing the impact on the main optical path while ensuring the quality of the reference signal.

[0029] In addition, to reduce interference from stray light and background noise, the reference optical path is typically equipped with a narrowband optical filter, allowing only optical signals matching the laser wavelength to pass through. In high-precision applications, the reference optical path may also include an adjustable attenuator to adjust the reference signal strength as needed, maintaining the optimal signal-to-noise ratio.

[0030] In summary, this step establishes a stable and reliable internal reference standard through the above-described zero-meter reference optical path configuration, which is beneficial for subsequent time delay measurements and error calibration.

[0031] Step S12: Based on the zero-meter reference optical path configuration, after receiving the laser pulse signal of the reference optical path through the reference detector, a time counter is triggered to record the reference time point. After receiving the laser pulse signal returned by the measurement optical path through the target through the main detector, a time counter is triggered to record the receiving time point. The time difference between the reference time point and the receiving time point is calculated to eliminate the influence of the inherent delay of the laser driving circuit and obtain the time difference after eliminating the delay of the driving circuit.

[0032] In step S12, after receiving the laser pulse signal from the reference optical path, the reference SPAD detector immediately generates an electrical pulse signal. This electrical pulse signal is sent to a time counter (usually part of a TDC time-to-digital converter) and recorded as the reference time point T0. The reference time point T0 actually marks the precise moment when the laser pulse leaves the VCSEL and is detected; it includes the response time of the VCSEL drive circuit and the inherent system delays such as the delay caused by the laser.

[0033] Simultaneously, the laser pulse in the measurement optical path is emitted towards the external target after passing through the transmitting optical system. After being reflected by the target, it is then focused onto the main SPAD detector through the receiving optical system. When the main SPAD detector receives the returned optical signal, it also generates an electrical pulse signal. This signal is also sent to the time counter and recorded as the reception time point T1.

[0034] Understandably, the actual flight time of the optical signal is obtained by calculating the time difference between the reference time point T0 and the receiving time point T1. The key advantage of this differential timing method is its ability to eliminate the influence of the inherent delay of the laser drive circuit. Traditional dTOF systems typically use the trigger signal of the laser drive circuit as the timing starting point, which introduces measurement errors caused by circuit delay fluctuations. However, using a reference SPAD detector to directly detect the actually emitted light pulse effectively avoids this problem.

[0035] The delay in the driving circuit mainly stems from factors such as the response time of electronic components in the VCSEL driving circuit, temperature sensitivity, and power supply fluctuations. These factors cause an unstable time delay between receiving the trigger signal from the driving circuit and the actual generation of the laser pulse, typically ranging from hundreds of picoseconds to several nanoseconds. Using a reference SPAD to directly probe the optical pulse can bypass these circuit delays and directly obtain the actual emission time of the optical signal, thereby improving measurement accuracy.

[0036] In practical operation, the operating mode of the SPAD detector is crucial. SPADs typically operate in Geiger mode, a single-photon detection mode where the absorption of a single photon triggers an avalanche effect, generating a measurable current pulse. To ensure consistent time measurements, the SPAD detector needs to be reset to the same initial state before each measurement. This is achieved through a quenching circuit, which rapidly reduces the SPAD bias voltage after detecting a photon, terminating the avalanche process, and then restores the bias voltage to prepare for the next detection.

[0037] Typically, the triggering mechanism of the time counter also needs to be designed accordingly. In this embodiment, a rising edge triggering mode is adopted, that is, the time counting is triggered when the SPAD output voltage exceeds a preset threshold. The choice of threshold level directly affects the measurement accuracy; too low a threshold level is easily affected by noise interference, while too high a threshold level may introduce additional delay. This embodiment ensures that time jitter is minimized while maintaining high sensitivity by precisely controlling the threshold level.

[0038] Furthermore, to improve the reliability of the measurement, the validity of each received pulse signal is verified, filtering out possible noise or abnormal signals. Verification criteria include pulse width, pulse amplitude, and pulse shape characteristics. Only pulses that pass verification are used for time difference calculation.

[0039] Based on this, step S12 can accurately measure the time difference between the reference time point and the receiving time point, effectively eliminating the influence of the inherent delay of the laser driving circuit, obtaining more accurate flight time data, and providing a reliable basis for subsequent distance calculation.

[0040] Step S13: Based on the time difference after eliminating the driving circuit delay, statistically analyze the zero-meter reference measurement data, construct a reference detector response time histogram, analyze the mean and standard deviation of the reference detector response time histogram, obtain the response time error model, quantify the receiver circuit delay, and obtain the receiver circuit delay parameters.

[0041] In step S13, under the condition that the target distance is controlled to be zero (or a known fixed distance), a large number of repeated measurements are performed (usually 10,000-100,000 times), and the time difference of each measurement is recorded to construct a statistical data set. These data are grouped by time interval to construct a histogram of the reference detector response time.

[0042] It should be noted that the reference detector response time histogram is a statistical distribution graph. The horizontal axis represents the measurement time difference, and the vertical axis represents the frequency or count of the corresponding time difference. Normally, when measuring a target at point-blank range, the time difference should be a fixed value. However, due to the random fluctuations in the response time of the SPAD detector, the actual measurement results will exhibit a certain distribution width.

[0043] Constructing a time histogram requires determining an appropriate time resolution (bin width). Too coarse a resolution will lose detail, while too fine a resolution may result in too few samples in each bin, affecting statistical significance. In this embodiment, the time resolution is set to 1-2 times the basic TDC resolution, typically within the range of 10-50 picoseconds.

[0044] Next, after the histogram is constructed, its key statistical properties are calculated. The mean represents the average response time, which is the main component of the system's fixed delay; the standard deviation characterizes the degree of time jitter, reflecting the system's time resolution capability. In addition, the system also analyzes the shape characteristics of the histogram, such as skewness (indicating the degree of distribution asymmetry) and kurtosis (indicating the degree of distribution sharpness), which help identify potential anomalous patterns.

[0045] The histogram of the reference detector's response time typically approximates a Gaussian distribution, but may exhibit a slight skew. This skew primarily stems from the physical characteristics of SPADs, specifically the random delay between photon absorption and avalanche formation, a process that is not perfectly symmetrical. By analyzing this skew characteristic, the system's response time error model can be further refined.

[0046] Typically, after histogram analysis, a SPAD response time error model is established. This model usually consists of two main components: a deterministic component, mainly determined by the mean, representing a fixed delay; and a random component, determined by the standard deviation and distribution shape, representing time jitter characteristics.

[0047] In this step, receiver circuit delay refers to the delay introduced throughout the entire electronic link from the arrival of the photon at the SPAD to the final generation of the timestamp. This includes multiple components such as SPAD response delay, preamplifier delay, signal conditioning circuit delay, and TDC trigger delay. Quantifying these delays is crucial for system calibration.

[0048] Based on the response time error model, delay parameters of the receiving circuit are extracted, including: fixed delay time (typically on the order of a few nanoseconds), temperature-dependent variation coefficient (typically tens of picoseconds / °C), signal strength-dependent variation characteristics (generally, the stronger the signal, the shorter the delay), and random jitter statistical parameters (standard deviation typically in the range of tens to hundreds of picoseconds). Furthermore, the correlation between multiple measurements is analyzed to detect short-term drift or long-term trends. This helps identify potential temperature effects or power supply instability issues, further refining the delay parameter model.

[0049] Through the above analysis process, step S13 obtains detailed receiver circuit delay parameters. These parameters will be used in subsequent ranging calibration to compensate for system errors introduced by the receiver and improve measurement accuracy.

[0050] Step S14: Based on the delay parameters of the receiving end circuit, analyze the quantization characteristics of the time-to-digital converter, and control the quantization error within a preset range through internal oscillator frequency calibration and interpolation algorithm to obtain quantization error compensation parameters.

[0051] Understandably, the Time-to-Digital Converter (TDC) is a key component in a dTOF system responsible for high-precision time measurement, converting time intervals into digital values. However, TDC suffers from quantization error, meaning it can only divide time into the smallest discrete unit (called time resolution or LSB, typically in the range of 10-100 picoseconds). This discretization process introduces quantization error, which can reach up to half an LSB.

[0052] In step S14, the quantization characteristics of the TDC are analyzed in detail, including basic time resolution testing, nonlinear error testing, and temperature stability testing. The basic time resolution test determines the actual minimum resolution of the TDC by repeatedly measuring known standard time intervals. The nonlinear error test is divided into differential nonlinear error (DNL) and integral nonlinear error (INL). DNL represents the deviation between the actual time interval and the ideal interval between adjacent digital codes, while INL represents the deviation between the actual output value and the ideal linear output. The temperature stability test evaluates the temperature dependence of the TDC performance by performing the above tests at different temperatures.

[0053] Based on the test results, a TDC quantization error model was established. This model includes not only the error of the ideal quantization process (maximum half an LSB), but also the nonlinear error and temperature dependence that may exist in actual TDC.

[0054] Meanwhile, in order to improve the time resolution of TDC and reduce quantization error, this embodiment adopts two key technologies: internal oscillator frequency calibration and time interpolation.

[0055] Specifically, internal oscillator frequency calibration addresses the accuracy and stability of the TDC's internal clock source. TDCs typically use a ring oscillator as their time reference, but the ring oscillator frequency varies with temperature, voltage, and manufacturing process. The calibration process consists of two phases: initial calibration and runtime calibration.

[0056] Typically, initial calibration is performed at the system's factory or during initialization, comparing and calibrating the ring oscillator frequency against an external high-precision clock source (such as a temperature-compensated crystal oscillator, TCXO). The specific steps are: generating a precisely known time interval (controlled by the TCXO), allowing the TDC to perform multiple measurements within this interval, calculating the average of the measurement results, and then adjusting the TDC's internal parameters to match the measured values ​​with the actual values. This process is usually repeated over multiple time intervals to ensure accuracy across the entire measurement range.

[0057] It should be noted that runtime calibration is performed continuously during system operation, adjusting compensation parameters by monitoring changes in the ring oscillator frequency. The system periodically uses an internal reference delay chain or an external reference clock to detect the oscillator frequency; if a frequency deviation exceeds a preset threshold (typically 20 ppm), the calibration parameters are updated. Through this dynamic calibration, the system can improve oscillator frequency stability to better than ±20 ppm, significantly reducing drift errors in time measurements.

[0058] Specifically, temporal interpolation algorithms are a key technology for achieving sub-LSB resolution. Their basic principle is to utilize the statistical distribution characteristics of multiple measurements to achieve more refined temporal measurements than the basic resolution. This embodiment employs a temporal interpolation method based on statistical histograms, and the specific implementation steps are as follows: Repeat the measurement multiple times (usually 100-1000 times) at the same time interval and record the digital output value of each measurement; construct a statistical histogram of these output values ​​and analyze its distribution characteristics; by analyzing the sample distribution ratio between two adjacent quantization levels on the histogram, estimate the precise location of the actual time point within the quantization interval.

[0059] For example, if the actual time point falls exactly in the middle of two quantization levels, then the frequencies of these two quantization levels should be approximately equal in multiple measurements; if the actual time point is closer to a quantization level, then the frequency of that quantization level will be higher. By analyzing this statistical distribution, the system can infer the precise location of the actual time point within the quantization interval, achieving sub-LSB resolution.

[0060] Typically, under optimal conditions, this time interpolation method can improve the effective resolution by 2-4 times. That is, if the basic resolution of TDC is 50 picoseconds, an effective resolution of 12.5-25 picoseconds can be achieved through interpolation.

[0061] Through the above steps, this embodiment can control the TDC quantization error within a preset range (usually less than 5 picoseconds), which is significantly better than the basic resolution of TDC.

[0062] Specifically, this step yields the following quantization error compensation parameters: the actual value of the TDC basic resolution, the DNL calibration table, the INL calibration table, the oscillator frequency calibration coefficient, the temperature compensation coefficient, and the time interpolation algorithm parameters. These parameters together constitute the TDC quantization error compensation mechanism, which can be used to improve the accuracy of time measurement in actual measurements.

[0063] Step S15: Based on the zero-meter reference measurement data, the receiver circuit delay parameters, and the quantization error compensation parameters, calculate the integrated system error calibration parameter set to obtain the system reference error calibration parameter set.

[0064] Step S15 involves integrating the parameters obtained from the previous steps to construct a comprehensive system error calibration parameter set. It should be noted that this process is not a simple parameter aggregation, but rather considers the potential mutual influence and coupling effects between multiple error sources to establish a unified error compensation framework.

[0065] In this step, firstly, a reference offset value is established based on the zero-meter reference measurement data. This reference offset value represents the time difference measured under ideal zero-distance conditions; theoretically, it should consist entirely of the system's internal delay and should not include any actual optical path propagation time. This reference offset value will serve as the reference zero point for all distance measurements.

[0066] Typically, calculating the baseline offset involves statistical analysis of a large amount of zero-meter measurement data. This includes not only calculating the average value but also detecting and filtering outliers (usually using the 3σ criterion or quartile-based methods). Furthermore, the temporal stability of the data is analyzed to detect any long-term drift trends; if present, a time-dependent compensation model may be necessary.

[0067] Secondly, by combining the receiver circuit delay parameters with the reference offset value, a complete receiver link delay model is constructed. This model describes the delay introduced throughout the entire process from the arrival of the photon at the SPAD to the final generation of the timestamp, including both fixed and variable components.

[0068] The receiver link delay model needs to consider the impact of signal strength. When the received optical signal strength changes, the SPAD's response time will change slightly; this phenomenon is called the walk effect. A model is typically established to model the relationship between signal strength (usually expressed as the SPAD's photon count rate or cumulative photon count) and time offset, used for accurate compensation under different signal conditions.

[0069] Then, by integrating the TDC quantization error compensation parameters, a time measurement error model is established. This model describes the errors introduced during time digitization and how to reduce these errors through compensation algorithms.

[0070] Typically, time measurement error models consist of two parts: deterministic errors (such as systematic biases caused by DNL and INL) and random errors (such as jitter caused by electronic noise). Deterministic errors can be accurately compensated for using lookup tables or function mappings; random errors are usually mitigated by increasing the number of measurements and performing statistical processing.

[0071] Finally, this step also considers the delays and errors introduced by the optical system, such as the dispersion effect of optical elements and the changes in the optical path caused by temperature, and incorporates these factors into the comprehensive error model.

[0072] It is understandable that optical system errors mainly arise from changes in the size and refractive index of optical components caused by temperature variations, which affect optical path length and signal propagation time. This embodiment establishes a temperature-dependent optical delay model to compensate for this error. Furthermore, if the system uses filters or other dispersive elements, the potential differences in propagation time for different wavelengths of light signals must also be considered.

[0073] Through the above steps, this embodiment generates a complete set of system reference error calibration parameters, including laser emission related parameters (emission delay, pulse shape characteristics, etc.), SPAD detector parameters (response time statistical characteristics, temperature variation characteristics, etc.), TDC parameters (quantization characteristics, nonlinear error compensation table, etc.) and optical system parameters (optical path delay, temperature sensitivity, etc.).

[0074] Typically, this parameter set is stored in the system as a data structure, containing multiple parts such as basic parameter values, calibration tables, temperature correlation coefficients, and algorithm parameters. During actual distance measurement, the corresponding error compensation value is queried or calculated based on the current measurement conditions (such as temperature, signal strength, etc.) and applied to the original measurement results to obtain a more accurate distance value.

[0075] Simultaneously, these parameters are stored in non-volatile memory, and a parameter version control and integrity check mechanism is designed to ensure the reliability and consistency of the parameters. In practical applications, calibration parameters are updated periodically based on long-term performance monitoring results to adapt to device aging and environmental changes, ensuring long-term measurement accuracy and stability.

[0076] In this embodiment, in step S2, based on the system reference error calibration parameter set, a multi-dimensional error compensation mathematical model is established by collecting calibration data under different measurement conditions, including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is then constructed using a discretization strategy to create a combination of basis functions, which is solved using an optimization algorithm to obtain the error compensation mapping function, including: Step S21: Based on the system reference error calibration parameter set, zero-meter reference calibration data and standard reflector ranging data are collected under different temperature conditions and different reflectivity conditions. Combining the quantization error compensation parameters and receiver circuit delay parameters in the system reference error calibration parameter set, a functional relationship model between temperature and time delay and a mapping relationship between reflection intensity and time offset are established to obtain temperature-time delay model parameters and reflectivity-time offset model parameters.

[0077] In this step, based on the system reference error calibration parameter set, systematic calibration data is collected and analyzed under different environmental conditions to establish model parameters that are adaptable to various measurement conditions.

[0078] To address temperature-related errors, tests are conducted in a temperature-controlled environment (temperature chamber or constant temperature chamber) at 5-10°C intervals within a temperature range of -40°C to 85°C. The test environment requires strict control of temperature uniformity and stability, typically requiring temperature fluctuations to be less than ±0.5°C, and a high-precision temperature sensor (accuracy better than ±0.1°C) is used to monitor the system temperature in real time. Before testing, the system needs to be stabilized at the target temperature for at least 30 minutes to ensure that the internal components reach temperature equilibrium.

[0079] Specifically, at each temperature point, two types of tests are performed: zero-meter benchmark testing and standard distance testing. Zero-meter benchmark testing utilizes an internal reference optical path, requiring no external target, and directly measures the system's internal delay. Standard distance testing uses a high-precision reflector placed at a precisely known distance (typically 1-5 meters, with a positional accuracy better than 0.1 mm) to measure actual ranging performance. The reflector is made of diffuse reflective material, has a high surface flatness and stable reflectivity, and is placed on a precision mechanical platform to ensure accurate and controllable positioning.

[0080] During testing, the system records temperature sensor readings, zero-meter reference time difference, standard distance measurements, and other auxiliary parameters (such as SPAD dark count rate and signal strength). For each temperature point, the system performs at least 10,000 repeated measurements to ensure statistical validity. The collected data undergoes preliminary filtering to remove obvious outliers (e.g., data points deviating from the mean by more than 3 standard deviations).

[0081] In this step, a functional model of the relationship between temperature and time delay is established by analyzing the data. Specific steps include: first, plotting a scatter plot of time delay versus temperature to visually observe the relationship pattern; then performing curve fitting, typically starting with a low-order polynomial (such as a second-order polynomial) to evaluate the fitting error; if the fitting accuracy is insufficient, attempting to increase the order or use a piecewise polynomial; finally, evaluating the model quality through residual analysis and cross-validation.

[0082] It should be noted that temperature-time delay models are typically expressed as piecewise polynomials, with different orders of polynomials used for fitting different temperature ranges. This model quantifies the change in time delay caused by a 1°C change in temperature, which is typically 10-50 picoseconds / °C in the mid-temperature region (15-35°C), and may be even greater in extreme temperature regions. Model parameters include the coefficients of each piecewise polynomial, the temperature values ​​at the piecewise points, and the calibration parameters of the temperature sensor.

