A short pulse driving method for a matrix-addressed lidar

Abstract

The application relates to a kind of short pulse driving methods of laser radar based on matrix addressing, specifically to the technical field of laser radar, the scheme is by constructing initial crosstalk matrix and adopting regularized non-negative matrix factorization algorithm to separate clean echo signal and crosstalk component, and then construct crosstalk risk tensor and global timing cost function, use the greedy initialization to combine simulated annealing fine tuning optimization lighting sequence and silence period, while the deconvolution residual is fed back to step S1 to update the crosstalk matrix by recursive least squares and form a closed loop feedback;Without increasing hardware overhead, the optical crosstalk and ghost signal are effectively identified and inhibited, the false points in the point cloud are significantly reduced, the ranging accuracy and target contour clarity are improved, the accuracy of obstacle identification and obstacle avoidance decision is ensured, while the frame rate and scanning uniformity are considered, and reliable operation of high-performance solid-state laser radar is realized.

Description

Technical Field

[0001] This invention relates to the field of lidar technology, and more specifically, to a short-pulse driving method for lidar based on matrix addressing. Background Technology

[0002] With the rapid development of autonomous driving, intelligent perception, and 3D imaging technologies, solid-state lidar, as a core sensor device, is evolving towards higher frame rates, longer detection distances, and higher spatial resolution. Among them, dot-matrix solid-state lidar based on two-dimensional VCSEL laser arrays achieves independent control of each emitting unit through matrix addressing, enabling spatial beam dot-matrix scanning in an extremely short time. In high frame rate mode, the driving circuit needs to rapidly cycle and illuminate different row and column combinations in the 32×16 array with nanosecond-level pulse widths and tens of ampere peak currents to achieve millions of ranging samples per second. However, this high-density, high-frequency dot-matrix emission method places extremely stringent requirements on the system's signal purity and timing accuracy. Especially in high-speed autonomous driving test scenarios, lidar needs to acquire accurate distance information of surrounding objects in real time in complex traffic environments, and any tiny signal interference may lead to deviations in the perception results.

[0003] Existing lidar driving technologies face significant crosstalk problems when dealing with high-density dot array emission. On the one hand, when the driving circuit rapidly switches between different row and column combinations, optical crosstalk exists between adjacent VCSEL units. That is, the laser emitted by one unit is scattered by the array substrate or packaging structure and misinterpreted by the detector as the echo signal of another unit. On the other hand, in nanosecond-level pulse emission mode, the residual echo of the previous pulse may overlap with the echo of the subsequent pulse in time, generating ghosting signals. The combined effect of these two types of crosstalk results in a large number of false points in the point cloud data, leading to blurred target contours, increased ranging errors, and in severe cases, directly affecting the accuracy of obstacle recognition and obstacle avoidance decisions, posing a safety hazard. Currently, the industry mainly suppresses crosstalk by adding physical isolation structures or using time-division multiplexing, but these methods either increase manufacturing costs and process complexity or sacrifice scanning frame rate and spatial resolution. Therefore, there is an urgent need for a technical solution that can identify and suppress crosstalk through intelligent timing arrangement and signal processing without increasing hardware overhead. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a short-pulse driving method for lidar based on matrix addressing, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps:

[0006] Step S1: The main control unit controls the horizontal drive module and the vertical drive module to light up each VCSEL light-emitting unit in the array one by one according to the predetermined test pattern, collects the original echo tensor of all detector channels, normalizes the response of each channel, and constructs an initial crosstalk matrix characterizing the optical and electromagnetic coupling strength between each unit.

[0007] Step S2: The main control unit models the original echo tensor acquired in normal working mode as a superposition of clean echo signal, crosstalk component and noise, and uses the crosstalk matrix as physical constraint. It then uses a regularized non-negative matrix factorization algorithm to deconvolve the original echo tensor to separate the clean echo signal, crosstalk component, signal-to-noise ratio of each channel and deconvolution residual.

[0008] Step S3: The main control unit constructs a crosstalk risk tensor based on the clean echo signal, crosstalk components and signal-to-noise ratio of each channel, and constructs a global time-series cost function based on the crosstalk risk tensor, which includes spatial crosstalk term, time decoupling term, scan integrity term and power consumption equalization term.

[0009] Step S4: The main control unit uses the global temporal cost function as the objective and employs a hybrid strategy of greedy initialization combined with simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame. The optimized lighting order and silent period are then written into the horizontal driving module and the vertical driving module. At the same time, the deconvolution residual output in step S2 is fed back to step S1, which is used to recursively update the crosstalk matrix with least squares each time step S1 is executed, so as to obtain the crosstalk matrix with real-time calibration and form a closed-loop feedback.

[0010] In a preferred embodiment, in step S1, when the main control unit constructs the initial crosstalk matrix, for each VCSEL light-emitting unit, the detector channel corresponding to the position of the VCSEL light-emitting unit is taken as the main channel, the echo signal of the main channel is taken as the reference template, and the amplitude square coherence function of the echo signal of the remaining detector channels and the reference template within the effective frequency band preset by the laser pulse spectrum is calculated.

[0011] Then, the amplitude squared coherence function is integrated over the effective frequency band, and the integration result is divided by the integral value of the self-coherence function of the main channel to obtain the off-diagonal elements of the initial crosstalk matrix. The integral value of the self-coherence function of the main channel is always equal to the effective frequency band width, thus ensuring that the diagonal elements of the initial crosstalk matrix are 1, and the off-diagonal elements represent the relative coherence coupling strength between each channel, and the values ​​are between 0 and 1.

[0012] In a preferred embodiment, in step S1, when the main control unit acquires the original echo tensor, it sequentially applies a single short pulse drive to each VCSEL light-emitting unit and synchronously records the time-domain echo signals of all M detector channels at T time sampling points to form a three-dimensional original echo tensor.

[0013] Then, for each VCSEL light-emitting unit, the time-domain echo signal is extracted from the main channel as a reference template, and the cross power spectral density of the time-domain echo signal of each other channel with the reference template and their respective self-power spectral density are calculated, thereby obtaining the amplitude squared coherence function.

[0014] The lower and upper limits of the effective frequency band are preset by the laser pulse spectrum to ensure that only the frequency components with concentrated signal energy are retained, and to eliminate the interference of out-of-band noise on the coupling strength estimation.

[0015] In a preferred embodiment, in step S2, the cost function optimized by the regularized nonnegative matrix factorization algorithm used by the main control unit consists of three terms:

[0016] The first item is the data fidelity item, which measures the degree of difference between the original echo tensor and the reconstructed echo tensor. The reconstructed echo tensor is composed of the crosstalk matrix multiplied by the convolution result of the clean echo signal and the standard pulse waveform template, plus the crosstalk component.

[0017] The second term is the total variation regularization term, which is used to constrain the sum of the absolute values ​​of the gradients of the clean echo signal in the time direction, so as to suppress the oscillation artifacts generated by deconvolution and preserve the rising edge of the echo.

[0018] The third term is the timing overlap penalty term, which is used to square the energy of the VCSEL light-emitting unit corresponding to the main channel and its neighboring VCSEL light-emitting units determined by the crosstalk matrix within the theoretical echo arrival time window according to the lighting order of the current frame, so as to avoid crosstalk energy being incorrectly allocated to the clean echo signal.

[0019] In a preferred embodiment, in step S2, the main control unit uses the alternating direction multiplier method to iteratively solve the cost function. The iterative process includes:

[0020] Initialize the clean echo signal and crosstalk components to be zero matrices, and initialize the Lagrange multipliers to be zero matrices;

[0021] In each iteration, the crosstalk components and Lagrange multipliers are first fixed, and the clean echo signal is updated using the near-end gradient descent method, where the step size of the near-end gradient descent method is set to a preset value and the number of inner loop iterations is fixed; then the updated clean echo signal and Lagrange multipliers are fixed, and the crosstalk components are updated using the closed-form solution, which includes rectified linear unit functions to ensure the non-negativity of the crosstalk components; then the Lagrange multipliers are updated.