[0083] To address reflectivity-related errors, this embodiment uses a variety of standard reflective materials with different reflectivities for testing, ranging from 5% (similar to a black object) to 95% (highly reflective materials). At least 5-7 standard materials with different reflectivities are prepared, and the actual reflectivity of each material is accurately measured using a calibrated reflectometer. These materials are placed at the same distance (typically 3-10 meters, within the system's typical operating distance range), and the system records the received light intensity and the corresponding measurement time difference.

[0084] Typically, to ensure consistent testing conditions, factors such as ambient light intensity, target distance, and incident angle must be strictly controlled. The light intensity signal is usually represented by the photon count rate or cumulative photon count of the SPAD, and requires normalization to eliminate the influence of distance.

[0085] This embodiment establishes a mapping relationship between reflectance and time offset by analyzing the relationship between reflectance and measurement time offset. First, the scatter plot of light intensity versus time offset is analyzed to identify key change points and trends. Then, based on the data distribution characteristics, a suitable mapping function is selected, which may be an exponential function, a logarithmic function, or a piecewise function. Finally, the function parameters are determined through data fitting, and residual analysis is performed to verify the accuracy of the model.

[0086] The reflectivity-time migration model primarily describes the travel time effect, where variations in light signal intensity lead to subtle differences in the SPAD triggering time. For high-reflectivity targets, the detector typically triggers earlier; while for low-reflectivity targets, the triggering time may be delayed. This effect can result in time migrations ranging from tens to hundreds of picoseconds, corresponding to distance errors of several centimeters or even more. Model parameters include the reflectivity inflection point, maximum time migration, compensation coefficients for low-reflectivity regions, and compensation coefficients for high-reflectivity regions.

[0087] In this step, the parameters of the temperature-time delay model and the reflectivity-time offset model are evaluated for effectiveness through rigorous validation tests. The validation tests use a separate test set, different from the modeling dataset, including test data at intermediate temperature and reflectivity points. A model is considered effective and used in the subsequent integrated error compensation system only if it meets the preset accuracy requirements in the validation tests (typically requiring the model prediction error to be less than 30% of the basic accuracy of the measurement system).

[0088] Step S22: Based on the temperature-time delay model parameters and the reflectivity-time offset model parameters, analyze the time histogram characteristics of the received signal and identify the multi-peak distribution. Design a peak separation algorithm to distinguish between direct reflection signals and secondary reflection signals. Monitor the background count rate change of the detector and design an adaptive threshold algorithm to maintain the signal-to-noise ratio within the preset signal-to-noise ratio threshold range. Obtain the multipath effect detection model parameters and ambient light compensation model parameters.

[0089] Step S22 involves modeling and compensating for more complex measurement scenarios, such as multipath effects and ambient light interference.

[0090] Specifically, the multipath effect refers to the fact that the laser signal does not arrive at the detector through a unique path. In addition to the direct reflection path, there may also be indirect paths after reflection from other objects. These signals from different paths will form multiple time peaks at the receiver, leading to ambiguity or errors in the ranging results.

[0091] To study multipath effects, this embodiment employs various test scenarios, including corner reflections (corners formed by two or three orthogonal planes), narrow passages (passages with reflective objects on both sides), and multi-object scenarios (the main target is surrounded by other reflective objects). These scenarios are specifically designed to generate controllable multipath effects, facilitating systematic research and modeling.

[0092] Typically, in each test scenario, a large number of measurements are performed (usually 1,000-10,000), and the raw time data for each measurement is recorded, not just the final processed distance value. This raw data is used to construct high-resolution time histograms, typically using a time resolution (bin width) of 10-20 picoseconds, to ensure that subtle multi-peak structures are captured.

[0093] It should be noted that time histogram analysis is a key technique for detecting multipath effects. In the case of a single reflection path, the time histogram usually exhibits a single-peak Gaussian distribution; however, when multipath effects are present, the histogram may show a multi-peak structure, with the main peak representing the direct reflection path and the secondary peak representing secondary reflections or other indirect paths.

[0094] This embodiment employs a peak separation algorithm to identify and distinguish these peaks. The algorithm first determines significant peaks using an adaptive thresholding method. The adaptive threshold is dynamically adjusted based on the background noise level of the histogram, typically set to the background mean plus several times the standard deviation. Then, peak fitting techniques are applied to analyze the multi-peak structure, usually using a Gaussian mixture model (GMM) or a multi-peak fitting method. These methods treat the histogram as a superposition of multiple Gaussian distributions, determining the position, width, and intensity of each Gaussian component through iterative optimization.

[0095] For each identified peak, a series of characteristic parameters typically need to be calculated, including peak position (corresponding to flight time), peak height (corresponding to signal strength), peak width (corresponding to time jitter), and the relative relationships between peaks (such as time interval ratios). These parameters are used to evaluate the credibility of each peak and construct a peak priority scoring function. The scoring criteria include: peak intensity (the stronger the peak, the more credible it is), peak width (the narrower the peak, the more credible it is), and consistency with theoretical expectations (based on distance prior information).

[0096] Typically, multipath effect detection model parameters include peak detection threshold function parameters, Gaussian mixture model parameters, peak feature extraction parameters, and peak scoring function parameters. These parameters enable the system to accurately identify the reflected signals of real targets in complex environments and avoid ranging errors caused by multipath interference.

[0097] As those skilled in the art will know, ambient light interference is another challenge faced by dTOF systems. Strong ambient light (such as direct sunlight) increases the background count rate of the SPAD, reduces the signal-to-noise ratio, and affects measurement accuracy. To investigate the impact of ambient light, the system was tested under different ambient light conditions (from complete darkness to simulated strong sunlight), and the changes in the SPAD's background count rate, signal count rate, signal-to-noise ratio, and ranging accuracy were recorded.

[0098] The system continuously monitors the dark count rate (DCR) of the SPAD when there is no laser emission, which reflects the intensity of ambient light. The monitoring process involves short-term sampling during the intervals between laser pulses (typically a few microseconds), and the average DCR is calculated by accumulating multiple sampling results. When the background count rate increases, the system correspondingly raises the signal detection threshold to ensure that the real signal can be effectively distinguished from background noise.

[0099] It's important to note that the adaptive thresholding algorithm is the core of ambient light compensation. This algorithm dynamically calculates the optimal detection threshold based on the current background count rate. Under low background light conditions, a lower threshold can be used to improve sensitivity; under strong background light conditions, the threshold needs to be increased to avoid false detections. Threshold calculation is typically based on the constant false alarm rate (CFAR) principle in signal detection theory, which maintains the same false detection probability under different background conditions.

[0100] In addition to adjusting the threshold, other measures can be taken to cope with strong ambient light, such as increasing the number of measurements (from the standard hundreds to thousands); extending the integration time (maximizing signal accumulation within an acceptable refresh rate range); and adjusting SPAD operating parameters (such as overvoltage, quenching time, etc.) to optimize its performance under strong light. These strategies work together to ensure that the system maintains a sufficient signal-to-noise ratio (typically requiring SNR > 10 dB) under various lighting conditions.

[0101] Typically, ambient light compensation model parameters include background count rate baseline, threshold adjustment function parameters, minimum signal-to-noise ratio requirement, and measurement strategy parameters (such as pulse count and integration time) for different ambient light intensities. These parameters enable the system to maintain stable ranging performance under various lighting conditions.

[0102] In summary, the multipath effect detection model and the ambient light compensation model were validated and optimized through real-world scenario testing. Validation tests included typical application scenarios (such as indoor, outdoor, and complex environments) and extreme condition tests (such as strong backlighting and high-reflection interference). Model parameters were fine-tuned based on the validation results to ensure effectiveness and robustness in real-world application environments.

[0103] Step S23: Based on the temperature-time delay model parameters, the reflectivity-time offset model parameters, the multipath effect detection model parameters, and the ambient light compensation model parameters, integrate the temperature-time delay model, the reflectivity-time offset model, the multipath effect detection model, and the ambient light compensation model into a multi-dimensional error compensation mathematical model, construct a ranging error compensation function that includes temperature, reflection intensity, target distance, and background light intensity, and obtain a multi-dimensional calibration mathematical model.

[0104] It is understandable that step S23 integrates the previously established models into a unified multi-dimensional error compensation mathematical model to achieve comprehensive processing of multiple error sources.

[0105] The integration process first requires addressing the mutual influence and coupling effects between different error sources. For example, temperature not only directly affects system delay but also influences the SPAD's response characteristics to light intensity, thus affecting reflectivity-related errors; similarly, ambient light intensity also affects the SPAD's sensitivity and the accuracy of the reflectivity model. This complex interaction is difficult to express using a simple superposition model.

[0106] To understand these interaction effects, this embodiment designs orthogonal experiments, which involve varying one factor while controlling for other variables and measuring its impact. For example, under fixed reflectivity and ambient light conditions, the system response at different temperature points is tested; or under fixed temperature and distance conditions, the effect of different ambient light intensities is tested. Through these experiments, the system can establish an interaction matrix between variables, quantifying the degree of mutual influence among the factors.

[0107] Based on interaction matrix analysis, this step employs a multivariable function approach to construct a comprehensive error compensation function E=f(T,I, D, B), where T represents temperature, I represents received light intensity (reflecting target reflectivity), D represents target distance, and B represents background light intensity. This function can be expressed as a nonlinear combination of various sub-models, including individual influence terms and interactive influence terms.

[0108] Understandably, individual influence terms represent the independent effects of temperature, reflectance intensity, distance, and background light, while interaction influence terms represent the interactions between two or more factors. Based on interaction matrix analysis, this embodiment selectively retains those significant interaction terms, typically including the interaction between temperature and reflectance intensity, the interaction between reflectance intensity and background light, and the interaction between reflectance intensity and distance, while ignoring those higher-order interaction terms with smaller impacts.

[0109] Typically, a major challenge in constructing such multidimensional functions is the need for extensive test data covering the multidimensional parameter space, making comprehensive testing of all possible parameter combinations impractical. To address this issue, this embodiment employs an intelligent sampling strategy, specifically including the following strategies: Uniform Sampling: Selecting several key points evenly distributed across each dimension to form the basic test set. For example, selecting 5 temperature points, 5 reflectivity points, 3 distance points, and 3 ambient light conditions, resulting in 225 combined points for testing; Boundary Point Density: Increasing the sampling point density in boundary regions of each dimension (such as extremely high / low temperatures, extremely high / low reflectivity, etc.), as these regions typically exhibit more drastic changes and require more refined modeling; Random Validation Points: In addition to the regular sampling points, randomly selecting some combined points for testing to verify the model's generalization ability; Adaptive Sampling: Based on the preliminary test results, increasing sampling points in regions with large errors or drastic changes to optimize model accuracy.

[0110] Meanwhile, another challenge is that the complexity of the function makes computation and storage difficult. This embodiment employs mathematical approximation methods to simplify the function representation: Principal Component Analysis (PCA): Analyzes the principal components of each influencing factor, prioritizing the modeling of those components that contribute the most, thus simplifying model complexity; Piecewise Functions: Uses different function forms in different parameter ranges to balance accuracy and computational complexity. For example, a more refined model may be used in commonly used temperature ranges, while a simplified model is used at extreme temperatures; Lookup Tables and Interpolation: For particularly complex relationships, discrete lookup tables can be used to store the values ​​of typical points, and then multidimensional interpolation can be used to calculate the values ​​of intermediate points, avoiding complex function calculations.

[0111] It should be noted that, through these strategies, this embodiment ultimately constructs a multidimensional calibration mathematical model, which can calculate the corresponding error compensation value based on the current measurement conditions (temperature, reflection intensity, distance, and background light).

[0112] Specifically, the model parameters include baseline calibration parameters (from system baseline error calibration), temperature compensation function parameters (polynomial coefficients or piecewise function parameters), reflectivity compensation function parameters (usually nonlinear function parameters), distance nonlinearity compensation parameters (correcting distance-related errors), ambient light compensation parameters, and main interaction effect parameters (such as temperature-reflectivity interaction coefficients).

[0113] Model validation is the final step in the integration process. In this embodiment, model performance is evaluated using an independent test set. Validation testing typically includes the following parts: basic validation (using test points under normal operating conditions), boundary validation (testing performance under extreme conditions), and long-term stability validation (evaluating the model's stability after long-term use). Generally, only models that have passed comprehensive validation will be ultimately applied to the product.

[0114] In step S23, the resulting multidimensional calibration mathematical model is the core of the system to achieve high-precision ranging. It can adapt to a variety of complex measurement conditions and provide stable and reliable ranging results in practical applications.

[0115] Step S24: Based on the multidimensional calibration mathematical model, the function space of the multidimensional calibration mathematical model is adaptively partitioned. Different types of basis functions are constructed to form a heterogeneous basis function set for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters. An adaptive mesh refinement algorithm is designed based on the heterogeneous basis function set, and the optimized mesh distribution is obtained through iterative optimization. Based on the optimized mesh distribution, the conjugate gradient method and multigrid technique are used to solve the linear equation system generated by the segmented polynomial discretization to obtain the error compensation mapping function.

[0116] In step S24, the complex multidimensional calibration mathematical model needs to be transformed into a practically usable error compensation mapping function. It should be noted that due to the high dimensionality and complexity of the model, direct solution or expression may be very difficult; therefore, numerical methods are required for discretization.

[0117] First, the function space of the multidimensional calibration mathematical model is adaptively partitioned. Traditional uniform grid partitioning leads to the curse of dimensionality in high-dimensional spaces, where the number of grid points increases exponentially with the dimension, becoming difficult to manage. For example, if each dimension is divided into 10 grid points, 10,000 grid points would be needed in four-dimensional space (temperature, reflectivity, distance, ambient light), making computation and storage extremely difficult.

[0118] To address this issue, this step employs an adaptive partitioning strategy, using a denser grid in regions of rapid error variation and a sparser grid in regions of gradual error variation. The key to adaptive partitioning is evaluating the rate of error change for each region, typically measured by the gradient or second derivative of the error function.

[0119] The specific implementation steps include: First, an initial evaluation is performed using a relatively coarse uniform grid (e.g., 5-10 points per dimension), and the error function value is calculated at each grid point. Then, the rate of change of error between adjacent grid points is calculated, typically using a first-order difference to approximate the gradient. For regions where the rate of change exceeds a preset threshold (e.g., twice the average rate of change), the grid point density is increased to further subdivide the area.

[0120] The subdivision can be either equal subdivision (dividing a single mesh cell into multiple sub-cells) or unequal subdivision (adaptively adjusting the size of the sub-cells based on the gradient direction). The above process is then repeated on the subdivided mesh until the preset accuracy requirement is met or the maximum subdivision level is reached.

[0121] Understandably, this adaptive meshing strategy can significantly reduce the number of mesh points that need to be processed while maintaining accuracy in critical areas. For example, a denser mesh is used in areas where temperature changes drastically affect system performance (such as during low-temperature startup or in high-temperature operating regions), while a coarser mesh is used in intermediate-temperature regions where the temperature effect is relatively mild.

[0122] Next, on the partitioned function space grid, appropriate basis functions need to be selected to represent the error compensation function. Considering the differences in the physical characteristics of different error sources, the system constructs different types of basis functions for different types of errors, forming a heterogeneous set of basis functions.

[0123] Temperature errors typically exhibit relatively smooth changes, thus second- to third-order polynomial basis functions are used for characterization. Polynomial basis functions possess global smoothness and good analytical properties, making them suitable for expressing variables like temperature that change relatively smoothly and have physical continuity. Second-order polynomials can capture simple nonlinear relationships, while third-order polynomials can express more complex change patterns; however, higher orders may lead to overfitting.

[0124] Typically, reflectivity error exhibits piecewise characteristics, with different variation patterns across different reflectivity ranges. To address this, the system employs cubic spline basis functions. These functions consist of multiple cubic polynomials, ensuring the continuity of function values, first derivatives, and second derivatives at the connection points (called knots). Cubic splines combine local adaptability with global smoothness, accurately capturing piecewise variations without introducing overfitting issues.

[0125] Distance errors, especially those related to multipath effects, often exhibit local characteristics. This embodiment uses Gaussian radial basis functions (RBFs) to characterize this local response. A key feature of the Gaussian RBF is its local support; it exhibits a significant response near the center point, while decaying rapidly away from the center. This is more suitable for expressing error patterns that occur only within a specific distance range, such as multipath effects.

[0126] Based on the aforementioned heterogeneous basis function set, this embodiment designs an adaptive mesh refinement algorithm. This algorithm optimizes the mesh distribution through an iterative approach, balancing computational accuracy and efficiency.

[0127] Specifically, the main steps of the above algorithm include: (1) Initialization: Construct an initial error function approximation on the coarse grid using the above heterogeneous basis function set; (2) Error evaluation: Calculate the approximation error on a series of test points (including grid points and random verification points); (3) Grid refinement: Refine the grid for regions where the error exceeds the threshold and update the grid distribution; (4) Basis function adjustment: Adjust the basis function distribution according to the refined grid, which may require adding new basis functions or adjusting the parameters of existing basis functions; (5) Refitting: Recalculate the basis function coefficients on the new grid; Repeat steps (2)-(5) until the accuracy requirement is met or the maximum number of iterations is reached.

[0128] Meanwhile, to improve refinement efficiency, the algorithm employs an error metric-driven adaptive strategy, concentrating computational resources on the regions where accuracy improvement is most needed. Error metrics are typically based on the difference between model predictions and actual measurements, and can be absolute error, relative error, or weighted error (with different weights assigned to different regions depending on application requirements).

[0129] Typically, after mesh generation and basis function selection, the system needs to solve a system of linear equations generated by discretization of the partitioned polynomials to determine the weight coefficients of each basis function. Since the system of equations can be very large, direct solution would be computationally expensive. Therefore, the system employs an iterative optimization method, specifically the conjugate gradient method, to improve solution efficiency.

[0130] It should be noted that the conjugate gradient method is an efficient algorithm for solving large-scale sparse linear equation systems. It avoids the computational burden of direct inversion and instead approaches the optimal solution gradually through iteration. The core idea of ​​the algorithm is to search along specific directions (conjugate directions) in each iteration. These directions are orthogonal to each other, which can avoid repetition and oscillation during the search process and accelerate the convergence speed.

[0131] In practice, a linear system of equations Ax = b is first constructed, where A is the coefficient matrix (composed of the values ​​of the basis functions at the sampling points), b is the target value vector (the actual measured error value), and x is the weight coefficient of the basis functions to be solved. Then, a guessed solution x0 is initialized, and the initial residual r0 = b - Ax0 and the initial search direction p0 = r0 are calculated. In each iteration, the system moves a certain step along the current search direction, updates the solution vector and residual, and then calculates the new conjugate direction. This process is repeated until the residual is less than a preset threshold or the maximum number of iterations is reached.