[0022] Repeat the iteration until the relative change between two adjacent cost function values ​​is less than a preset threshold or the maximum number of iterations is reached. At this point, the final clean echo signal, crosstalk component, signal-to-noise ratio of each channel, and deconvolution residual are output.

[0023] In a preferred embodiment, the specific method by which the main control unit constructs the crosstalk risk tensor in step S3 is as follows:

[0024] For any two VCSEL emitting units, the main control unit traverses all detector channels, multiplies the relative coupling strength of the first VCSEL emitting unit to the current channel in the crosstalk matrix with the relative coupling strength of the second VCSEL emitting unit to the current channel, divides by the signal-to-noise ratio of the current channel and adds a preset positive number to obtain the contribution value of the current channel.

[0025] Then, the contribution values ​​of all channels are summed and multiplied by a spatial attenuation factor. The spatial attenuation factor is determined by the physical distance between the two VCSEL light-emitting units in the array and follows a Gaussian attenuation law. Finally, the crosstalk risk tensor element values ​​corresponding to the two VCSEL light-emitting units are obtained in the crosstalk risk tensor.

[0026] In a preferred embodiment, the specific method by which the master control unit constructs the global temporal cost function in step S3 is as follows:

[0027] The spatial crosstalk term is obtained by summing the crosstalk risk tensor element values ​​of all adjacent unit pairs in the lighting sequence of the next frame;

[0028] The time decoupling term is obtained by substituting the deviation between the actual silence period of each adjacent unit pair and the minimum time interval required for theoretical crosstalk attenuation into the Huber loss function and multiplying it by a preset first weighting coefficient, and then summing them up. The minimum time interval required for theoretical crosstalk attenuation is determined by the maximum half-life of the crosstalk component output in step S2 on the set of detector channels that are jointly affected.

[0029] The scan integrity term is obtained by multiplying the scan ratio of each sub-region of the current frame by the KL divergence of an ideally uniform distribution by a preset second weighting coefficient;

[0030] The power consumption equalization term is obtained by multiplying the sum of the squares of all quiet periods by a preset third weighting coefficient; the sum of the four terms yields the global timing cost function.

[0031] In a preferred embodiment, in step S4, the main control unit uses a hybrid strategy of greedy initialization combined with simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame in the following way:

[0032] First, using the crosstalk risk tensor as the distance matrix, the initial lighting order is constructed using the nearest neighbor heuristic. That is, a starting point is randomly selected from all VCSEL light-emitting units to be lit in the next frame. Then, the unselected VCSEL light-emitting unit with the smallest crosstalk risk tensor element value with the current last VCSEL light-emitting unit is selected as the next lighting unit. All silent periods are initialized to the minimum time interval required for theoretical crosstalk attenuation plus a preset offset.

[0033] Then, in the simulated annealing fine-tuning stage, the initial temperature, the termination temperature, and the cooling rate are set. A fixed number of Markov chain iterations are performed at each temperature. In each iteration, a 2-opt subsequence reversal operation is randomly selected for the lighting order or a random replacement operation is performed for the silent period. The decision to accept a new solution is made based on the exponential function of the cost increment and the current temperature. The process is repeated after cooling until the termination temperature is reached.

[0034] In a preferred embodiment, in step S4, the main control unit also performs hard constraint verification and penalty function backfeeding during the simulated annealing fine-tuning process:

[0035] After each new solution is received, the total frame time, which is the sum of all silent periods, is checked to see if it exceeds the preset upper limit. If it does, a frame rate penalty term proportional to the square of the excess is added to the global temporal cost function.

[0036] At the same time, it checks whether the scan ratio of each sub-region and the KL divergence of the ideal uniform distribution exceed the preset threshold. If they exceed the threshold, a coverage uniformity penalty term proportional to the square of the excess is added to the global temporal cost function.

[0037] The cost function with the added penalty term is used for acceptance judgment in subsequent simulated annealing.

[0038] In a preferred embodiment, in step S4, the main control unit feeds back the deconvolution residual output in step S2 to step S1 for recursive least squares update of the crosstalk matrix in the following specific way:

[0039] Calculate the product of the deconvolution residual matrix and the transpose matrix of the clean echo signal, then divide by the square of the Frobenius norm of the clean echo signal plus a preset positive number to obtain the initial update amount;

[0040] The initial update amount is multiplied by a preset learning rate, and then multiplied element-wise by a confidence weighted matrix to obtain the final update amount.

[0041] Add the final update value to the current crosstalk matrix to obtain the updated crosstalk matrix;

[0042] Each element of the confidence weighting matrix is ​​determined by the crosstalk risk tensor element values ​​of the two corresponding VCSEL emitting units in the crosstalk risk tensor. The element values ​​of the confidence weighting matrix are independent of the detector channel number. The VCSEL emitting unit with higher crosstalk risk has a smaller weight.

[0043] The beneficial effects of this invention are as follows: By constructing an initial crosstalk matrix and using a regularized nonnegative matrix factorization algorithm to separate the clean echo signal from the crosstalk component, a crosstalk risk tensor and a global temporal cost function are constructed. Greedy initialization combined with simulated annealing is used to fine-tune and optimize the lighting sequence and quiet period. At the same time, the deconvolution residual is fed back to step S1 to recursively update the crosstalk matrix using least squares to form a closed-loop feedback. Without increasing hardware overhead, it effectively identifies and suppresses optical crosstalk and ghost signals, significantly reduces false points in the point cloud, improves ranging accuracy and target contour clarity, ensures the accuracy of obstacle recognition and obstacle avoidance decisions, and simultaneously takes into account frame rate and scanning uniformity, thus achieving reliable operation of high-performance solid-state lidar. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0047] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0048] Example 1: This example provides the following... Figure 1 The method for driving short pulses of a lidar based on matrix addressing is shown, and specifically includes the following steps:

[0049] Step S1: The main control unit controls the horizontal drive module and the vertical drive module to light up each VCSEL light-emitting unit in the array one by one according to the predetermined test pattern, collects the original echo tensor of all detector channels, normalizes the response of each channel, and constructs an initial crosstalk matrix characterizing the optical and electromagnetic coupling strength between each unit.

[0050] Step S2: The main control unit models the original echo tensor acquired in normal working mode as a superposition of clean echo signal, crosstalk component and noise, and uses the crosstalk matrix as a physical constraint. The initial crosstalk matrix is ​​the initial crosstalk matrix constructed in step S1, and is subsequently updated by step S4 to the crosstalk matrix calibrated in real time. The original echo tensor is deconvolved using a regularized nonnegative matrix factorization algorithm to separate the clean echo signal, crosstalk component, signal-to-noise ratio of each channel and deconvolution residual; where the deconvolution residual is the difference between the original echo tensor and the reconstructed echo tensor.

[0051] Step S3: The main control unit constructs a crosstalk risk tensor based on the clean echo signal, crosstalk components and signal-to-noise ratio of each channel, and constructs a global time-series cost function based on the crosstalk risk tensor, which includes spatial crosstalk term, time decoupling term, scan integrity term and power consumption equalization term.

[0052] Step S4: The main control unit uses the global temporal cost function as the objective and employs a hybrid strategy of greedy initialization combined with simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame. The optimized lighting order and silent period are then written into the horizontal driving module and the vertical driving module. At the same time, the deconvolution residual output in step S2 is fed back to step S1, which is used to recursively update the crosstalk matrix with least squares each time step S1 is executed, so as to obtain the crosstalk matrix with real-time calibration and form a closed-loop feedback.

[0053] In this embodiment, it is specifically necessary to explain that in step S1, when the main control unit constructs the initial crosstalk matrix, for each VCSEL emitting unit, the detector channel corresponding to the position of the VCSEL emitting unit is taken as the main channel, and the echo signal of the main channel is taken as the reference template. The amplitude square coherence function of the echo signals of the other detector channels and the reference template within the effective frequency band preset by the laser pulse spectrum is calculated. The amplitude square coherence function is used to measure the linear correlation between two signals at each frequency component. The specific calculation process of the amplitude square coherence function is as follows: First, for the reference template... The echo signals from the reference template and the channel to be compared are subjected to Fast Fourier Transform (FFT) to obtain their respective complex frequency domain representations. Then, the cross-power spectral density (CPS) of the two is calculated, which is the square of the modulus of the conjugate of the reference template frequency domain signal multiplied by the frequency domain signal of the channel to be compared. Next, the auto-power spectral density (ASPD) of the reference template and the channel to be compared is calculated, which is the square of the modulus of their respective frequency domain signals. Finally, the cross-power spectral density is divided by the product of the auto-power spectral densities to obtain the amplitude squared coherence function value at each frequency point. This function value is between 0 and 1, where 1 indicates complete linear correlation and 0 indicates complete uncorrelation.