[0132] To further improve solution efficiency, this embodiment incorporates a multigrid technique. The multigrid method utilizes a hierarchical structure of grids with different resolutions. An approximate solution is first obtained on a coarse grid, and then the result is interpolated onto a fine grid as initial values ​​before further solving on the fine grid. This method can significantly reduce the number of iterations, especially when the convergence speed of high-frequency error components is very slow.

[0133] Specifically, the implementation steps of the multigrid technique include: first, constructing a grid hierarchy, typically starting with the finest grid and creating increasingly coarser grid layers by merging adjacent grid points; then, solving the equations on the coarsest grid to obtain an initial solution; next, interpolating this solution to a finer-level grid as an initial value for several iterative optimizations; repeating this process until the finest grid is reached to obtain the final solution. This method combines the advantages of fast convergence on coarse grids and high accuracy on fine grids, significantly improving solution efficiency.

[0134] Through the above solution process, this embodiment finally obtains the error compensation mapping function, which consists of the compensation value at each grid point, as well as the basis functions and weighting coefficients used for interpolation. This mapping function can quickly calculate the corresponding error compensation value based on the current measurement conditions (temperature, reflection intensity, distance, background light) without repeating the complex numerical calculation process.

[0135] Typically, to facilitate real-time applications, this mapping function can be implemented in different forms: for embedded processors, a lookup table and simple interpolation may be used to pre-calculate the values ​​of key points, and only simple interpolation needs to be performed at runtime; for more powerful processing platforms, the basis function expression can be used directly for calculation; or dedicated hardware (such as FPGA) may be used to implement fast calculation.

[0136] In summary, this error compensation mapping function is the final result of the system calibration process, which can provide accurate distance measurement under various complex measurement conditions, thus facilitating the high-precision application of dTOF lidar.

[0137] In this embodiment, step S24, which involves constructing different types of basis functions to form a heterogeneous basis function set for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters, includes: Step S241: Based on the temperature-time delay model parameters, analyze the physical characteristics of the temperature error, and use second- to third-order polynomial basis functions to characterize the smooth change relationship between temperature and time delay, thereby obtaining the low-order polynomial basis function of the temperature error.

[0138] In this step, it is necessary to first analyze in depth the mechanism and characteristics of the influence of temperature on the measurement accuracy of the dTOF system in order to determine the most suitable mathematical function form to characterize the temperature-time delay relationship. This embodiment, through experiments and common knowledge in the field, found that the main mechanisms by which temperature affects the measurement system include the temperature characteristics of the laser emitter (VCSEL), the temperature sensitivity of the SPAD detector response time, the temperature drift of the internal clock source of the TDC, and the thermal expansion of optical components.

[0139] Specifically, the temperature characteristics of a VCSEL laser emitter are reflected in the changes in its threshold current, output power, and wavelength with temperature. In particular, the change in threshold current alters the delay time between the driving signal and actual emission, typically manifesting as an increase in delay with increasing temperature; a typical temperature coefficient is approximately 0.5–2 nanoseconds / °C. This relationship stems from the physical property of semiconductor materials where the bandgap changes with temperature; increased temperature reduces the semiconductor bandgap, leading to an increase in the lasing threshold current and thus extending the emission delay time.

[0140] The temperature sensitivity of SPAD detectors is mainly manifested in the fact that high temperatures increase the dark count rate (DCR) and slightly reduce the photon detection efficiency (PDE), while also affecting the timing characteristics of avalanche triggering. The temperature coefficient is typically in the range of tens of picoseconds / °C. This is primarily because temperature affects the thermal generation rate of charge carriers and the dynamics of the avalanche process in SPADs. At high temperatures, the increased thermally generated charge carriers lead to an increase in DCR, while enhanced charge carrier scattering slows down the avalanche process.

[0141] Temperature drift in a TDC manifests as changes in the internal clock source and circuit delays with temperature, resulting in a change in the time measurement scale, typically around 10-100 ppm / °C. This is primarily due to the temperature dependence of the electrical characteristics of electronic components (transistors, capacitors, resistors, etc.), especially since ring oscillator-type clock sources are particularly sensitive to temperature.

[0142] This embodiment involves conducting detailed temperature characteristic tests within a temperature control chamber, ranging from -40°C to 85°C, in 5°C increments. The system collected comprehensive data on the temperature-time delay relationship. Analysis of this data revealed that the temperature-time delay relationship exhibits a relatively smooth change within the normal operating temperature range, with the curve typically displaying a slight curvature, indicating a certain degree of nonlinearity, although this nonlinearity is limited.

[0143] Based on this, this embodiment selects second- to third-order polynomial basis functions to characterize the temperature-time delay relationship. The reasons for choosing these functions include: first-order polynomials (linear functions) are too simple to capture the nonlinear characteristics of the temperature effect, especially the accelerating trend in extreme temperature regions; second-order polynomials can express the basic curvature of the temperature curve, suitable for describing the variation patterns in the moderate temperature range, and their second-order terms can capture the accelerating or decelerating trends of the temperature effect; third-order polynomials offer greater flexibility, capable of expressing more complex curvature changes, and are particularly suitable for characterizing the inflection point behavior that may occur in extreme temperature regions.

[0144] Those skilled in the art will know that although fourth-order or higher-order polynomials theoretically provide stronger fitting capabilities, actual tests have shown that increasing to fourth-order or higher usually improves fitting accuracy by less than 5%, while significantly increasing computational complexity and the risk of overfitting. Therefore, the system does not use higher-order polynomials.

[0145] Typically, the low-order polynomial basis functions for temperature error include a constant term (representing a fixed offset independent of temperature), a first-order term (representing linear temperature dependence), a second-order term (representing a quadratic temperature effect), and a third-order term when a more accurate fit is required. The system fits the coefficients of these basis functions using the least squares method and evaluates whether the third-order term is needed based on the fitting error (e.g., root mean square error RMSE). If adding the third-order term improves the fitting error by less than a preset threshold (usually 5%), only the second-order term is retained to simplify model complexity.

[0146] To improve the stability of numerical calculations, this embodiment uses an offset value relative to a reference temperature (e.g., 25°C) to represent temperature, rather than an absolute temperature value. This makes the order of magnitude of the polynomial coefficients more uniform and reduces rounding errors. In some cases, a single polynomial may not accurately represent the variation characteristics of the entire temperature range, especially since abnormal changes may occur at extreme temperatures. To address this, this embodiment employs a piecewise polynomial method, dividing the temperature range into three segments: a low-temperature region (-40°C to 0°C), a medium-temperature region (0°C to 50°C), and a high-temperature region (50°C to 85°C). Each segment uses an independent second- or third-order polynomial.

[0147] In summary, the low-order polynomial basis function for temperature error obtained in this embodiment is computationally simple and efficient, has few parameters, and clear physical meaning, making it more suitable for implementing real-time temperature compensation in resource-constrained embedded systems. The model's simplicity also facilitates the analysis of the main mechanisms affecting temperature, providing guidance for hardware optimization.

[0148] Step S242: Based on the reflectivity-time offset model parameters and the low-order polynomial basis function of the temperature error, analyze the piecewise characteristics of the reflectivity error, construct a piecewise continuous smooth function in different reflectivity intervals using cubic spline basis functions and ensure the continuity of the first and second derivatives to obtain the piecewise spline basis function of the reflectivity error.

[0149] Step S242 analyzed the complex relationship between reflectivity and measurement error, finding that this relationship exhibits obvious piecewise characteristics, requiring more flexible mathematical tools for accurate expression. Reflectivity error is closely related to the reflectivity characteristics of the target surface, with different reflectivity ranges displaying different error variation patterns.

[0150] This embodiment uses standard reflective materials with reflectivity ranging from 5% to 95% for detailed testing, including low-reflectivity materials (black velvet, matte black coating, 5-15%), medium-reflectivity materials (gray cardstock, standard gray card, 30-60%), and high-reflectivity materials (white cardstock, reflective material, 70-95%). These materials are placed at precisely known distances (typically the midpoint of the system's nominal operating distance, approximately 10 meters), and the system records the received light intensity and the corresponding measurement time difference. To ensure consistent testing conditions, all materials are kept at the same distance, angle, and ambient light conditions, and at least 10,000 measurements are performed for each material to obtain statistical significance.

[0151] After analyzing the experimental data, three distinct regional characteristics were observed: In the low reflectivity region (<10%), the signal is weak, the detector operates in a single-photon sensitive region, and the error-reflectivity relationship changes drastically. Typically, the measurement delay decreases rapidly with increasing reflectivity, because the weaker the signal, the later the effective trigger time, forming the so-called time-travel effect; In the medium reflectivity region (10%-60%), the detector operates in a stable single-photon detection mode, the response is relatively stable, the error-reflectivity relationship is approximately linear or low-order nonlinear, and the rate of change is relatively gentle; In the high reflectivity region (>60%), due to the excessive signal strength, the detector may saturate or a photon accumulation effect may occur. In this region, the error-reflectivity relationship becomes steep again, but in the opposite direction to that in the low reflectivity region, with the measurement time further advanced as reflectivity increases.

[0152] Typically, this piecewise characteristic is difficult to accurately represent using a single polynomial function. While using higher-order polynomials (e.g., 5th order or higher) can achieve good accuracy at the fitting points, it often produces unnatural oscillations near the piecewise points, leading to inaccurate predictions in the interpolation region. Considering these characteristics, the system chooses cubic spline basis functions as the best method to characterize reflectivity error.

[0153] Understandably, a cubic spline is a piecewise function composed of multiple cubic polynomials, characterized by maintaining the continuity of function values, first derivatives, and second derivatives at connection points (nodes). This continuity ensures that the error compensation function changes smoothly across the entire reflectivity range, avoiding abrupt changes or oscillations. Specifically: first, based on reflectivity-time offset test data, the locations of key nodes are determined, including the low reflectivity boundary (approximately 5%), the low-to-medium reflectivity transition point (10-15%), the medium-to-high reflectivity transition point (60-70%), and the high reflectivity boundary (approximately 95%). In regions with drastic changes (such as low-reflectivity and high-reflectivity areas), the node density is higher; in the moderately changing medium reflectivity area, the nodes can be sparser.

[0154] After determining the node positions, the corresponding time offset value is recorded at each node, serving as the basic data points for constructing the spline function. Cubic spline construction involves solving a system of linear equations to determine the coefficients of each cubic polynomial, ensuring that the overall function satisfies node value constraints and continuity conditions.

[0155] In practical implementation, the system uses either natural splines or complete splines as boundary conditions. Natural splines require the second derivative to be zero at the boundary points, which means that the curve transitions naturally and smoothly at the boundary; complete splines, on the other hand, require specifying the first derivative value at the boundary, which is usually determined based on the extrapolation trend of the physical model or experimental data.

[0156] After the cubic spline construction is completed, its performance is further evaluated through cross-validation. Specifically, a subset of points from the original data is randomly selected as the test set and not used in spline construction. Then, the remaining data is used to construct the spline function and predict the values ​​of the test points. The difference between the predicted values ​​and the actual measured values ​​is compared to evaluate the model's generalization ability. If the cross-validation error exceeds expectations, it may be necessary to adjust the node positions or increase the number of nodes and reconstruct the spline function.

[0157] Based on the above description, using cubic spline basis functions to characterize reflectivity error has the following advantages: it can accurately capture piecewise variation characteristics, avoiding the problem that a single function cannot take into account the characteristics of different regions; it effectively ensures the smoothness of the function, ensuring that error compensation transitions smoothly when reflectivity changes, without introducing artificial oscillations; it has local adjustment capabilities, modifying a node only affects the nearby region, not the distant region; it is not sensitive to noise and outliers, and is not prone to overfitting; it has high computational efficiency, given a reflectivity value, it is only necessary to determine the interval and apply the corresponding cubic polynomial.

[0158] In summary, this step establishes a piecewise spline basis function for reflectivity error using the above method, accurately expressing the time shift under different reflectivity conditions, and providing a key component for subsequent comprehensive error compensation.

[0159] Step S243: Based on the parameters of the multipath effect detection model, the low-order polynomial basis function of the temperature error, and the piecewise spline basis function of the reflectivity error, analyze the local characteristics of the distance error, and use the Gaussian radial basis function to characterize the local response characteristics of the distance-related error to obtain the Gaussian radial basis function of the distance error.

[0160] Step S243 involves modeling distance-related errors, particularly those caused by multipath effects. These errors exhibit significant local characteristics and require specialized mathematical tools for accurate representation. Multipath effects refer to the fact that the laser signal does not reach the detector via a unique path; in addition to the direct reflection path, there are indirect paths through reflections from other objects, leading to ranging errors.

[0161] Based on this, this embodiment employs various experimental environments to study distance-related errors, including standard distance testing and multipath scenario testing. The standard distance test utilizes a high-precision displacement platform, placing standard reflectors at multiple distance points within the system's measurement range (1-50 meters) to measure the system's error at different distances. The multipath scenario test designs special environments such as corner reflections (corners formed by two or three perpendicular walls), narrow passages (corridors with reflective objects on both sides), and multiple targets in front and behind (multiple targets at different distances existing simultaneously) to specifically induce multipath effects.

[0162] Specifically, distance-related errors have the following characteristics: at most distance points, the systematic error is relatively stable, but there may be some slowly changing systematic deviations; at specific distance points or intervals, local error peaks or abrupt changes may occur, which are usually related to multipath effects or internal system resonances; the error caused by multipath effects is highly dependent on the specific environment, and the same distance may exhibit completely different error patterns under different environmental structures.

[0163] Typically, this localization characteristic manifests as follows: at certain specific distance points or intervals, the measurement error may suddenly increase or exhibit a pattern different from the surrounding area, while in other areas the error may remain relatively stable. Traditional polynomial or spline basis functions struggle to efficiently express this localization characteristic, either requiring very high orders (leading to overfitting) or a large number of nodes (increasing computational complexity).

[0164] Specifically, the Gaussian radial basis function (RBF) is an ideal choice for solving such problems. Understandably, a Gaussian RBF is a function centered at a specific point that decays rapidly with increasing distance from that point. Its core characteristic is local support; it exhibits a significant response near the center point, while rapidly decaying to near zero further away. This property makes it well-suited for representing localized error patterns.

[0165] Understandably, the basic form of a Gaussian RBF (Regression-Based Function) is centered around a central point, with the function value decaying according to a Gaussian distribution as the distance from the central point increases. The parameter controlling the function width determines the function's influence range; the smaller the parameter, the narrower the function and the higher the degree of localization; the larger the parameter, the wider the function and the lower the degree of localization.

[0166] Based on the analysis of multipath effect detection model parameters and experimental data, distance locations where significant errors may occur are identified, and these locations are set as the centers of the Gaussian radial basis function. Specific steps include: analyzing time histogram data to identify distance ranges where multi-peak distributions frequently occur; testing target scenes with different structures and materials in an experimental environment and recording distance points with abnormal measurement errors; and combining theoretical analysis and experimental data to determine the set of center locations of the Gaussian radial basis function.

[0167] Typically, an appropriate width parameter is determined for each center location. The general principle is to use larger parameter values ​​(narrow functions) in regions of rapid error change to capture fast changes, and smaller parameter values ​​(wide functions) in regions of relatively gradual error change to ensure a smooth transition. Parameter values ​​are usually determined through iterative optimization to find the parameter set that best fits the experimental data.

[0168] A key characteristic of Gaussian radial basis functions (RBFs) is their ability to achieve local corrections without affecting values ​​at distant locations. For example, if the system detects a specific error peak at 5 meters, an RBF centered at 5 meters can be added to precisely correct for this error without significantly affecting measurements at 1 or 10 meters. This ability to locally correct allows the system to accurately model and correct errors at specific distances, making it suitable for handling localized problems such as multipath effects.

[0169] In addition, RBF has good function approximation ability and provides high flexibility, allowing the number of basis functions to be increased or decreased according to actual needs, balancing accuracy and computational complexity.

[0170] In practical applications, the total number of basis functions is usually limited (typically 20-50) to control computational burden. The placement strategy of basis functions is also crucial: they are densely placed in known problem regions and sparsely placed in general regions. Furthermore, the contribution of each basis function is periodically evaluated, and those with minimal contributions are removed to maintain model simplicity.

[0171] In summary, the Gaussian radial basis function for distance error, together with the aforementioned polynomial basis function for temperature error and spline basis function for reflectivity error, constitutes a heterogeneous basis function set. Each basis function is optimized for specific types of error characteristics, and together they achieve high-precision comprehensive error compensation.

[0172] Step S244: Based on the low-order polynomial basis function of the temperature error, the piecewise spline basis function of the reflectivity error, and the Gaussian radial basis function of the distance error, construct a composite basis function and calculate the weighting coefficients to obtain the heterogeneous basis function set.

[0173] Step S244 integrates the three types of basis functions constructed previously into a unified set of heterogeneous basis functions to express the complex relationships in the multidimensional error space. It should be noted that this integration is not a simple juxtaposition or superposition, but rather considers the potential interactions between the error sources to construct more complex composite basis functions.

[0174] Specifically, the construction of composite basis functions is based on a thorough analysis of the interaction relationships among error sources. This embodiment investigates the interaction effects among three main factors—temperature, reflectivity, and distance—through orthogonal experiments. The orthogonal experiments use the controlled variable method, changing one factor while keeping other factors constant, measuring its impact, and then comparing the measurement results under different combinations to quantify the strength of the interaction effect.

[0175] The results indicate the existence of several significant interaction effects: the interaction between temperature and reflectivity (temperature changes affect SPAD sensitivity and response time, which in turn affect the system's response characteristics to targets with different reflectivities); the interaction between reflectivity and distance (targets with different reflectivities may exhibit different error patterns at different distances); and the interaction between temperature and distance (temperature changes may affect the system's multipath resolution capability).

[0176] Based on the understanding of the above-mentioned interaction relationships, the system employs various methods to construct composite basis functions. Among them, direct multiplication is a commonly used method, which directly multiplies basis functions of different dimensions to form high-dimensional basis functions. However, this method leads to a rapid increase in the number of basis functions. If temperature has 5 basis functions, reflectivity has 7 basis functions, and distance has 10 basis functions, direct multiplication will produce 350 composite basis functions, which may result in excessive computation.

[0177] To control complexity, this embodiment typically employs a selective combination strategy, constructing composite basis functions only for dimension combinations that exhibit significant interactions, rather than mechanically constructing all possible combinations. For example, if a significant interaction is found between temperature and reflectivity, while the interaction between temperature and distance is weak, only a composite basis function for temperature and reflectivity might be constructed, instead of a composite basis function for temperature and distance.

[0178] Those skilled in the art will recognize that in high-dimensional spaces, sparse grid methods are often employed to further control the number of basis functions. Sparse grids are a multidimensional function approximation technique that uses a selected set of points to construct multidimensional interpolations instead of a full Cartesian product grid. This method significantly reduces the number of points required while maintaining good approximation accuracy. For three-dimensional problems (temperature, reflectivity, and distance), if there are 10 basic points in each dimension, a full grid would require 1000 points, while a sparse grid method might only require 100-200 points to achieve near-perfect accuracy.