[0054] Then, the amplitude-squared coherence function is integrated over the effective bandwidth, and the result is divided by the integral value of the self-coherence function of the main channel to obtain the off-diagonal elements of the initial crosstalk matrix. The integral value of the self-coherence function of the main channel is always equal to the effective bandwidth, thus ensuring that the diagonal elements of the initial crosstalk matrix are 1. The off-diagonal elements represent the relative coherence coupling strength between channels, and their values ​​range between 0 and 1. The effective bandwidth is preset using the following method: based on the time-domain width τp of the laser pulse (in nanoseconds), its spectral main lobe width is determined by Fourier transform. Typically, the lower limit of the effective bandwidth is 0.5 divided by τp, and the upper limit is 3 divided by τp. For example, when the pulse width is 7 nanoseconds... At that time, the effective frequency band has a lower limit of approximately 71.4 MHz and an upper limit of approximately 428.6 MHz. This frequency band covers more than 90% of the pulse energy, while eliminating low-frequency DC drift and high-frequency noise. The integration process uses the trapezoidal numerical integration method, which divides the effective frequency band into 1024 sub-intervals and accumulates the product of the coherence function value and the frequency step size in each sub-interval. The self-coherence function of the main channel is always 1 at all frequency points, so its integral value is the effective frequency band width, for example, 357.2 MHz. The diagonal elements of the initial crosstalk matrix are set to 1, and the off-diagonal elements are the integration ratios mentioned above, finally resulting in a 512-row, 512-column real number matrix, with each element between 0 and 1.

[0055] In step S1, when the main control unit acquires the raw echo tensor, it sequentially applies a single short pulse drive to each VCSEL emitting unit, synchronously recording the time-domain echo signals of all M detector channels at T time sampling points to form a three-dimensional raw echo tensor. The specific acquisition process of the three-dimensional raw echo tensor is as follows: the main control unit selects VCSEL emitting units one by one according to the predetermined test pattern through the horizontal drive module and the vertical drive module; when each emitting unit is lit, the drive circuit applies a constant short pulse with a peak current of 10 amperes and a pulse width of 7 nanoseconds; all M detector channels (M equals 512) start sampling simultaneously, the sampling clock frequency is 2 GHz, each channel records T sampling points, T is 256, corresponding to a 128 nanosecond observation window; the sampling start time is synchronized with the pulse emission time, with a delay of no more than 0.5 nanoseconds; the acquired data is stored in the buffer in the form of a three-dimensional array, the first dimension is the emitting unit number, the second dimension is the detector channel number, and the third dimension is the time sampling point number;

[0056] Then, for each VCSEL emitting unit, the time-domain echo signal is extracted from the main channel as a reference template, and the cross-power spectral density and the auto-power spectral density of the time-domain echo signals of other channels with the reference template are calculated, thus obtaining the amplitude squared coherence function. The specific calculation method of cross-power spectral density and auto-power spectral density is as follows: for each emitting unit i, the sequence of 256 time-domain sampling points of its main channel (i.e., the i-th detector channel) is used as the reference template xi(t); for any other channel j, its time-domain sequence is xj(t); first, Hanning windowing is applied to xi(t) and xj(t) to reduce spectral leakage, and then 2 The frequency domain sequences Xi(f) and Xj(f) are obtained by 56-point Fast Fourier Transform. The cross-power spectral density Pij(f) is calculated as follows: Pij(f) is equal to the conjugate of Xi(f) multiplied by the square of the magnitude of Xj(f); the auto-power spectral density Pii(f) is equal to the square of the magnitude of Xi(f), and Pjj(f) is equal to the square of the magnitude of Xj(f); the amplitude squared coherence function Cij(f) is equal to the square of Pij(f) divided by the product of Pii(f) and Pjj(f); to ensure statistical stability, the same emitting unit is measured 16 times, and the cross-power spectral density and auto-power spectral density obtained from the 16 measurements are averaged before the coherence function is calculated.