[0179] Specifically, adaptive importance sampling is another method for controlling complexity. The system adaptively adjusts the basis function distribution density based on the contribution of each dimension and combination to the final error, using more basis functions in important regions and fewer in less important regions. Importance assessment is typically based on sensitivity analysis, calculating the derivative or rate of change of the error function with respect to each factor.

[0180] Using these methods, this embodiment constructs a heterogeneous basis function set containing both one-dimensional and multi-dimensional composite basis functions. Based on this set, the system calculates the weight coefficients of each basis function to construct a complete error compensation model.

[0181] The calculation of weight coefficients typically employs the following method: First, calibration data covering multiple combinations of conditions is collected to form a training set; then, a least squares problem is constructed with the goal of finding a set of weight coefficients that makes the weighted sum of the basis functions as close as possible to the error value of the actual measurement; considering that the number of basis functions may be large, direct solution may lead to overfitting, so L2 regularization (ridge regression) is usually introduced to prevent overfitting by penalizing large weight coefficients; finally, numerical optimization methods are used to solve the optimization problem to obtain the optimal weight coefficients.

[0182] In summary, step S244 ultimately yields a set of heterogeneous basis functions and their corresponding weight coefficients, which constitute a complete multidimensional error compensation mapping function.

[0183] Understandably, this mapping function can quickly calculate the corresponding error compensation value based on the current measurement conditions (temperature, reflectivity, distance, etc.), achieving high-precision ranging. To improve computational efficiency, the system may pre-calculate values ​​at specific grid points to form a lookup table, and calculate the value of any point at runtime through multidimensional interpolation, avoiding the need to calculate the complete basis function combination every time, thus significantly improving real-time performance.

[0184] In this embodiment, step S3, which involves acquiring real-time laser pulse time-of-flight data and environmental parameters, inputting the environmental parameters into the error compensation mapping function to calculate the error compensation value, applying the error compensation value to the time-of-flight data, and processing it through a filter to obtain the calibrated measurement result, includes: Step S31: Obtain the emission time and reception time of each laser pulse, calculate the original flight time, and simultaneously record the received light intensity, current ambient temperature, and background light intensity to obtain the original measurement data set.

[0185] This step marks the beginning of the real-time ranging process for dTOF lidar, requiring the acquisition of complete raw measurement data to lay the foundation for subsequent error compensation and data processing. First, the VCSEL laser emitter is triggered to emit a laser pulse. Simultaneously, some of the light energy is received by the reference SPAD detector through the reference optical path, generating a reference time point T0, marking the actual moment the light pulse leaves the system. The laser pulse in the main optical path is then directed towards the target through the optical system, reflected back by the target, and received by the main SPAD detector, generating a reception time point T1.

[0186] The system records these two time points using a high-precision time-of-flight record (TDC). The TDC resolution is typically in the 10-50 picosecond range, corresponding to a distance resolution of approximately 1.5-7.5 millimeters. Original flight time ΔT raw The time difference between T1 and T0 is calculated, which includes the actual propagation time of the optical signal and the fixed internal delay of the system.

[0187] Meanwhile, the system records environmental parameters related to the current measurement: received light intensity I (represented by the photon count rate or cumulative photon count recorded by the SPAD within the measurement window, reflecting the target's reflection characteristics and distance), ambient temperature T (measured by a high-precision temperature sensor built into the system, with an accuracy typically better than ±0.5°C), and background light intensity B (evaluated by the SPAD's dark count rate or ambient light count when the laser is not emitting).

[0188] The acquisition of these data requires precise timing control, and the system typically uses a main controller (MCU or FPGA) to manage the entire measurement process. The measurement process is divided into a preparation phase (setting SPAD bias and initializing TDC), a laser emission phase (triggering VCSEL and recording reference time), a signal reception phase (waiting for and recording return signals), and a data reading phase (acquiring data such as time difference and light intensity).

[0189] Typically, for each measurement, the system checks the data validity and filters out obviously abnormal data (such as weak signals or times outside the reasonable range). To improve measurement reliability, the system performs multiple measurements consecutively within a short period (usually a few milliseconds), forming a measurement batch, typically containing 100-1000 independent measurements. This batch measurement method can effectively suppress the influence of random noise and improve the statistical stability of the data.

[0190] Specifically, the system will measure the raw flight time ΔT each time. raw Data such as received light intensity I, ambient temperature T, and background light intensity B are organized into a structured set of raw measurement data to prepare for subsequent statistical analysis and error compensation.

[0191] Meanwhile, considering the need for real-time processing, the system typically adopts a circular buffer or sliding window data structure to retain only the measurement data within the most recent period, rather than accumulating it indefinitely. This satisfies the needs of statistical analysis while avoiding excessive consumption of storage resources.

[0192] Step S32: Based on the original measurement data set, construct a real-time time histogram for the flight time data of one hundred to one thousand consecutive measurements in the original measurement data set, analyze the main peak position and distribution width of the real-time time histogram, and obtain the statistical characteristics of the time histogram.

[0193] It should be noted that time histogram analysis is a key step in improving the ranging accuracy of the dTOF system. This embodiment extracts more accurate time information than a single measurement by analyzing the statistical distribution characteristics of a large amount of measurement data, while also evaluating measurement quality and identifying abnormal patterns.

[0194] First, extract 100 to 1000 consecutive flight time data points from the original measurement dataset. The choice of data volume needs to balance statistical validity and real-time performance: too little data (<100 times) may result in statistically insignificant characteristics, making it difficult to form a reliable distribution pattern; too much data (>1000 times) will increase processing latency and affect system response speed.

[0195] It should be noted that in most applications, 200-500 measurements is a good compromise, providing sufficient statistical samples while maintaining a good refresh rate (typically 10-30Hz).

[0196] Typically, when constructing a time histogram, the first step is to determine an appropriate time resolution or bin width. A resolution that is too coarse will lose detail, while a resolution that is too fine may result in an overly sparse histogram with unclear statistical characteristics. A general rule is to set the resolution based on the system's time jitter level, usually choosing 1-2 times the basic TDC resolution or 1 / 5-1 / 10 of the data standard deviation. For typical dTOF systems, a bin width of 20-50 picoseconds is a common choice.

[0197] Specifically, the process of constructing a time histogram involves assigning each flight time data point to a corresponding time interval (bin) and counting the number of data points within each interval. The horizontal axis of the histogram represents flight time, and the vertical axis represents the count or frequency of the corresponding time interval. To facilitate comparison of histograms of datasets of different sizes, the vertical axis is sometimes standardized to frequency or percentage.

[0198] Typically, a completed time histogram will exhibit a Gaussian-like distribution, but it may contain non-ideal features such as skewness, multiple peaks, or elongated tails. The system needs to extract key statistical features from this histogram, mainly including the location of the main peak and the width of the distribution.

[0199] The peak position represents the most likely time of flight and is typically determined by the bin position with the highest count in the histogram. This position should generally correspond to the actual time of flight of the optical signal. To improve accuracy, the system may use interpolation methods to determine the peak position instead of simply using the center values ​​of discrete bins. Common interpolation methods include parabolic fitting (applying a quadratic polynomial to the peak bin and its adjacent bins) and centroid methods (calculating a weighted average of the region near the peak).

[0200] Distribution width reflects the degree of temporal uncertainty or jitter in a measurement and is an important indicator for evaluating measurement accuracy. Commonly used distribution width metrics include standard deviation (reflecting the dispersion of the distribution) and full width at half maximum (FWHM, which is the width at which the histogram height drops to half its maximum value). Standard deviation is suitable for approximate Gaussian distributions, while FWHM makes fewer assumptions about the shape of the distribution and has a wider range of applications.

[0201] Typically, in addition to basic location and width features, this embodiment also analyzes other statistical properties of the histogram, including: skewness (representing the degree of distribution asymmetry), which can help identify nonlinear effects in the system; kurtosis (representing the degree of distribution sharpness), which can assess noise characteristics; and multimodal index (such as the ratio of secondary peaks to primary peaks), which is used to detect multipath effects.

[0202] Time histogram analysis can also identify anomalous measurement patterns. For example, a clearly bimodal or multimodal histogram structure may indicate multipath effects; an unusually wide distribution or long-tailed characteristics may indicate interference signals or detector anomalies. This information is crucial for assessing measurement quality and selecting appropriate processing strategies.

[0203] In summary, this step organizes the extracted statistical features (peak location, distribution width, skewness, kurtosis, etc.) into a structured time histogram statistical feature set, providing a basis for subsequent error compensation and quality assessment. These features are not only used to optimize current measurement results, but may also be used to monitor long-term system performance trends and support adaptive calibration functions.

[0204] Step S33: Based on the time histogram statistical characteristics, input the received light intensity, current ambient temperature, and background light intensity from the original measurement data set, along with the time histogram statistical characteristics, into the error compensation mapping function to calculate the corresponding error compensation value, thereby obtaining the real-time error compensation parameters.

[0205] In this step, a targeted error compensation value is calculated based on the current measurement conditions and time histogram characteristics to achieve real-time adaptive calibration. This process utilizes the multidimensional error compensation mapping function established in the aforementioned embodiments to map specific measurement conditions to corresponding error compensation values.

[0206] First, key environmental parameters are extracted from the raw measurement dataset: received light intensity I (representing the target's reflectivity), current ambient temperature T, and background light intensity B. These parameters need to be preprocessed and standardized to meet the input requirements of the error compensation mapping function. For example, light intensity may need to be normalized to the 0-1 range; temperature is typically expressed as an offset relative to a reference temperature (e.g., 25°C); and background light intensity may be expressed as a relative value or a logarithmic scale.

[0207] Simultaneously, the time histogram statistical features extracted in the previous step are also used as input, including the main peak position μ. main (representing average flight time), distribution width σ main These statistical characteristics include (indicating the degree of time jitter), skewness, kurtosis, and possible multimodal indicators. These characteristics provide important information about the current measurement quality and properties, helping to select the most appropriate compensation strategy.

[0208] Next, these parameters are input into the error compensation mapping function to calculate the corresponding error compensation value. The error compensation mapping function is a multidimensional function established during the system calibration phase using a large amount of test data, capable of predicting system errors based on input conditions. As mentioned earlier, a heterogeneous set of basis functions, including temperature polynomial basis functions, reflectivity spline basis functions, and distance Gaussian radial basis functions, was used in the function construction process.

[0209] Typically, the calculation process needs to be executed efficiently because it needs to be completed during real-time measurement. To this end, this embodiment can employ optimization strategies such as lookup table method (pre-calculating a series of error compensation values ​​under typical conditions and storing them in a lookup table, then calculating them through multidimensional interpolation when needed in real time), hierarchical calculation (first calculating the compensation values ​​of the main influencing factors, and then adding compensation for secondary factors as needed), parallel calculation (calculating the contributions of different basis functions in parallel on a platform that supports parallel processing), or hardware acceleration (using dedicated hardware such as FPGA or DSP to implement fast calculation of the error compensation function).

[0210] Typically, the calculated error compensation value includes several components: reference calibration compensation E base (From system baseline error calibration, compensating for inherent system delay), temperature-dependent compensation E temp (Delayed adjustment with temperature variation), reflectivity-related compensation E refl (Travel time effect compensation for targets with different reflectivities), multipath effect compensation E multi(If a multi-peak distribution is detected) and ambient light compensation E amb (Adjustments to account for the effects of background light). These compensation values ​​together constitute the real-time error compensation parameter set.

[0211] This step adjusts the compensation strategy based on the characteristics of the time histogram. For example, if a significant multimodal distribution is detected, multipath effect compensation may be enhanced; if the distribution width is abnormally large, the smoothing intensity may be increased; if the signal-to-noise ratio is low, the compensation step size may be reduced to avoid overcorrection. This adaptive strategy ensures optimal compensation results under various conditions.

[0212] In summary, real-time error compensation parameters are not only used to correct current measurement results, but are also recorded to monitor long-term system performance changes and adaptive parameter updates. They may also be used to analyze the changing trends of compensation parameters, identify possible hardware aging or environmental change patterns, and perform preventative maintenance or parameter adjustments in advance.

[0213] Step S34: Based on the real-time error compensation parameters, apply the real-time error compensation parameters to the original flight time in the original measurement data set for error correction. At the same time, design an adaptive Kalman filter based on the statistical characteristics of the time histogram to smooth the measurement results after error correction, reduce random fluctuations while retaining rapidly changing effective information, and obtain the filtered measurement data.

[0214] In this step, the calculated real-time error compensation parameters are applied to the raw measurement data, and the results are further optimized using adaptive Kalman filtering technology to balance measurement accuracy and response speed.

[0215] First, error compensation is applied to the raw flight time data. This process typically involves subtracting the error compensation value, ΔT, from the raw flight time. corrected =ΔT raw -E total E total It is the sum of all compensation components: E total =E base +E temp +E refl +E multi +E amb This compensation is not only applied to the location of the main peak in the time histogram, but may also be applied to each individual measurement point for more refined statistical analysis.

[0216] Even after error correction, the measurement results still contain a certain degree of random fluctuation, which usually requires further optimization through filtering techniques. The system employs an adaptive Kalman filter, an advanced filtering technique that can automatically adjust the filtering parameters according to the measurement characteristics, making it particularly suitable for processing signals with varying dynamic characteristics.

[0217] It's important to note that the core of the adaptive Kalman filter is a state-space model, which describes the measurement process as a dynamic system, including state equations and observation equations. The state equations describe how the system state (such as target distance and velocity) evolves over time; the observation equations describe the relationship between the measured values ​​and the system state. The Kalman filter estimates the true system state recursively, updating the estimate each time new measurement data is available.

[0218] Specifically, in dTOF distance measurement applications, the state variables of the Kalman filter typically include distance and velocity (and optionally acceleration), while the observation variables are the corrected distance measurements. In the prediction step, the filter estimates the current state based on the state from the previous moment, and then in the update step, it adjusts the estimate using new measurement data, iterating this process continuously.

[0219] Typically, key parameters of a Kalman filter include the process noise covariance Q (representing the degree of random variation in the system state) and the measurement noise covariance R (representing the uncertainty of the measurement). In traditional Kalman filters, these parameters are fixed, but in adaptive Kalman filters, they are dynamically adjusted based on real-time measurement characteristics.

[0220] The adaptive mechanism, based on the statistical characteristics of the time histogram, mainly considers the following factors: measurement noise assessment (dynamically adjusting the R value according to the width of the time histogram distribution; the wider the distribution, the greater the measurement uncertainty, and the R value increases accordingly); signal dynamics assessment (assessing the target motion state through the change patterns between continuous measurements; reducing the velocity component in the Q value when rapid changes are detected, enabling the filter to track changes faster); multipath detection response (when the time histogram shows multi-peak characteristics, it may increase the dependence on historical data and reduce the weight of potentially unreliable current measurements); outlier handling (increasing the measurement noise value of measurement points that deviate significantly from the predicted value, or ignoring them completely in extreme cases).

[0221] It's important to note that adaptive Kalman filtering balances filter strength and response speed. For example, when measurements are stable, the filter enhances smoothing and reduces random fluctuations; when a valid, rapid change is detected, the filter responds quickly, reducing hysteresis. This adaptability enables the system to provide both high-precision static measurements and accurate tracking of dynamic targets.

[0222] To further improve the robustness of the filter, this embodiment may employ advanced techniques such as the multi-model approach (running multiple filters with different parameters simultaneously and selecting the best result based on performance) or particle filtering (which performs better in nonlinear or non-Gaussian noise conditions).

[0223] In summary, after filtering, this embodiment obtains filtered measurement data, specifically including optimized distance estimates and possible velocity estimates. It is understood that these data smooth out random fluctuations while retaining meaningful variation information, providing a high-quality foundation for the final ranging results.

[0224] Step S35: Based on the filtered measurement data, calculate the calibrated flight time and distance values, and calculate the confidence score based on the statistical features of the time histogram to obtain the calibrated measurement results and the confidence score.

[0225] In step S35, the final ranging result is calculated based on the aforementioned filtered measurement data, and a quality evaluation index is assigned to each result so that the upper-layer application can make decisions as needed.

[0226] First, calculate the final distance value based on the filtered time-of-flight data: D = c·ΔT filtered / 2, where c is the speed of light, ΔT filtered It is the filtered flight time, divided by 2 because the light needs to travel to and from the target.

[0227] In practical applications, the system may need to consider the effect of atmospheric refractive index on the speed of light, especially in long-distance measurements or extreme environmental conditions. For indoor applications or short-distance measurements, it is generally assumed that the speed of light in air is close to the speed of light in a vacuum.

[0228] In addition to calculating distance values, this embodiment also evaluates the reliability of each measurement result, generating a confidence score. The confidence score is a comprehensive quality indicator that reflects the credibility of the measurement result, helping upper-level applications determine whether the result should be adopted. Confidence scores are typically normalized to a range of 0-100% or 0-1 for ease of understanding and use.

[0229] The confidence score is calculated based on several factors, primarily derived from the statistical characteristics of the time histogram and information from the filtering process: signal strength (higher received light intensity generally leads to more reliable measurements without saturation); distribution width (narrower time histogram distribution indicates higher measurement accuracy and corresponding confidence); peak shape (ideally, the time histogram should approximate a Gaussian distribution; the system evaluates the degree of matching between the actual distribution and the ideal Gaussian); filter residual (during Kalman filtering, the difference between the measured and predicted values ​​reflects measurement consistency; smaller residuals indicate a better fit between the measurement and the system model, resulting in higher confidence); and environmental conditions (measurement reliability may decrease in extreme temperatures, strong background light, or rapidly changing environments).

[0230] Understandably, these factors are combined in a weighted manner to form the final confidence score. The weights can be adjusted according to specific application requirements and system characteristics. For example, in environments with poor lighting conditions, the weight of the signal strength factor may be increased; when tracking fast-moving targets, the weight of the filter residual factor may be increased.

[0231] Typically, the calculated distance values ​​and confidence scores are organized into a structured final output, which may also include timestamps, target IDs (if the system supports multi-target tracking), and optional velocity estimates. These outputs can be transmitted to upper-layer applications via standard interfaces (such as UART, SPI, Ethernet, etc.) or used directly for control decisions within the system.

[0232] Specifically, confidence scores enable higher-level applications to implement more intelligent decision-making logic. For example, an autonomous driving system might react immediately to obstacle detection with high confidence, while waiting for more data confirmation for detections with low confidence; a robot navigation system might adjust positioning weights based on the confidence of map points; and an industrial measurement system might automatically filter valid data based on confidence.

[0233] This embodiment also enables historical tracking and analysis of measurement results, identifying long-term trends and anomaly patterns, and supporting predictive maintenance and automatic calibration. For example, if the confidence score is found to be continuously declining over time, it may indicate that the system needs cleaning or recalibration; if the confidence score is consistently low under specific environmental conditions, it may be necessary to optimize system parameters for those conditions.