[0057] The lower and upper limits of the effective frequency band are preset by the laser pulse spectrum to ensure that only the frequency components with concentrated signal energy are retained, eliminating the interference of out-of-band noise on the coupling strength estimation. The specific settings for the lower and upper limits of the effective frequency band are: the lower limit is set to 50 MHz and the upper limit is set to 500 MHz. This range is determined based on the spectral characteristics of the 7 nanosecond pulse, where the main lobe width is approximately 143 MHz and the sidelobe energy is concentrated within 500 MHz. The frequency components below 50 MHz mainly come from detector dark current drift and low-frequency flicker noise, while the components above 500 MHz are mainly high-frequency thermal noise and amplifier noise. By integrating within this frequency band, the signal-to-noise ratio of the crosstalk matrix estimation can be improved by approximately 12 dB, effectively suppressing the contamination of coupling strength calculation by incoherent interference.

[0058] In this embodiment, it should be specifically noted that in step S2, the cost function optimized by the regularized nonnegative matrix factorization algorithm used by the main control unit consists of three terms:

[0059] The first term is the data fidelity term, which measures the difference between the original echo tensor and the reconstructed echo tensor. The reconstructed echo tensor is constructed by multiplying the crosstalk matrix by the convolution result of the clean echo signal and the standard pulse waveform template, and then adding the crosstalk component. The specific calculation process of the data fidelity term is as follows: First, each row of the clean echo signal is convolved with the standard pulse waveform template. The convolution kernel length is 14 sampling points, corresponding to the waveform of a 7 nanosecond pulse at a sampling rate of 2 GHz. After convolution, it is truncated to 256 sampling points to obtain the convolution result matrix. Then, the convolution result matrix is ​​left-multiplied by the crosstalk matrix to obtain an intermediate matrix of 512 rows and 256 columns. Finally, the intermediate matrix is ​​added to the crosstalk component to obtain the reconstructed echo tensor. The value of the data fidelity term is equal to half the sum of the squares of each element of the difference between the original echo tensor and the reconstructed echo tensor. This difference matrix is ​​called the reconstruction error matrix.

[0060] The second term is the total variation regularization term, which is used to constrain the sum of the absolute values ​​of the gradients of the clean echo signal in the time direction to suppress the oscillation artifacts generated by deconvolution and preserve the rising edge of the echo. The specific calculation process of the total variation regularization term is as follows: For each row of the clean echo signal, that is, the time-domain sequence corresponding to each VCSEL emitting unit, the difference between adjacent time sampling points is calculated, and the absolute values ​​of all differences are summed; the sum of the absolute values ​​of all 512 rows is accumulated and then multiplied by the first regularization parameter; the preset value of the first regularization parameter is 0.01, which is determined by cross-validation, that is, on the offline test dataset, with the goal of maximizing the signal-to-noise ratio of the point cloud, 10 candidate values ​​are selected in the range of 0.001 to 0.1 at logarithmic intervals, and the parameter that makes the validation set performance optimal is selected;

[0061] The third term is the timing overlap penalty term, which is used to square the energy within the theoretical echo arrival time window of the VCSEL emitting unit corresponding to the main channel and its neighboring VCSEL emitting units determined by the crosstalk matrix, based on the lighting order of the current frame. This is to prevent crosstalk energy from being incorrectly allocated to the clean echo signal. The specific calculation process of the timing overlap penalty term is as follows: First, based on the lighting order of the current frame, determine the theoretical echo arrival time of each lit VCSEL emitting unit. The theoretical echo arrival time is equal to the pulse emission time plus the estimated value of the optical flight time, where the optical flight time is based on the minimum and maximum detection ranges of the lidar. The distance is determined, typically ranging from 10 nanoseconds to 200 nanoseconds; then, for each detector channel, the VCSEL emitting unit corresponding to its main channel and the adjacent VCSEL emitting units with coupling strength greater than 0.1 are determined according to the crosstalk matrix; the theoretical echo arrival time of these units is extended by 5 sampling points before and after (corresponding to a 2.5 nanosecond time window) to form a time window set; finally, the values ​​of the elements in the crosstalk component located within these time windows are squared and summed, and then multiplied by the second regularization parameter; the preset value of the second regularization parameter is 0.05, and its selection method is similar to that of the first regularization parameter, determined in the range of 0.01 to 0.1 through cross-validation;