[0234] In summary, this embodiment, through its comprehensive data processing and quality assessment mechanism, enables the dTOF lidar system to provide accurate and reliable ranging results, meeting the needs of various demanding applications.

[0235] In this embodiment, step S32 involves constructing a real-time time histogram from the flight time data of one hundred to one thousand consecutive measurements in the original measurement data set, analyzing the main peak position and distribution width of the real-time time histogram, and obtaining the statistical characteristics of the time histogram, including: Step S321: Obtain the flight time data of one hundred to one thousand consecutive laser pulse measurements in the original measurement data set, set the time resolution as the time interval width of the histogram, and obtain the time histogram construction parameters.

[0236] Step S321 is the preparatory stage for constructing the time histogram. This step requires determining an appropriate amount of data and time resolution to lay the foundation for subsequent statistical analysis. Specifically, the system extracts 100 to 1000 consecutive flight time data points from the original measurement dataset. The selection of this number requires balancing multiple factors.

[0237] Statistical validity requires a sufficiently large sample size to establish a stable statistical distribution. For a typical dTOF system, at least 100 measurements are needed to form a meaningful distribution pattern. Increasing the sample size can improve statistical stability, reduce the influence of random noise, and improve the accuracy of peak location and distribution width estimation.

[0238] Temporal resolution is directly related to the amount of data; increasing the amount of data means requiring a longer acquisition time, which may lead to a decrease in temporal resolution. In real-time applications, the system needs to provide measurement results at a sufficiently fast refresh rate, which limits the amount of data available for each statistical analysis. For example, if the system measures at a frequency of 100 kHz (10 microseconds per measurement), to achieve a refresh rate of 10 Hz, each analysis can contain a maximum of 10,000 measurements.

[0239] The dynamics of the target also affect the choice of data volume. If the target is stationary or moves slowly, a larger data volume (e.g., 500-1000 times) can be used to improve accuracy; if the target moves rapidly, the data volume needs to be reduced (e.g., 100-200 times) to avoid motion blur effect, i.e., the target position has changed significantly within the statistical window.

[0240] System resource limitations (such as processing power and memory capacity) also affect the amount of data that can be processed. In resource-constrained embedded systems, processing large amounts of data can lead to excessive computational burden and affect system response speed.

[0241] Taking these factors into account, an adaptive strategy is typically used to select the amount of data: increase the amount of data in static measurement scenarios or scenarios requiring high precision; reduce the amount of data in dynamic scenarios or scenarios with high real-time requirements. Some systems also implement multi-level processing, simultaneously providing fast results based on a small amount of data and high-precision results based on a larger amount of data.

[0242] After acquiring an appropriate amount of time-of-flight data, it is necessary to set the time resolution or bin width as the time interval width of the histogram. Too coarse a resolution will lead to information loss and an inability to capture subtle distribution features; too fine a resolution may result in excessively scattered data, with too few samples in each bin, making the statistical significance unclear. A general guideline is to set the bin width based on the system's inherent time uncertainties (such as SPAD jitter, TDC resolution, etc.), typically 1 / 5 to 1 / 10 of the measurement standard deviation or 1 to 2 times the TDC resolution.

[0243] For a typical dTOF system, a suitable bin width is usually in the range of 10–50 picoseconds. For example, if the system's measurement standard deviation is approximately 100 picoseconds, a reasonable bin width might be 20 picoseconds; if the basic resolution of the TDC is 30 picoseconds, the bin width might be chosen to be 30 or 60 picoseconds.

[0244] To implement an adaptive bin width strategy, it is typically dynamically adjusted based on the actual measurement distribution characteristics. For example, when the signal strength is high and the distribution is narrow, a finer bin width can be used to improve accuracy; when the signal is weak and the distribution is wide, the bin width can be increased to ensure that there are enough samples in each bin.

[0245] Typically, in addition to bin width, the coverage of the histogram also needs to be determined. The coverage should be wide enough to include all reasonable measurements, usually set to the mean of the expected measurements ± a few standard deviations. For cases where multimodality may exist (such as multipath effects), the coverage may need to be wider to capture all possible peaks.

[0246] In this step, parameters such as the selected data volume, bin width, and coverage are organized into time histogram construction parameters, preparing for the next step of histogram construction. These parameters may be dynamically adjusted according to measurement conditions to adapt to different application requirements and environmental conditions.

[0247] Step S322: Based on the time histogram construction parameters, divide the flight time data into corresponding time intervals, count the value of each time interval, construct a real-time time histogram, and obtain an initial time histogram.

[0248] In step S322, the flight time data is organized into a histogram for visualization and analysis based on the time histogram construction parameters determined earlier. This forms the basis for subsequent statistical analysis.

[0249] Histogram construction is a process of data classification and counting. The system first determines the time axis range of the histogram, usually based on the minimum and maximum values ​​of the data, and then appropriately expands the boundaries to ensure that all data points fall within the range.

[0250] Then, based on the set bin width, the entire time range is divided into a series of consecutive time intervals (bins). For example, if the range is 10.0-10.7 nanoseconds and the bin width is 20 picoseconds (0.02 nanoseconds), then the histogram will contain 35 bins, corresponding to time intervals such as [10.00-10.02) and [10.02-10.04).

[0251] Next, iterate through each data point in the time-of-flight dataset, determine which time interval it falls into, and increment the count for that interval. This process can be represented as: for each measured flight time t, calculate the index i of its corresponding bin = floor((tt) min ) / bin width Then increment the count value of the i-th bin by 1.

[0252] To improve processing efficiency, especially with large amounts of data, this embodiment can employ parallel processing or streaming processing techniques. Parallel processing divides the data into multiple batches, calculates the local histogram of each batch simultaneously, and then merges the results; streaming processing updates the histogram in real time as data arrives, without waiting for all data to be collected.

[0253] It should be noted that the constructed histogram is a discrete frequency distribution table, recording the number of measurements within each time interval. To facilitate comparison and analysis, the system may standardize the histogram, converting the original count values ​​into frequency (count value divided by the total number of measurements) or probability density.

[0254] In practical applications, histograms may exhibit some undesirable characteristics, requiring preprocessing: if the data is sparse, some bins may lack samples, leading to discontinuous histograms; if outliers exist, the histogram range may be too large, obscuring the main features. To address these issues, techniques such as smoothing (e.g., moving average) or outlier filtering can be used to optimize the histogram.

[0255] For histograms that may contain multiple peaks, sufficient detail needs to be preserved so that subsequent analysis can distinguish between different peaks. This may require the use of finer bin widths or specialized multi-peak analysis techniques.

[0256] The constructed histogram data is stored in a structured format, typically including an array of bin center values, corresponding count values ​​or frequency arrays, and related metadata (such as bin width, total number of samples, etc.). This data structure is called the initial time histogram, which forms the basis for subsequent peak analysis and feature extraction.

[0257] In some advanced systems, histogram construction may be an iterative process, adjusting parameters (such as bin width or range) based on preliminary analysis results to reconstruct a more optimized histogram. For example, if the peaks in the initial histogram are found to be too narrow, occupying very few bins, the bin width may be reduced to increase resolution; if the data is found to be distributed in a very small area, the histogram range may be reduced to increase the resolution of that area.

[0258] Step S323: Based on the initial time histogram, identify the position of the main peak in the initial time histogram using a peak detection algorithm, calculate the count value corresponding to the main peak position, determine the main peak position, and obtain the main peak position parameter.

[0259] Understandably, the core task of this step is to accurately identify the position of the main peak in the constructed time histogram. This position usually represents the most likely time of flight of the light signal, and accurate identification of the main peak position is crucial for improving ranging accuracy.

[0260] It should be noted that peak detection is a class of algorithms used to identify local maxima in a data sequence. In histogram analysis of dTOF systems, peak detection needs to consider factors such as noise, the possibility of multiple peaks, and accuracy requirements, and usually employs more complex methods than simply finding the maximum value.

[0261] The usual approach is to find points that meet the following conditions: the value of that point is greater than that of its neighboring points and greater than a certain preset threshold. This method is simple but easily affected by noise, and may produce multiple false peaks when the histogram fluctuates greatly.

[0262] To improve robustness, this embodiment typically employs more advanced peak detection algorithms, such as gradient-based methods, sliding window maximum methods, or wavelet transform methods. These methods are better able to resist noise interference and identify the true signal peaks.

[0263] Gradient-based peak detection first calculates the first-order difference (gradient) of the histogram, then finds points where the gradient changes from positive to negative; these points represent the transition from an increasing to a decreasing function, corresponding to local maxima. To reduce the influence of noise, the histogram can be smoothed first, such as by applying a Gaussian filter.

[0264] The sliding window maximum method slides a fixed-width window across the histogram to find the maximum value within the window. If this maximum value is located near the center of the window and is sufficiently large, it is considered a peak. The choice of window width is crucial: too narrow a window may capture noisy peaks, while too wide a window may miss narrower, more accurate peaks.

[0265] Wavelet transform decomposes the histogram at different scales to analyze the multi-resolution characteristics of the signal, effectively distinguishing between peak values ​​and noise in the true signal. While this method has high computational complexity, it offers excellent noise resistance and peak localization accuracy.

[0266] In multi-peak scenarios, it is necessary to determine which peak is the dominant peak. Commonly used criteria include peak height (the peak with the largest count), peak area (the peak with the largest area under its base), or prior knowledge (such as the peak closest to the expected location). In dTOF systems, the dominant peak is usually the earliest salient peak because it is most likely to correspond to a direct reflection path.

[0267] Understandably, once the bin corresponding to the main peak is determined, the count value (or standardized frequency value) of that bin is calculated. This value reflects the significance of the main peak and is an important indicator for assessing the reliability of the peak. The higher the peak value, the more significant it is relative to the background or other peaks, generally indicating a more reliable measurement.

[0268] To improve the accuracy of peak location, interpolation methods are typically used to estimate the precise location of the peak. Commonly used interpolation methods include parabolic fitting (applying a quadratic polynomial to the peak bin and three points of its left and right adjacent bins, and then calculating the extreme points of the fitted curve), centroid method (calculating the weighted average of the peak region, with the weights being the count values ​​of each bin), and Gaussian fitting (applying a Gaussian function to the peak region, and then using the fitting parameters to determine the precise peak location).

[0269] Through these refined positioning methods, this embodiment can achieve peak position accuracy superior to that of the bin width, significantly improving ranging resolution. For example, if the bin width is 30 picoseconds, interpolation may improve the peak position accuracy to 5-10 picoseconds. By recording the identified main peak position, peak count, and possible peak shape parameters as main peak position parameters, a foundation is provided for subsequent analysis.

[0270] Step S324: Based on the main peak position parameter, locate the data distribution area around the main peak position in the initial time histogram, analyze the peak shape characteristics around the main peak position using the full width at half maximum (FWHM) method or standard deviation calculation, calculate the distribution width around the main peak, and obtain the distribution width parameter.

[0271] After determining the location of the main peak in the previous step, step S324 involves analyzing the shape characteristics of the peak, particularly its distribution width. This reflects the temporal uncertainty of the measurement and is an important indicator for evaluating measurement accuracy and quality. First, the system needs to locate the data distribution area around the main peak in the initial time histogram. This area should be large enough to completely encompass the shape of the main peak, but not so large as to include other peaks or irrelevant background areas.

[0272] Currently, the common practice is to use the main peak as the center and extend a certain range to both sides. The size of this range may be based on the estimated distribution width (such as 3-5 times the typical time jitter value of the system) or the distance from the initially observed peak to the background level. After determining the analysis area, the system uses the full width at half maximum (FWHM) method or standard deviation calculation to analyze the peak shape characteristics and quantify the distribution width.

[0273] Understandably, half-height full width is an intuitive and widely used measure of distribution width. It is defined as the width where the histogram height drops to half its maximum value. The specific calculation steps include: determining the peak maximum count value, Max. Count ; Calculate the half-height threshold (Half) Max =Max Count / 2; Search to the left from the main peak position to find the first count value less than or equal to Half. Max The bin file records its location as Left. Edge Search to the right from the main peak position to find the first count value less than or equal to Half. Max The bin directory records its location as Right. Edge ; Calculate FWHM=Right Edge -Left Edge .

[0274] Typically, to improve accuracy, bin boundaries are not simply used as half-height points; instead, more precise half-height point positions are calculated through linear interpolation. For example, if the count value between two adjacent bins is greater than Half... Max Become less than Half Max Based on the count values ​​of these two bins, the exact half-height point position will be calculated using linear interpolation.

[0275] The full width at half maximum (FWHM) method is applicable to various peak shapes due to its intuitiveness, computational simplicity, and fewer assumptions about the distribution shape. However, it only uses information from a limited number of points in the distribution and may be unstable when there is significant noise or the distribution is irregular.

[0276] Another common method is to calculate the standard deviation, which provides better resistance to noise by utilizing information from the entire distribution. Typically, the standard deviation calculation involves: calculating the weighted average (mean) of all bins within the peak region, with the weights being the count values ​​of each bin; calculating the squared deviation of each bin from the mean, multiplying it by the count value of that bin; summing these weighted squared deviations, dividing by the total value, and taking the square root to obtain the standard deviation.

[0277] Standard deviation is more suitable for approximate Gaussian distributions, providing a statistical measure of the distribution's dispersion. For a perfect Gaussian distribution, FWHM is approximately equal to 2.355 times the standard deviation, and the system may use this relationship to test the Gaussian nature of the distribution.

[0278] In addition to FWHM and standard deviation, other shape characteristic parameters may also be calculated: skewness (measures the degree of asymmetry in the distribution), kurtosis (measures the sharpness of the distribution), and peak signal-to-noise ratio (the ratio of the height of the main peak to the level of background noise), etc.

[0279] Through these analyses, this embodiment obtains complete distribution width parameters, including FWHM and / or standard deviation, as well as other possible shape characteristic indicators. These parameters are used to evaluate the accuracy of the current measurement and may also be used to adaptively adjust system parameters or select the optimal processing strategy.

[0280] Step S325: Based on the main peak position parameter and the distribution width parameter, identify whether there is a multi-peak distribution in the initial time histogram, evaluate the signal-to-noise ratio based on the count value corresponding to the main peak position parameter and the background noise count value of the initial time histogram, determine the measurement quality, and obtain the statistical characteristics of the time histogram.

[0281] In step S325, a more comprehensive time histogram analysis will be performed, with particular attention to multi-peak detection and signal-to-noise ratio evaluation, to comprehensively assess the measurement quality and provide a basis for subsequent processing. Multi-peak detection is the process of identifying multiple significant peaks that may exist in the histogram, which is particularly important for detecting multipath effects and interference signals. Typically, the time histogram should have only one obvious peak, corresponding to the reflected signal of the real target. Multiple peaks may indicate the existence of multiple reflection paths or that the system is subject to interference.

[0282] Multi-peak detection is typically performed after the main peak analysis is complete, searching for other possible peaks in the remaining histogram after excluding the main peak region. The detection method is similar to that of main peak detection, but may use stricter criteria to avoid misclassifying noise fluctuations as valid peaks.

[0283] Currently, a common method for multi-peak detection is to set a height threshold relative to the main peak. For example, it might focus only on secondary peaks whose height exceeds the main peak by 20%, assuming that fluctuations below this threshold are simply noise. Another approach is to evaluate peaks by combining their width and shape, requiring effective peaks to have certain peak quality, such as area, half-width at half-maximum (FWHM) ratio, and height.

[0284] For each detected secondary peak, the system records its location, height, width, and other characteristics, as well as its relationship to the primary peak (e.g., time difference, height ratio). This information is used for multipath effect analysis and compensation. In particular, if a secondary peak with a significant time interval from the primary peak is found, it may indicate the presence of secondary reflections or other indirect paths, and the system may need to adopt special processing strategies.

[0285] Signal-to-noise ratio (SNR) evaluation is another crucial step, quantifying the strength of the effective signal relative to noise, directly impacting measurement reliability. In a time histogram, the peak represents the effective signal, while the count level in the background region represents the noise baseline. SNR calculation typically involves: determining the peak signal strength S (usually the peak height minus the background noise level); estimating the background noise level N (calculating the average count value far from the peak region, or extracting the noise baseline after baseline fitting across the entire histogram); and calculating the SNR = S / N, sometimes expressed in decibels: SNR dB =10 log10(SNR).

[0286] A high signal-to-noise ratio (SNR) indicates that the signal is significantly stronger than the background noise, making the measurement more reliable; a low SNR indicates that the signal may be overwhelmed by noise, and the measurement results may be unreliable. Generally speaking, an SNR > 10 (10dB) is usually considered the minimum acceptable standard, while in practical applications, a higher standard, such as an SNR > 20 or higher, may be required.

[0287] Based on multi-peak detection and signal-to-noise ratio (SNR) evaluation, a comprehensive judgment is made on the quality of the current measurement. Judgment criteria may include: clear single peak (only one significant peak exists, with no obvious secondary peaks, indicating a single signal path and reliable measurement); multiple peaks but a prominent main peak (secondary peaks exist, but the main peak is clearly dominant, such as a height ratio greater than 5:1, allowing for reliable extraction of the main signal); multiple peaks with similar peak values ​​(multiple peak intensities are similar, making it difficult to determine which is the true signal, potentially leading to unreliable measurement); high SNR (SNR exceeding a preset threshold, such as 20dB, indicating minimal noise impact and more reliable results); and low SNR (SNR below the minimum requirement, suggesting the signal may be masked by noise, resulting in unreliable measurement).

[0288] Next, all the above analysis results are integrated into structured time histogram statistical features, including main peak location parameters (location, height), distribution width parameters (FWHM, standard deviation), multi-peak characteristics (presence of multi-peak and secondary peak characteristics), signal-to-noise ratio (SNR, signal purity), and comprehensive quality assessment. These features are used to optimize the current measurement results and will also be input into the error compensation mapping function to guide the system in selecting the most appropriate processing strategy.

[0289] In summary, this embodiment, through comprehensive time histogram analysis, can extract maximum information from the raw measurement data, identify multiple possible sources of error, provide a scientific basis for subsequent error compensation and data filtering, and ultimately achieve high-precision and high-reliability distance measurement.

[0290] In this embodiment, in step S4, the calibrated measurement results are converted into three-dimensional point cloud data, geometric feature points are extracted from the three-dimensional point cloud data for point cloud registration, and system extrinsic parameters are optimized to establish a system drift monitoring index. When the system drift monitoring index exceeds a preset threshold, the calibration parameters are automatically updated to obtain a self-optimized system parameter set, including: Step S41: Based on the calibrated measurement results and the confidence score, convert them into three-dimensional point cloud data, extract the geometric feature points of corners, edges and planes from the three-dimensional point cloud data, establish a feature descriptor based on local geometric descriptors, and obtain the feature point set and the feature descriptor.