[0062] In step S2, the main control unit uses the alternating direction multiplier method to iteratively solve the cost function. The iterative process includes:

[0063] The clean echo signal and crosstalk components are initialized to zero matrices, and the Lagrange multipliers are also initialized to zero matrices. Specifically, the clean echo signal is initialized to a 512-row, 256-column all-zero matrix; the crosstalk components are initialized to a 512-row, 256-column all-zero matrix; the Lagrange multipliers are initialized to a 512-row, 256-column all-zero matrix; and the penalty parameter is initialized to 0.1.

[0064] In each iteration, the crosstalk components and Lagrange multipliers are fixed, and the clean echo signal is updated using the proximal gradient descent method. The step size of the proximal gradient descent method is set to a preset value, and the number of iterations in the inner loop is fixed. The specific steps of updating the clean echo signal using the proximal gradient descent method are as follows: First, the gradient of the cost function with respect to the clean echo signal under the current clean echo signal is calculated. This gradient consists of three parts: the derivative of the data fidelity term with respect to the clean echo signal, the derivative of the total variational regularization term with respect to the clean echo signal, and the Lagrange multipliers. The derivative of the data fidelity term with respect to the clean echo signal is calculated. The derivative of the data fidelity term is obtained by the inverse convolution operation of the transpose of the crosstalk matrix and the reconstruction error matrix. The derivative of the total variation regularization term is obtained by calculating the second difference of the clean echo signal in the time direction. Then, the clean echo signal is updated along the negative gradient direction with a step size of 0.01. Next, a non-negativity constraint is applied to the updated clean echo signal, setting all negative values ​​to zero. The above update process is repeated 20 times as the inner loop to obtain the updated clean echo signal value for this iteration.

[0065] After fixing the updated clean echo signal and Lagrange multipliers, the crosstalk components are updated using a closed-form solution. The closed-form solution includes a rectified linear unit function to ensure the non-negativity of the crosstalk components. Then, the Lagrange multipliers are updated. The specific calculation process for updating the crosstalk components using the closed-form solution is as follows: First, a temporary matrix is ​​calculated, which is equal to the original echo tensor minus the matrix resulting from the convolution of the crosstalk matrix and the clean echo signal with the standard pulse waveform template, plus the Lagrange multipliers divided by the penalty parameter. Then, each element of the temporary matrix is ​​divided by 1, and the corresponding element of the time window indicator function matrix is ​​multiplied by the penalty parameter, plus twice the second regularization parameter. The time window indicator function matrix is ​​a 512-row, 256-column matrix, with a value of 1 for positions belonging to the time window set and 0 otherwise. Finally, the rectified linear unit function is applied to the division result to set all negative values ​​to zero, resulting in the updated crosstalk components. The update formula for the Lagrange multipliers is: the new Lagrange multipliers are equal to the old Lagrange multipliers plus the penalty parameter multiplied by the reconstruction error matrix.

[0066] The iteration is repeated until the relative change of the cost function values ​​between two adjacent iterations is less than a preset threshold or the maximum number of iterations is reached. At this point, the final clean echo signal, crosstalk component, signal-to-noise ratio of each channel, and deconvolution residual are output. The specific convergence criteria are set as follows: the relative change of the cost function values ​​between two adjacent iterations is less than 10 to the power of negative 4, or the number of iterations reaches 100. The relative change of the cost function value is defined as the absolute value of the cost function value of the current iteration minus the cost function value of the previous iteration divided by the cost function value of the previous iteration. When either condition is met, the iteration stops. The final output deconvolution residual is equal to the original echo tensor minus the matrix multiplied by the crosstalk matrix and the convolution result of the clean echo signal with the standard pulse waveform template, minus the crosstalk component.