[0291] This step involves converting the calibrated ranging results into 3D point cloud data. For dTOF modules with single-point ranging, this requires a scanning mechanism. Common scanning methods include mechanical rotation (such as rotating a mirror or rotating the entire sensor), MEMS mirror scanning, or solid-state electronic scanning arrays. The scanning mechanism provides orientation information (horizontal and vertical angles) for each distance measurement point, which, together with the distance value, determines the point's 3D coordinates.

[0292] The coordinate transformation follows a mapping from spherical to Cartesian coordinates, calculating the 3D coordinates of each point based on distance and corresponding angle. During the transformation, a confidence score is considered; typically, only points with a confidence score exceeding a preset threshold (e.g., 70% or 80%) are retained to ensure point cloud quality. Points with high confidence scores but below the threshold may be marked as low-confidence points and given lower weight in subsequent processing.

[0293] During point cloud generation, the system also records additional attributes for each point, such as reflectivity (based on received light intensity), timestamp, and original confidence level. This information is helpful for subsequent feature extraction and registration. Reflectivity information is particularly useful, as it can help distinguish surfaces of different materials and improve the accuracy of feature point recognition.

[0294] After generating the point cloud, geometric feature points are extracted from it. These are points in the point cloud that have significant geometric characteristics and are easy to identify and match, mainly including corner points, edge points, and planar feature points. These feature points are usually relatively stable in the environment and are suitable for point cloud registration and pose estimation.

[0295] Corner detection typically targets geometric corners in point clouds, where curvature or principal curvature changes significantly. Common corner detection methods include principal curvature-based methods, Harris 3D corner detection, or SIFT-3D. The system calculates the curvature features within the neighborhood of each point. If the curvature exceeds a threshold and satisfies a local extremum condition, it may be marked as a corner. The choice of neighborhood size is crucial in corner detection; too small a neighborhood may be sensitive to noise, while too large a neighborhood may miss small-scale features.

[0296] Edge points are points located at discontinuities in the contours or surfaces of an object, typically exhibiting abrupt changes in point cloud density or curvature along a certain direction. Edge extraction can be achieved by calculating the eigenvalues ​​of the point's neighborhood covariance matrix; if the largest eigenvalue is significantly larger than other eigenvalues, it indicates that the point is near an edge. Another method is to calculate the change in the normal vector within the point's neighborhood; the normal vector of an edge point usually differs significantly from that of its surrounding points.

[0297] Planar feature points are those points located in flat regions but possessing sufficient discriminative power, typically situated at planar boundaries or in areas with distinctive textures. Planar feature detection can be based on the plane fitting residual within the point's neighborhood, or by combining reflectivity information to find texture features on the plane.

[0298] For example, planar features (such as walls, floors, and tabletops) are widely present in indoor and urban environments and serve as important registration references.

[0299] Typically, after feature point extraction, a local geometric descriptor is created for each feature point, enabling them to be identified and matched across different point clouds. The descriptor should be rotation- and scale-invariant, and also possess some resistance to changes in point cloud density and noise.

[0300] Currently, commonly used local geometry descriptors include Point Feature Histogram (PFH), Fast Point Feature Histogram (FPFH), Spin Image, and Signature Histogram (SHOT). Taking FPFH as an example, it constructs a multi-dimensional histogram to describe the local surface structure by calculating the angular relationships between pairs of points in the neighborhood of a feature point. This description method captures the geometric relationships within the neighborhood, is insensitive to changes in the spatial arrangement of points, and is suitable for point cloud registration.

[0301] Another important descriptor is the Normal Distribution Histogram (NDH), which statistically analyzes the distribution of normal vectors within the neighborhood of a feature point. For planar features, the normal vectors are concentrated; for edges, they are distributed along the two principal directions; and for corners, the distribution is more dispersed. This distribution characteristic helps distinguish between different types of features.

[0302] Typically, different descriptors can be used for different types of feature points. For example, for corner points, a point distribution-based descriptor such as PFH might be used; for planar features, more attention might be paid to normal vectors and curvature information; and for edges, directional information and local shape descriptions might be combined.

[0303] Through the above steps, this embodiment obtains a set of feature points and their corresponding feature descriptors. These feature points typically account for a small portion (approximately 1-5%) of the total point cloud, but contain most of the geometric information, sufficient to support subsequent registration and localization tasks. The system stores these feature points and descriptors in structured data, including the point's 3D coordinates, feature type (corner / edge / plane), descriptor vector, and possible quality scores (based on feature saliency and the confidence level of the original point).

[0304] Step S42: Based on the feature point set and the feature descriptor, the iterative nearest point algorithm is used to register and match the three-dimensional point cloud data collected at different times, calculate the rotation matrix and translation vector, and obtain the point cloud registration result.

[0305] Step S42 requires precisely aligning the point cloud data acquired at different times to assess system stability and calibrate extrinsic parameters. This involves point cloud registration techniques, among which the Iterative Closest Point (ICP) algorithm is a fundamental and effective method.

[0306] Understandably, the goal of point cloud registration is to find an optimal rigid body transformation (rotation matrix R and translation vector t) that minimizes the distance between corresponding points in two point clouds. In dTOF lidar self-calibration applications, the currently acquired point cloud is typically registered with a reference point cloud (such as the point cloud during initial system calibration or the point cloud from the previous moment), and the drift of system parameters is evaluated based on the registration results.

[0307] Currently, the basic process of the standard ICP algorithm includes four main steps: corresponding point search, weighted allocation, transformation estimation, and iterative solution. However, in practical applications, the system usually adopts an improved feature-enhanced ICP algorithm to make full use of the previously extracted feature information and improve registration accuracy and efficiency.

[0308] Typically, the registration process begins with initialization, which involves determining the initial relative poses of the two point clouds. In the case of continuous scanning, the previous registration result or motion estimate can be used as the initial value; for cases without initial poses, the system may use feature matching to obtain a coarse alignment. Feature matching is based on the previously established feature descriptors and uses methods such as nearest neighbor search or RANSAC to find potentially corresponding feature point pairs in the two point clouds.

[0309] In feature matching, for each feature point in the source point cloud (the point cloud to be registered), a search is conducted in the feature point set of the target point cloud (the reference point cloud) for points with the most similar descriptors. Similarity is typically measured using Euclidean distance or cosine similarity between descriptor vectors. To improve matching reliability, a bidirectional matching strategy may be employed, meaning that matches from A to B and from B to A must be identical for a match to be considered valid. Additionally, a similarity threshold may be set to filter out matches with low similarity.

[0310] Specifically, after obtaining the initial correspondence, the iterative optimization phase of ICP begins. In each iteration, based on the currently estimated transformation, the nearest point in the target point cloud is found for each feature point in the source point cloud, forming a corresponding point pair. Unlike standard ICP, feature-enhanced ICP incorporates feature descriptor information, searching for correspondences only between points with similar features, thus avoiding incorrect matching.

[0311] Typically, different weights may be assigned to each pair of corresponding points, based on factors such as the feature quality of the points, descriptor similarity, and spatial distance. For example, corner points usually have higher positioning accuracy and can be given higher weights; corresponding points that are too far apart may be mismatches and should be given lower weights or be eliminated directly.

[0312] Then, the optimal transformation parameters are estimated based on weighted pairs of corresponding points. This is typically achieved by minimizing the sum of weighted squared distances between pairs of corresponding points. The transformation estimate can be quickly computed using closed-form algorithms such as singular value decomposition (SVD) or quaternion methods to obtain the rotation matrix R and translation vector t.

[0313] Typically, the source point cloud position is updated using the newly estimated transformation, and then the next iteration begins, re-finding the corresponding point and optimizing the transformation parameters. The iterative process continues until a certain termination condition is met, such as reaching the maximum number of iterations, the change in transformation parameters being less than a threshold, or the residual improvement being insignificant.

[0314] To improve the robustness and convergence speed of the ICP algorithm, this embodiment can adopt the following improvement strategies: multi-resolution strategy (a coarse-to-fine registration approach, first using downsampled point clouds for fast alignment, and then using the original resolution for fine optimization); point pair filtering (using methods such as statistical outlier removal and consistency checks to filter out possible erroneous matches); local optimization (focusing on registration of feature-rich regions to avoid the dominant role of flat regions in the registration results); and hybrid metric (combining a hybrid objective function of point-to-point distance and point-to-plane distance to better handle planar and edge features).

[0315] In summary, after the ICP iteration, the optimal transformation parameters between the two point clouds are obtained, namely the rotation matrix R and the translation vector t. This transformation represents the relative motion of the system between the two acquisitions (which should be close to a unit transformation if the system is fixed) or the accumulation of errors. Simultaneously, the registration error is also calculated, typically the root mean square distance between corresponding point pairs, as a measure of registration quality.

[0316] It should be noted that the final point cloud registration results include: optimal transformation parameters (R and t), registration error metric, number of effective matching point pairs, and possible registration uncertainty estimates. These results will be used for subsequent extrinsic parameter optimization and system drift monitoring.

[0317] Step S43: Based on the point cloud registration result, the system extrinsic parameters are optimized by maximizing the mutual information between the three-dimensional point cloud data using the continuously acquired three-dimensional point cloud data. The system self-calibration is achieved without the need for a dedicated calibration target, and the optimized system extrinsic parameters are obtained.

[0318] Step S43 optimizes system extrinsic parameters using continuously acquired point cloud data, achieving automatic calibration without the need for a dedicated calibration device. This method is particularly suitable for systems already deployed in real-world environments, allowing for continuous parameter optimization and accuracy maintenance during system operation.

[0319] Typically, system extrinsic parameters mainly include the transformation relationship between the lidar coordinate system and the global reference coordinate system, as well as the relative positions and attitudes of multiple sensor components. For dTOF lidar, extrinsic parameters may include the relative positions of the transmitter and receiver, the installation deviation of the scanning mechanism, and the transformation relationships between sensors in a multi-sensor fusion system. Accurate extrinsic parameters are crucial for correct 3D point cloud generation and multi-sensor data fusion.

[0320] Currently, traditional extrinsic parameter calibration typically requires dedicated calibration targets (such as checkerboard or specially shaped targets), which are complex and require manual intervention. In contrast, self-calibration methods based on mutual information utilize the geometric characteristics of the environment itself to automatically adjust and optimize extrinsic parameters by maximizing the mutual information between continuously acquired point clouds, without the need for dedicated calibration equipment.

[0321] As is understandable, mutual information is an information theory concept used to measure the degree of interdependence between two random variables. In point cloud registration, it measures the structural similarity between two point clouds; their mutual information should be maximized when the two point clouds are perfectly aligned. A major advantage of mutual information is that it can handle data from different modalities (such as data from different sensor types), without requiring the data representation to be completely identical, focusing only on the information structure contained in the data.

[0322] Specifically, the extrinsic parameter optimization process based on mutual information first requires converting point cloud data into a form that can compute mutual information. Currently, a common method is to convert the 3D point cloud into a voxel-based probability distribution or feature distribution. For example, the space can be divided into a regular voxel grid, and then the density, normal vector distribution, or other feature statistics of points within each voxel can be calculated to form a 3D feature distribution map.

[0323] After preparing the data, we set up an optimization problem to find a set of extrinsic parameters that maximizes the mutual information between the point cloud generated based on these parameters and the reference point cloud. The variables in this optimization problem are extrinsic parameter vectors, including rotation and translation components.

[0324] Typically, optimization processes employ iterative methods, such as gradient ascent, Nelder-Mead simplex method, particle swarm optimization, or genetic algorithms. In each iteration, a new set of extrinsic parameter values ​​is tried, the corresponding mutual information is calculated, and then the extrinsic parameters are adjusted according to the rules of the optimization algorithm, moving in the direction that increases mutual information.

[0325] Calculating mutual information requires estimating the joint and marginal distributions of point cloud data. In practice, this is typically achieved through kernel density estimation or histogram methods. For example, a two-dimensional histogram can be constructed, where the two axes correspond to certain features (such as height, normal vector angle, etc.) in two point clouds, and the histogram values ​​represent the frequency of co-occurrence of these two features. Mutual information is then calculated using the histogram, measuring the degree of interdependence between features.

[0326] To improve the stability and efficiency of optimization, the system can employ a multi-scale strategy, first quickly finding an approximate solution at a coarse scale, and then precisely optimizing at a fine scale. Furthermore, the system may introduce a regularization term to avoid overfitting or unreasonable parameter values.

[0327] In practical applications, extrinsic parameter optimization may not be performed after each point cloud acquisition, but rather periodically based on system status assessments. For example, the extrinsic parameter optimization process may be triggered when a significant increase in point cloud registration error is detected, or when the system experiences significant temperature changes or vibrations.

[0328] In summary, after optimization, a set of updated extrinsic parameter values ​​will be obtained. These values ​​can better describe the actual geometric relationships between the various components of the LiDAR, as well as their transformation relationships with the reference coordinate system. The system will evaluate the changes in mutual information and point cloud registration error before and after optimization to determine the effectiveness of the optimization. If the optimization brings significant improvement, the new extrinsic parameters will be adopted and updated in the system configuration; if the improvement is limited or nonexistent, the system may retain the original parameters or conduct more in-depth analysis.

[0329] In step S43, through this self-calibration method that does not require a dedicated calibration target, the dTOF lidar system can continuously maintain and optimize its calibration status in the actual use environment, adapt to environmental changes and system drift, and ensure long-term stable measurement accuracy.

[0330] Step S44: Based on the optimized system extrinsic parameters and historical measurement data, establish a system drift monitoring index system including feature point position offset, point cloud registration accuracy, and measurement consistency. Periodically evaluate the system status, and trigger a recalibration process when the key indicators exceed the preset threshold to obtain the system drift monitoring results; wherein the key indicators include feature point position offset, point cloud registration accuracy, and measurement consistency.

[0331] It should be noted that system drift refers to the deviation of the measurement parameters of a dTOF lidar system from their intended use over time or due to environmental changes. This deviation may be caused by factors such as temperature variations, mechanical stress, aging of electronic components, or environmental interference. To detect and correct these drifts in a timely manner, step S44 establishes a complete system drift monitoring index system.

[0332] First, based on optimized system extrinsic parameters and historical measurement data, several key monitoring indicators are constructed. Feature point position offset is the first key indicator, measuring the degree to which the position of stable feature points in the environment changes over time. The system selects a set of anchor points from historical point cloud data. These points are typically stable geometric features in the environment, such as building corners, edges of specific structures, or feature points on fixed objects that are difficult to move. The system records the precise positions of these anchor points in the reference point cloud and then tracks the positional changes of these points in each new measurement.

[0333] Typically, the calculation of feature point position offset includes the following steps: First, using a feature matching method, find the feature points corresponding to the reference anchor points in the newly acquired point cloud; then, calculate the Euclidean distance between each pair of feature points to obtain a set of offset values; finally, the system may use statistical indicators of these offset values ​​(such as mean, median, standard deviation, etc.) as the overall offset indicator.

[0334] To improve the reliability of feature point tracking, dynamic evaluation of feature point quality can be employed to automatically exclude feature points that become unstable or difficult to identify in new environments. Simultaneously, the system may also discover and add new stable feature points in new scans, maintaining a sufficient number of anchor points.

[0335] It should be noted that point cloud registration accuracy is the second key indicator. It directly measures the quality of the point cloud alignment process and reflects the stability of the system parameters. Registration accuracy is usually expressed as root mean square error (RMSE) or the mean / median of Euclidean distance. These measures the average distance between corresponding point pairs. The smaller the distance, the more accurate the registration and the more stable the system.

[0336] Typically, registration accuracy metrics include several aspects: global RMSE (overall error of all corresponding point pairs), local RMSE (error of a specific region or type of feature), inlier ratio (proportion of corresponding points within the expected error range), and stability of registration transformation parameters (changes in rotation and translation between consecutive registrations). By analyzing these different dimensions of registration accuracy metrics, the system can identify different types of drift patterns.

[0337] Measurement consistency is the third key metric, assessing the stability of a system under repeated measurements under identical conditions. Typically, if the environment and system state remain constant, multiple measurements should produce nearly identical results. The measurement consistency metric is calculated by comparing measurements of the same scene or object at different time points, and can detect random fluctuations or gradual drift in the system.

[0338] Specifically, measurement consistency analysis includes: the standard deviation of static target distance values ​​(assessing measurement stability), the consistency of reflectivity measurements (assessing signal processing stability), the repeatability of feature extraction (the proportion of feature points that can be repeatedly identified in the same scene), and the temperature correlation of measurement results (the degree to which measured values ​​change with temperature). These indicators combined can comprehensively evaluate the stability and reliability of the system.

[0339] The system periodically (possibly every time it is powered on, daily, or after a preset number of measurements) evaluates these metrics and compares them to preset thresholds. These thresholds are typically determined based on the system's design specifications, application requirements, and historical performance data. For example, if the system's design accuracy is ±1 cm, the threshold for feature point position offset might be set to 5 mm; the threshold for point cloud registration RMSE might be set to 3 mm; and the threshold for the standard deviation of measurement consistency might be set to 2 mm.

[0340] It should be noted that when a key indicator exceeds a preset threshold, a recalibration process of the corresponding level will be triggered. Different indicators may correspond to different recalibration responses: minor drift may only require fine-tuning of parameters; moderate drift may require a complete self-calibration process; severe drift may require issuing a warning and recommending manual inspection or maintenance.

[0341] The system also records the results of each assessment, the triggered recalibration operations, and their effects, forming a system drift monitoring record. These records are not only used for current system status assessment but also for long-term trend analysis, predicting potential system failures or aging, and guiding preventative maintenance.

[0342] In summary, through this multi-index, periodically evaluated system drift monitoring mechanism, the dTOF lidar system can maintain long-term stable measurement accuracy, promptly detect and correct possible parameter drift, and improve the system's reliability and service life.

[0343] Step S45: Based on the system drift monitoring results, design a hierarchical recalibration strategy, from parameter adjustment to comprehensive system calibration, automatically select an appropriate calibration level according to the degree of drift, update the system calibration parameter library, establish a parameter self-optimization mechanism to predict parameter change trends, and obtain the self-optimized system parameter set.

[0344] Based on the drift monitoring results obtained above, step S45 adopts an intelligent calibration strategy to automatically correct parameter deviations and establish a parameter change prediction model through long-term learning to achieve continuous optimization of system performance.

[0345] Understandably, a tiered recalibration strategy is a resource-efficient approach. It employs different levels of calibration based on the nature and extent of drift, avoiding unnecessary full calibration and saving system resources and time. Typically, three to four calibration levels are designed, from light to heavy: parameter fine-tuning, partial recalibration, full system calibration, and hardware diagnostics.

[0346] Among these, parameter fine-tuning is the lightest calibration method, suitable for minor system drift. When a small drift is detected, the system compensates through simple parameter adjustments. This typically involves minor corrections to external parameters (such as fine-tuning rotation angles or translation vectors) or updating temperature compensation parameters. Parameter fine-tuning can usually be completed within seconds without significantly disrupting normal system operation.

[0347] Partial recalibration targets moderate system drift, which may involve a specific subsystem or a particular type of parameter. For example, if the main drift is found to originate from temperature-related parameters, the system will specifically recalibrate the temperature model; if the drift is mainly manifested within a specific distance range, the parameters within that range will be calibrated. Partial recalibration typically requires more data acquisition and processing, and may take several minutes, but can still be completed during normal system operation.

[0348] It's important to note that comprehensive system calibration is a more thorough process, suitable for situations where multiple metrics deteriorate simultaneously or parameters change significantly. It re-evaluates and optimizes all system calibration parameters, including baseline error calibration parameters, multi-dimensional error compensation model parameters, and extrinsic parameters. Comprehensive calibration may require specific procedures, such as guiding the system to scan a feature-rich environment or conducting tests under different conditions. This process can take anywhere from ten to several tens of minutes and is typically performed during system downtime or dedicated maintenance periods.

[0349] Hardware diagnostics is the highest level of response, triggered when a serious anomaly is detected and software calibration cannot effectively resolve the issue. This may indicate a hardware problem, such as component aging, optical system contamination, or mechanical deformation. The system will perform a series of hardware self-tests to attempt to locate the problem and may issue an alert recommending manual inspection or maintenance.

[0350] This embodiment can automatically select an appropriate calibration level based on drift monitoring results, typically based on a decision tree or rule engine, considering the overall situation of multiple indicators. For example, if only the point cloud registration accuracy is slightly reduced but other indicators are normal, parameter fine-tuning may be selected; if the feature point position shift is significant and has a clear directionality, partial recalibration may be performed; if multiple indicators deteriorate simultaneously, full calibration may be required.

[0351] Typically, after the calibration process is completed, the calibration parameter library, a structured parameter storage system, is updated, containing calibration parameter sets from different periods and conditions. Usually, new parameters generated during each calibration are stored along with timestamps, environmental conditions, and system status information, forming a historical record of the parameters. Therefore, multiple versions of parameter sets may be retained, automatically selecting the most suitable set under different conditions, or reverting to a previous stable version if new parameters are ineffective.

[0352] It should be noted that the parameter self-optimization mechanism is a long-term learning and prediction system that builds a predictive model of parameter change trends based on historical data in a parameter database. This mechanism may employ time series analysis, regression models, or machine learning methods to analyze the changing patterns of parameters with factors such as time, temperature, and usage intensity. For example, the system may discover that certain parameters exhibit seasonal change patterns or show a linear decay trend with usage time.

[0353] Understandably, based on this predictive trend, this embodiment can achieve predictive calibration, that is, proactively adjusting parameters before they reach critical values ​​to avoid performance degradation. Simultaneously, the system may also identify abnormal parameter change patterns, detecting potential problems in advance, such as signs of impending component failure.

[0354] Typically, parameter self-optimization also includes parameter sensitivity analysis to assess the impact of different parameters on system performance, focusing on the critical parameters with the greatest impact and optimizing calibration resource allocation. For example, if it is found that temperature compensation parameters have the most significant impact on accuracy, the system may increase the monitoring and optimization frequency of these parameters.

[0355] Based on this, this embodiment generates a self-optimized system parameter set, which is a dynamically updated set of parameters that can adapt to changes in system state and environmental conditions, maintaining optimal performance. Compared with the static parameter set, the self-optimized parameter set has better adaptability and stability, enabling the dTOF lidar system to operate stably for a long time under various conditions.

[0356] In this embodiment, before acquiring the real-time laser pulse time-of-flight data and environmental parameters, the method further includes near-range blind zone elimination and photon accumulation effect suppression steps, including: Step S301: Obtain a preliminary estimate of the target distance. Based on detector gating technology, dynamically adjust the detector's receiving window opening time and duration according to the preliminary estimate. Simultaneously control the laser pulse width to be variably adjustable within the range of 50ps to 5ns. When the preliminary estimate is less than a preset near-range threshold, adjust the laser pulse width to 50ps to 500ps and set the receiving window delay to 0.5 nanoseconds to 2 nanoseconds. When the preliminary estimate is greater than a preset far-range threshold, adjust the laser pulse width to 1ns to 5ns and set the receiving window delay to 10 nanoseconds to 50 nanoseconds. Obtain the adaptively adjusted laser pulse parameters and receiving window parameters.

[0357] Those skilled in the art will know that the near-range blind zone problem and the long-range measurement accuracy problem are the main challenges of dTOF systems.

[0358] In this step, the system first obtains a preliminary estimate of the target distance, which may come from a variety of sources: previous measurement results (in the case of continuous measurement), a coarse prediction model (based on the target's movement trend), auxiliary sensors (such as low-precision but full-range ranging sensors), or preliminary results from multiple rapid scans.

[0359] In cases of initial startup or lack of prior information, the system may employ a multi-mode measurement strategy, configuring parameters suitable for different distance ranges, and then determining the most probable distance interval from the results.

[0360] Typically, the accuracy requirement for the initial target distance estimation is not high; it is sufficient to determine the approximate distance range of the target (e.g., near, medium, or far). This rough estimate is adequate to support subsequent adaptive parameter adjustments.

[0361] Based on the initial distance estimate, detector gating technology is used to dynamically adjust the receiver window parameters. Detector gating achieves selective detection of optical signals arriving at specific times by precisely controlling the detector's sensitive time window. In dTOF systems, this is typically achieved by controlling the enable signal or bias voltage of the SPAD detector.

[0362] Currently, the receiving window is mainly controlled by two parameters: the opening time (the delay relative to the laser emission time) and the duration (the length of time the window remains open). These two parameters directly determine the range of distances the system can detect; the opening time corresponds to the minimum measurable distance, and the opening time plus the duration corresponds to the maximum measurable distance. By adjusting these two parameters, the system can focus on a specific distance range, improving signal quality and reducing interference.

[0363] Typically, in conjunction with the adjustment of the receiving window, the system also controls the laser pulse width to be variably adjusted within the range of 50 ps to 5 ns. Pulse width is a key parameter affecting the range resolution and maximum measurement range of a dTOF system. Narrow pulses (e.g., 50 ps-500 ps) provide higher range resolution, suitable for precise short-range measurements; wide pulses (e.g., 1 ns-5 ns) provide more photons, enhancing signal strength, and are suitable for long-range measurements. The system can dynamically adjust the pulse width according to the target distance to optimize measurement performance.

[0364] Specifically, when the initial estimated value is less than the preset near-range threshold (typically set to 1-3 meters), a parameter configuration optimized for near-range operation is adopted: the laser pulse width is adjusted to a narrow pulse of 50 ps to 500 ps, ​​while the receiving window delay is set to a shorter 0.5 nanoseconds to 2 nanoseconds. The narrow pulse provides higher temporal (distance) resolution, enabling the differentiation of subtle distance changes at close range; the short-delay receiving window can capture early return signals reflected from near-range targets, solving the near-range blind zone problem of traditional dTOF systems; understandably, an appropriate receiving window delay can also avoid direct interference from the transmitted pulse, improving the signal-to-noise ratio.

[0365] Typically, when the initial estimated value exceeds the preset long-distance threshold (usually set to 20-50 meters, depending on system design specifications), switch to long-distance optimized configuration: adjust the laser pulse width to a wide pulse of 1ns to 5ns, while setting the receiver window delay to a longer delay of 10 nanoseconds to 50 nanoseconds. The wide pulse contains more light energy, enhancing the long-distance signal strength and increasing the probability of detecting long-distance targets; the longer receiver window delay matches the time-of-flight of the long-distance optical signal, ensuring that the receiver window is open when the long-distance signal returns.

[0366] For intermediate distance ranges (between the near and far distance thresholds), the system can employ transitional configurations such as medium pulse width (500ps-1ns) and medium window delay.

[0367] In summary, this parameter adaptive adjustment mechanism based on preliminary distance estimation maintains optimal performance across different distance ranges, effectively eliminating near-range blind spots while maximizing long-range measurement capabilities. The system will then apply these adaptively adjusted laser pulse parameters and receiver window parameters to subsequent real-time measurement processes.

[0368] Step S302: Based on the adaptively adjusted laser pulse parameters and the receiving window parameters, the flight time data and received light intensity information corresponding to different pulse widths are fused, and the light intensity distribution characteristics within the receiving window are combined with the flight time data through an intensity-assisted distance judgment algorithm to obtain the full-range distance measurement capability.

[0369] In this step, the previously adaptively adjusted parameters are applied to the actual measurement, and through the intelligent fusion of multiple measurement results, a high-precision distance measurement capability covering the entire range is achieved.

[0370] It should be noted that, based on the adaptively adjusted laser pulse parameters and receiver window parameters, the system performs real-time measurements to acquire time-of-flight data and received light intensity information under different parameter configurations. For applications requiring high-precision measurements across the entire range, the system can employ a multi-mode measurement strategy, i.e., continuously measuring using multiple parameter configurations within a short period (typically on the order of milliseconds), and then fusing these results.

[0371] Specifically, the multi-mode measurement includes a near-range mode (narrow pulse, short delay window), a mid-range mode (medium pulse width, medium delay window), and a long-range mode (wide pulse, long delay window). Each mode is optimized for a specific distance range to produce the most accurate measurement results within that range.

[0372] Next, time-of-flight data and corresponding received light intensity information were collected under different modes. It is understandable that time-of-flight data directly reflects the round-trip time of the optical signal, while received light intensity indirectly reflects the target distance and reflection characteristics. Light intensity decreases inversely with the square of the distance; therefore, under the same target reflectivity conditions, light intensity can serve as a supplementary indicator of distance.

[0373] It should be noted that multi-mode data fusion is a key technology for the system to achieve full-range measurement. The fusion process first requires normalization and calibration of data from different modes to ensure data comparability.

[0374] For example, different pulse widths result in different received signal energies, and the system needs to normalize the light intensity according to the pulse width. Similarly, different receiver window configurations may introduce different system biases into time-of-flight measurements, which need to be compensated for through calibration parameters.

[0375] Secondly, after normalization and calibration, an intensity-assisted distance determination algorithm is applied for data fusion. This algorithm combines time and intensity measurements to improve the reliability and accuracy of distance estimation. Specifically, its purpose is to provide additional distance cues based on light intensity distribution characteristics in situations of measurement uncertainty, helping to resolve ambiguities or improve accuracy.

[0376] Typically, intensity-assisted distance estimation algorithms involve the following steps: comparing time-of-flight measurements under different modes to check for consistency (considering systematic biases between modes); analyzing the light intensity distribution characteristics within the receiving window, such as peak position, shape, and width, which reflect the relative position of the light pulse within the receiving window and indirectly provide distance information; and combining time-of-flight measurements and light intensity distribution characteristics to estimate the most probable true distance using statistical models or machine learning algorithms.

[0377] In cases of inconsistency between modes, the system can select the most reliable mode result based on the light intensity level and measurement confidence level. For example, if the target is at close range, the close-range mode will typically produce a stronger return signal and a higher confidence level. In boundary cases (such as when the target is located at the boundary between two modes), a weighted fusion strategy may be used to perform a weighted average of the results based on the relative confidence levels of each mode.

[0378] For example, for particularly complex scenarios (such as multiple reflections or the presence of transparent objects), it may be possible to analyze the complete time histogram and light intensity distribution, identify multiple possible targets, and report complex target structures.

[0379] In summary, this embodiment, through its multi-mode measurement and intensity-assisted distance judgment algorithm, effectively fuses measurement results under different parameter configurations, overcoming the limitations of a single mode and achieving full-range high-precision measurement capabilities covering distances from near (e.g., less than 0.5 meters) to far (e.g., over 100 meters). This embodiment provides the fused final distance estimate and its confidence level as output to subsequent point cloud generation and processing modules. This full-range measurement capability enables dTOF lidar to adapt to various complex environments and application scenarios, providing more accurate and reliable distance information for both close-range precision operations and long-range obstacle detection.

[0380] Step S303: Based on the full-range distance measurement capability, for high reflectivity targets identified in the full-range distance measurement capability whose received light intensity is greater than a preset high reflectivity threshold, a multi-zone detector array is used to achieve multi-channel parallel sampling, dispersing the photon arrival time received by a single detector to multiple detection channels, thereby obtaining a dispersed photon arrival time distribution.

[0381] Those skilled in the art will know that photon piling-up is a common problem in the measurement of high-reflectivity targets. Specifically, photon piling-up is a phenomenon that occurs when a SPAD detector receives high-intensity light signals, which can cause the ranging result to be premature, affecting the measurement accuracy.

[0382] In this step, the system first identifies targets with received light intensity exceeding a preset high reflectivity threshold, based on the previously established full-range distance measurement capability. These high reflectivity targets may include reflective materials (such as traffic signs and reflective stickers), metal surfaces, glass, or water surfaces—objects with specular reflective properties. Typically, the high reflectivity threshold is set based on the system's optical characteristics and the performance of the SPAD detector, and may be set to 3-5 times the normal reflectivity or a specific value close to the SPAD saturation level.

[0383] Specifically, when a high-reflectivity target is detected, a processing mechanism is triggered, employing a multi-zone detector array to achieve multi-channel parallel sampling. This multi-zone detector array is a special SPAD detector design that divides the photosensitive area into multiple physically isolated but optically synchronized sub-regions, each connected to an independent electronic readout channel. This design allows photons incident on the detector to be distributed across multiple independent detection areas, with each area recording the photon arrival time separately.

[0384] Specifically, the working principle of a multi-region detector array (SPAD) is based on the spatial distribution characteristics of photons. The incident beam typically covers the entire detector area, with photons randomly distributed across various sub-regions. Each sub-region operates independently; when a photon is detected, its arrival time is recorded, and the array enters a quenching period (dead time, the period during which the SPAD cannot detect the next photon after detecting one). Because each sub-region is independent, one region entering the quenching period does not affect other regions from continuing to detect photons.

[0385] Based on this, the above design effectively solves the photon stacking effect problem. The photon stacking effect is essentially caused by the quenching characteristics of SPADs. When multiple photons arrive almost simultaneously, only the first photon is recorded, and subsequent photons are not detected during the quenching period. This causes the timing record to be biased towards the earliest arriving photon, resulting in a pre-acceleration bias. The multi-zone array design, by distributing photons across multiple independent channels, significantly reduces the probability of a single channel receiving multiple photons, thereby mitigating the stacking effect.

[0386] In practical applications, 4x4, 8x8, or larger SPAD arrays can be used, with the appropriate size selected based on application requirements and hardware resources. Each SPAD unit is connected to an independent Time-to-Digital Converter (TDC) channel or shares a TDC but has independent trigger control. These channels operate in parallel, simultaneously recording photon arrival times.

[0387] This embodiment uses this multi-channel parallel sampling mechanism to obtain a distributed photon arrival time distribution. This distribution contains photon time records from all channels and typically contains more complete information than single-channel sampling, especially information about subsequent photons that would be blocked by the quenching period in a single-channel design.

[0388] Specifically, the decentralized photon arrival time distribution has the following important characteristics: the leading edge of the distribution (the earliest arriving photon) is similar to that of a single channel, but the main body of the distribution is more complete and closer to the true photon arrival time distribution; the recording of subsequent photons makes the shape of the distribution more accurate, which helps to correctly identify the true peak position; the statistical results of multiple channels improve the sampling efficiency, and more photon information can be obtained in the same measurement time, thus improving the signal-to-noise ratio.

[0389] In summary, while this multi-zone detector design may increase hardware complexity and cost, it is suitable for applications requiring high precision, especially those dealing with high-reflectivity targets. In cost-constrained applications, the multi-zone detection function may only be activated when a high-reflectivity target is detected, balancing performance and resource consumption. This embodiment integrates photon arrival time data collected from multiple channels to form a more complete and accurate time distribution record, providing a foundation for subsequent stacking effect correction.

[0390] Step S304: Based on the dispersed photon arrival time distribution, a corrected histogram is constructed by statistically counting the photons in each channel. The statistical histogram correction algorithm is applied to identify the characteristic peak of the accumulation effect in the dispersed photon arrival time distribution where the peak position is earlier than the actual flight time, and the time offset is calculated for compensation and correction. The compensated and corrected flight time data is used as the flight time data of the real-time laser pulse to obtain the ranging result after suppressing the photon accumulation effect.

[0391] It should be noted that this step is the core of suppressing the photon stacking effect. Through statistical analysis and algorithm correction, the time measurement deviation caused by the stacking effect is compensated to obtain more accurate time-of-flight data.

[0392] Based on the previously obtained distributed photon arrival time distribution, the photon counts for each channel are first calculated to construct a corrected histogram. The corrected histogram is a statistical representation of the original photon arrival times, but unlike the standard histogram, it integrates data from multiple channels, containing more complete photon timing information.

[0393] In this step, the process of constructing the corrected histogram includes the following steps: time alignment (due to possible slight time deviations between different channels, the system first precisely aligns the data of each channel to ensure they reference the same time base); count normalization (there may be efficiency differences in SPADs of different channels, and the system normalizes the counts of each channel according to pre-calibrated efficiency parameters); time binning (the system groups the aligned and normalized photon arrival time data by time interval to form a time histogram); multi-channel merging (the system merges the time histogram data of all channels to form a unified corrected histogram); and background noise filtering (the system may apply filtering techniques to remove random background noise and improve the clarity of signal peaks).

[0394] Typically, a well-constructed corrected histogram provides a comprehensive view of photon arrival times. However, in the case of high-reflectivity targets, this histogram may still exhibit a stacking effect, where the peak position is earlier than the actual flight time. This is because even with multi-channel dispersion, high-intensity signals can still cause each channel to record a bias towards the earliest arriving photon.

[0395] This embodiment applies a statistical histogram correction algorithm to identify and compensate for this stacking effect. Understandably, the core of the correction algorithm is to identify the characteristics of the stacking effect and estimate the true peak location. This typically includes the following methods: theoretical model fitting (the system uses a theoretical distribution model of photon arrival time to fit the observed histogram, recovering the true parameters from the distorted observed distribution); control group analysis (possibly measuring the same target at different intensities simultaneously, establishing a model of the relationship between intensity and time offset by comparing the measurement results at low and high intensities); statistical moment correction (analyzing the statistical moments of the histogram and estimating the offset caused by the stacking effect based on the variation patterns of these moments); and machine learning methods (possibly using a pre-trained model to predict the degree of the stacking effect based on histogram features).

[0396] Using the method described above, the time offset caused by the stacking effect can be calculated. This offset typically increases with signal strength, potentially reaching hundreds of picoseconds in extreme cases, corresponding to a distance error of tens of centimeters. The system applies the calculated offset to the original time-of-flight data for compensation and correction. This correction can be a simple time addition (adding the estimated offset back to the measured flight time) or a more complex model correction that considers signal strength, SPAD characteristics, and environmental factors.

[0397] In summary, the corrected time-of-flight data can effectively eliminate the influence of photon stacking effect and provide a more accurate target distance estimate. This corrected data can be used as the time-of-flight data of real-time laser pulses for subsequent distance calculation and point cloud generation.

[0398] Ultimately, the range measurement results after suppressing the photon stacking effect are output, achieving high-precision measurement in the case of high reflectivity targets. This enables the dTOF lidar to accurately measure targets with various surface characteristics, from dark objects with low reflectivity to metals or reflective materials with high reflectivity, significantly improving the system's adaptability and reliability.

[0399] The laser emitter, reference detector, main detector, and time-to-digital converter used in this embodiment will be described below.

[0400] In this embodiment, the laser emitter is a VCSEL laser emitter, the reference detector and the main detector are both SPAD detectors, and the time-to-digital converter is a TDC time-to-digital converter.

[0401] In practical implementation, a VCSEL (Vertical-Cavity Surface-Emitting Laser) is used as the laser emission source, which has advantages such as low power consumption, narrow divergence angle, and high modulation bandwidth. A SPAD (Single-Photon Avalanche Diode) is selected as both the reference and main detectors, offering single-photon-level sensitivity and higher response speed. A TDC (Time-to-Digital Converter) is used to achieve high-precision time measurement, with a time resolution of less than 10 ps, ​​supporting millimeter-level distance measurement accuracy.

[0402] Example 2: Figure 2 A schematic diagram of the ranging calibration system based on dTOF lidar in this embodiment is shown, as follows: Figure 2 As shown, this embodiment also provides a ranging calibration system based on dTOF lidar, the system comprising: The zero-meter reference calibration module 201 is used to acquire the laser pulse signal emitted by the laser emitter, divide the laser pulse signal into a reference optical path and a measurement optical path through an optical beam splitter, receive the laser pulse signal of the reference optical path through a reference detector and record the reference time point, receive the laser pulse signal reflected by the target through the measurement optical path through a main detector and record the reception time point, calculate the time difference between the reference time point and the reception time point and statistically analyze the zero-meter reference measurement data, and obtain the system reference error calibration parameter set by combining the quantization characteristics analysis of the time-to-digital converter. Error modeling module 202 is used to establish a multi-dimensional error compensation mathematical model based on the system reference error calibration parameter set, by collecting calibration data under different measurement conditions, including temperature model, reflectivity model, multipath effect model and ambient light compensation model, and to construct a combination of basis functions for the multi-dimensional error compensation mathematical model by discretization strategy and solve it by optimization algorithm to obtain the error compensation mapping function. The real-time calibration module 203 is used to acquire real-time laser pulse time-of-flight data and environmental parameters, input the environmental parameters into the error compensation mapping function to calculate the error compensation value, apply the error compensation value to the time-of-flight data and process it through a filter to obtain the calibrated measurement result; The self-optimization module 204 is used to convert the calibrated measurement results into three-dimensional point cloud data, extract geometric feature points from the three-dimensional point cloud data for point cloud registration and optimize system extrinsic parameters, establish system drift monitoring indicators, and automatically update calibration parameters when the system drift monitoring indicators exceed a preset threshold, thereby obtaining a self-optimized system parameter set.

[0403] The system in this embodiment includes four main functional modules: the zero-meter reference calibration module 201 is responsible for establishing the system reference error calibration parameter set; the error modeling module 202 is responsible for constructing a multi-dimensional error compensation mathematical model and generating an error compensation mapping function; the real-time calibration module 203 is responsible for real-time calibration and filtering of the measurement data; and the self-optimization module 204 is responsible for monitoring system drift and automatically updating the calibration parameters.

[0404] It should be noted that each functional module of the system corresponds one-to-one with the corresponding steps in the aforementioned method embodiment 1. Through the coordinated work of the above modules, this embodiment achieves high-precision and high-reliability dTOF lidar ranging calibration.

[0405] The above description is merely a specific implementation of the embodiments of this disclosure, but the protection scope of the embodiments of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this disclosure should be included within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the protection scope of the claims.

Claims

1. A ranging calibration method based on dTOF lidar, characterized in that, include: The laser pulse signal emitted by the laser emitter is acquired, and the laser pulse signal is divided into a reference optical path and a measurement optical path by an optical beam splitter. The laser pulse signal of the reference optical path is received by a reference detector and a reference time point is recorded. The laser pulse signal reflected by the target in the measurement optical path is received by a main detector and the reception time point is recorded. The time difference between the reference time point and the reception time point is calculated, and the zero-meter reference measurement data is statistically analyzed. Combined with the quantization characteristics analysis of the time-to-digital converter, the system reference error calibration parameter set is obtained. Based on the system's reference error calibration parameter set, calibration data is collected under different measurement conditions to establish a multi-dimensional error compensation mathematical model that includes a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is discretized to construct a combination of basis functions and solved by an optimization algorithm to obtain the error compensation mapping function. The time-of-flight data and environmental parameters of the real-time laser pulse are acquired. The environmental parameters are input into the error compensation mapping function to calculate the error compensation value. The error compensation value is applied to the time-of-flight data and processed by a filter to obtain the calibrated measurement result. The calibrated measurement results are converted into three-dimensional point cloud data. Geometric feature points are extracted from the three-dimensional point cloud data for point cloud registration and system extrinsic parameters are optimized. System drift monitoring indicators are established. When the system drift monitoring indicators exceed a preset threshold, the calibration parameters are automatically updated to obtain a self-optimized system parameter set.

2. The method according to claim 1, characterized in that, The laser pulse signal is split into a reference optical path and a measurement optical path using an optical beam splitter. A reference detector receives the laser pulse signal from the reference optical path and records the reference time point. A main detector receives the laser pulse signal reflected from the target in the measurement optical path and records the reception time point. The time difference between the reference time point and the reception time point is calculated, and zero-meter reference measurement data is statistically analyzed. Combined with the quantization characteristics analysis of the time-to-digital converter, a system reference error calibration parameter set is obtained, including: The laser pulse signal emitted by the laser emitter is divided into a reference optical path and a measurement optical path by an optical beam splitter, and a zero-meter reference optical path system is established to obtain the zero-meter reference optical path configuration. Based on the zero-meter reference optical path configuration, after receiving the laser pulse signal of the reference optical path through the reference detector, a time counter is triggered to record the reference time point. After receiving the laser pulse signal returned by the measurement optical path through the target through the main detector, a time counter is triggered to record the receiving time point. The time difference between the reference time point and the receiving time point is calculated to eliminate the influence of the inherent delay of the laser driving circuit and obtain the time difference after eliminating the driving circuit delay. Based on the time difference after eliminating the driving circuit delay, the zero-meter reference measurement data is statistically analyzed to construct a reference detector response time histogram. The mean and standard deviation of the reference detector response time histogram are analyzed to obtain the response time error model, quantify the receiver circuit delay, and obtain the receiver circuit delay parameters. Based on the delay parameters of the receiving end circuit, the quantization characteristics of the time-to-digital converter are analyzed. The quantization error is controlled within a preset range through internal oscillator frequency calibration and interpolation algorithm to obtain quantization error compensation parameters. Based on the zero-meter reference measurement data, the receiver circuit delay parameters, and the quantization error compensation parameters, a comprehensive system error calibration parameter set is calculated to obtain the system reference error calibration parameter set.

3. The method according to claim 1, characterized in that, Based on the system's reference error calibration parameter set, a multi-dimensional error compensation mathematical model is established by collecting calibration data under different measurement conditions. This model includes a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is then discretized to construct a combination of basis functions, which are solved using an optimization algorithm to obtain the error compensation mapping function, including: Based on the system reference error calibration parameter set, zero-meter reference calibration data and standard reflector ranging data were collected under different temperature conditions and different reflectivity conditions. Combining the quantization error compensation parameters and receiver circuit delay parameters in the system reference error calibration parameter set, a functional relationship model between temperature and time delay and a mapping relationship between reflection intensity and time offset were established to obtain temperature-time delay model parameters and reflectivity-time offset model parameters. Based on the temperature-time delay model parameters and the reflectivity-time offset model parameters, the time histogram characteristics of the received signal are analyzed and the multi-peak distribution is identified. A peak separation algorithm is designed to distinguish between direct reflection signals and secondary reflection signals. The background count rate of the detector is monitored and an adaptive threshold algorithm is designed to maintain the signal-to-noise ratio within a preset signal-to-noise ratio threshold range. Thus, the multipath effect detection model parameters and the ambient light compensation model parameters are obtained. Based on the temperature-time delay model parameters, the reflectivity-time offset model parameters, the multipath effect detection model parameters, and the ambient light compensation model parameters, the temperature-time delay model, reflectivity-time offset model, multipath effect detection model, and ambient light compensation model are integrated into a multi-dimensional error compensation mathematical model. A ranging error compensation function including temperature, reflection intensity, target distance, and background light intensity is constructed to obtain a multi-dimensional calibration mathematical model. Based on the multidimensional calibration mathematical model, the function space of the multidimensional calibration mathematical model is adaptively partitioned. Different types of basis functions are constructed to form a heterogeneous basis function set for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters. An adaptive mesh refinement algorithm is designed based on the heterogeneous basis function set, and the optimized mesh distribution is obtained through iterative optimization. Based on the optimized mesh distribution, the conjugate gradient method and multigrid technique are used to solve the linear equation system generated by the discretization of the segmented polynomial to obtain the error compensation mapping function.

4. The method according to claim 3, characterized in that, The method involves constructing different types of basis functions to form a heterogeneous basis function set, for the temperature error corresponding to the temperature-time delay model parameters, the reflectivity error corresponding to the reflectivity-time offset model parameters, and the distance error corresponding to the multipath effect detection model parameters, including: Based on the temperature-time delay model parameters, the physical characteristics of the temperature error are analyzed, and second- to third-order polynomial basis functions are used to characterize the smooth change relationship between temperature and time delay, thus obtaining the low-order polynomial basis function of the temperature error. Based on the reflectivity-time offset model parameters and the low-order polynomial basis function of the temperature error, the piecewise characteristics of the reflectivity error are analyzed. A piecewise continuous smooth function is constructed in different reflectivity intervals using cubic spline basis functions, while ensuring the continuity of the first and second derivatives, to obtain the piecewise spline basis function of the reflectivity error. Based on the parameters of the multipath effect detection model, the low-order polynomial basis function of the temperature error, and the piecewise spline basis function of the reflectivity error, the local characteristics of the distance error are analyzed, and the Gaussian radial basis function is used to characterize the local response characteristics of the distance-related error, thus obtaining the Gaussian radial basis function of the distance error. Based on the low-order polynomial basis function of temperature error, the piecewise spline basis function of reflectivity error, and the Gaussian radial basis function of distance error, a composite basis function is constructed and the weighting coefficients are calculated to obtain the heterogeneous basis function set.

5. The method according to claim 1, characterized in that, The process of acquiring real-time laser pulse time-of-flight data and environmental parameters, inputting the environmental parameters into the error compensation mapping function to calculate the error compensation value, applying the error compensation value to the time-of-flight data and processing it through a filter to obtain the calibrated measurement result includes: The emission and reception times of each laser pulse are obtained, the original flight time is calculated, and the received light intensity, current ambient temperature and background light intensity are recorded to obtain the original measurement data set. Based on the original measurement data set, a real-time time histogram is constructed from the flight time data of one hundred to one thousand consecutive measurements in the original measurement data set. The main peak position and distribution width of the real-time time histogram are analyzed to obtain the statistical characteristics of the time histogram. Based on the time histogram statistical characteristics, the received light intensity, current ambient temperature, and background light intensity from the original measurement data set, along with the time histogram statistical characteristics, are input into the error compensation mapping function to calculate the corresponding error compensation value, thereby obtaining the real-time error compensation parameters. Based on the real-time error compensation parameters, the original flight time in the original measurement data set is corrected by applying the real-time error compensation parameters. At the same time, an adaptive Kalman filter is designed based on the statistical characteristics of the time histogram to smooth the measurement results after error correction, reducing random fluctuations while retaining rapidly changing effective information, and obtaining filtered measurement data. Based on the filtered measurement data, the calibrated flight time and distance values ​​are calculated, and the confidence score is calculated based on the statistical characteristics of the time histogram to obtain the calibrated measurement results and the confidence score.

6. The method according to claim 5, characterized in that, A real-time time histogram is constructed from the flight time data of 100 to 1000 consecutive measurements in the original measurement dataset. The main peak position and distribution width of the real-time time histogram are analyzed to obtain the statistical characteristics of the time histogram, including: Obtain flight time data from one hundred to one thousand consecutive laser pulse measurements in the original measurement data set, set the time resolution as the time interval width of the histogram, and obtain the time histogram construction parameters. Based on the time histogram construction parameters, the flight time data is divided into corresponding time intervals, the count value of each time interval is counted, a real-time time histogram is constructed, and an initial time histogram is obtained. Based on the initial time histogram, the position of the main peak in the initial time histogram is identified by a peak detection algorithm, the count value corresponding to the main peak position is calculated, the main peak position is determined, and the main peak position parameter is obtained. Based on the main peak position parameter, locate the data distribution area around the main peak position in the initial time histogram, analyze the peak shape characteristics around the main peak position by the full width at half maximum (FWHM) method or standard deviation calculation, calculate the distribution width around the main peak, and obtain the distribution width parameter. Based on the main peak position parameter and the distribution width parameter, the presence of a multi-peak distribution is identified in the initial time histogram. The signal-to-noise ratio is evaluated based on the count value corresponding to the main peak position parameter and the background noise count value of the initial time histogram to determine the measurement quality and obtain the statistical characteristics of the time histogram.

7. The method according to claim 1, characterized in that, The process involves converting the calibrated measurement results into three-dimensional point cloud data, extracting geometric feature points from the three-dimensional point cloud data for point cloud registration, optimizing system extrinsic parameters, establishing system drift monitoring indicators, and automatically updating calibration parameters when the system drift monitoring indicators exceed a preset threshold, resulting in a self-optimized system parameter set, including: Based on the calibrated measurement results and the confidence score, the data is converted into three-dimensional point cloud data. Geometric feature points of corners, edges and planes are extracted from the three-dimensional point cloud data. A feature descriptor based on local geometric descriptors is established to obtain a set of feature points and the feature descriptor. Based on the feature point set and the feature descriptor, the iterative nearest point algorithm is used to register and match the three-dimensional point cloud data collected at different times, calculate the rotation matrix and translation vector, and obtain the point cloud registration result. Based on the point cloud registration results, the system extrinsic parameters are optimized by maximizing the mutual information between the three-dimensional point cloud data using continuously acquired three-dimensional point cloud data. This enables system self-calibration without the need for a dedicated calibration target, resulting in optimized system extrinsic parameters. Based on the optimized system extrinsic parameters and historical measurement data, a system drift monitoring index system is established, including feature point position offset, point cloud registration accuracy, and measurement consistency. The system status is evaluated periodically, and a recalibration process is triggered when key indicators exceed preset thresholds to obtain system drift monitoring results. The key indicators include feature point position offset, point cloud registration accuracy, and measurement consistency. Based on the system drift monitoring results, a hierarchical recalibration strategy is designed, from parameter adjustment to comprehensive system calibration. The appropriate calibration level is automatically selected according to the degree of drift, the system calibration parameter library is updated, a parameter self-optimization mechanism is established to predict the parameter change trend, and the self-optimized system parameter set is obtained.

8. The method according to claim 1, characterized in that, Before acquiring the real-time laser pulse time-of-flight data and environmental parameters, the process also includes near-range blind zone elimination and photon accumulation effect suppression steps, including: A preliminary estimate of the target distance is obtained. Based on detector gating technology, the opening time and duration of the detector's receiving window are dynamically adjusted according to the preliminary estimate. The laser pulse width is simultaneously controlled to be variably adjustable within the range of 50 ps to 5 ns. When the preliminary estimate is less than a preset near-range threshold, the laser pulse width is adjusted to 50 ps to 500 ps and the receiving window delay is set to 0.5 nanoseconds to 2 nanoseconds. When the preliminary estimate is greater than a preset far-range threshold, the laser pulse width is adjusted to 1 ns to 5 ns and the receiving window delay is set to 10 nanoseconds to 50 nanoseconds. The adaptively adjusted laser pulse parameters and receiving window parameters are obtained. Based on the adaptively adjusted laser pulse parameters and the receiving window parameters, the flight time data and received light intensity information corresponding to different pulse widths are fused, and the light intensity distribution characteristics within the receiving window are combined with the flight time data through an intensity-assisted distance judgment algorithm to obtain full-range distance measurement capability. Based on the full-range distance measurement capability, for high reflectivity targets identified in the full-range distance measurement capability whose received light intensity is greater than a preset high reflectivity threshold, a multi-zone detector array is used to achieve multi-channel parallel sampling, dispersing the photon arrival time received by a single detector to multiple detection channels, thereby obtaining a dispersed photon arrival time distribution. Based on the dispersed photon arrival time distribution, a corrected histogram is constructed by statistically counting the photons in each channel. The statistical histogram correction algorithm is applied to identify the characteristic peak of the accumulation effect in the dispersed photon arrival time distribution where the peak position is earlier than the actual flight time, and the time offset is calculated for compensation and correction. The compensated and corrected flight time data is used as the flight time data of the real-time laser pulse to obtain the ranging result after suppressing the photon accumulation effect.

9. The method according to claim 1, characterized in that, The laser emitter is a VCSEL laser emitter, the reference detector and the main detector are both SPAD detectors, and the time-to-digital converter is a TDC time-to-digital converter.

10. A ranging calibration system based on dTOF lidar, characterized in that, include: The zero-meter reference calibration module is used to acquire the laser pulse signal emitted by the laser emitter, split the laser pulse signal into a reference optical path and a measurement optical path through an optical beam splitter, receive the laser pulse signal of the reference optical path through a reference detector and record the reference time point, receive the laser pulse signal reflected by the target through the measurement optical path through a main detector and record the reception time point, calculate the time difference between the reference time point and the reception time point and statistically analyze the zero-meter reference measurement data, and obtain the system reference error calibration parameter set by combining the quantization characteristics analysis of the time-to-digital converter. The error modeling module is used to establish a multi-dimensional error compensation mathematical model based on the system's reference error calibration parameter set, by collecting calibration data under different measurement conditions, including a temperature model, a reflectivity model, a multipath effect model, and an ambient light compensation model. The multi-dimensional error compensation mathematical model is discretized to construct a combination of basis functions and solved by an optimization algorithm to obtain the error compensation mapping function. The real-time calibration module is used to acquire real-time laser pulse time-of-flight data and environmental parameters, input the environmental parameters into the error compensation mapping function to calculate the error compensation value, apply the error compensation value to the time-of-flight data and process it through a filter to obtain the calibrated measurement result; The self-optimization module is used to convert the calibrated measurement results into three-dimensional point cloud data, extract geometric feature points from the three-dimensional point cloud data for point cloud registration and optimize system extrinsic parameters, establish system drift monitoring indicators, and automatically update calibration parameters when the system drift monitoring indicators exceed a preset threshold, thereby obtaining a self-optimized system parameter set.

Images