[0067] The signal-to-noise ratio (SNR) of each channel is calculated as follows: For each detector channel, calculate the energy of the clean echo signal on that channel (the sum of the squares of all time sampling points), divide it by the energy of the crosstalk component on that channel, add the noise variance estimate, and then take the logarithm to the base 10 and multiply by 10; the noise variance estimate is the square of the Frobenius norm of the deconvolution residual divided by 512 and multiplied by 256.

[0068] In this embodiment, it is specifically necessary to explain the method by which the main control unit constructs the crosstalk risk tensor in step S3:

[0069] For any two VCSEL emitting units, the main control unit traverses all detector channels, multiplies the relative coupling strength of the first VCSEL emitting unit to the current channel in the crosstalk matrix with the relative coupling strength of the second VCSEL emitting unit to the current channel, divides by the signal-to-noise ratio of the current channel, and adds a preset positive number to obtain the contribution value of the current channel; the preset positive number is 10 to the power of negative 6, which is used to prevent the denominator from overflowing when the signal-to-noise ratio is zero or extremely small; the signal-to-noise ratio of the current channel is provided by the signal-to-noise ratio vector of each channel output in step S2, in decibels, and the value range is usually between 0 and 40 dB;

[0070] Then, the contribution values ​​of all channels are summed and multiplied by a spatial attenuation factor. The spatial attenuation factor is determined by the physical distance between the two VCSEL emitting units in the array and follows a Gaussian attenuation law. This yields the crosstalk risk tensor element values ​​corresponding to the two VCSEL emitting units in the crosstalk risk tensor. The specific calculation method for the spatial attenuation factor is as follows: First, calculate the physical Euclidean distance between the two VCSEL emitting units in the array, in pixels; the array layout is 32 rows and 16 columns, with both the horizontal and vertical spacing between adjacent units being 250 micrometers; then, calculate the square of this distance divided by 2. The square of the spatial decay constant is multiplied by the sum of the squares of the squares. The spatial decay constant is preset to be 2 pixel spacing, corresponding to 500 micrometers. Finally, the natural exponential function is taken for the negative ratio to obtain the spatial decay factor. The value of this factor ranges from 0 to 1, with a value of 1 when the distance is zero and decaying according to a Gaussian curve as the distance increases. The sum of the contribution values ​​of all channels is multiplied by this factor to obtain the crosstalk risk tensor element values. This tensor is a 512-row, 512-column real number matrix. The diagonal elements are the crosstalk risk values ​​of the same cell with itself. However, since the same cell will not generate crosstalk with itself in actual applications, the diagonal elements are set to zero in subsequent calculations.

[0071] In step S3, the master control unit constructs the global time-series cost function in the following way:

[0072] The spatial crosstalk term is obtained by summing the crosstalk risk tensor element values ​​of all adjacent unit pairs in the lighting sequence of the next frame. The calculation process of the spatial crosstalk term is as follows: Let the number of light-emitting units to be lit in the next frame be L, and arrange them in the lighting order as the sequence π1, π2, ..., πL; from u equal to 1 to L minus 1, take out the element value of the πu-th row and π(u+1)-th column of the crosstalk risk tensor in turn, and add all these element values ​​to obtain the value of the spatial crosstalk term; the physical meaning of this term is to penalize adjacent strongly coupled unit pairs that are lit, and the larger the value, the higher the crosstalk risk.

[0073] The time decoupling term is obtained by substituting the deviation between the actual silence period of each adjacent unit pair and the minimum time interval required for theoretical crosstalk attenuation into the Huber loss function and multiplying it by a preset first weighting coefficient, and then summing the results. The minimum time interval required for theoretical crosstalk attenuation is determined by the maximum half-life of the crosstalk component output in step S2 on the set of detector channels that jointly influence it. The specific calculation method for the minimum time interval required for theoretical crosstalk attenuation is as follows: For two adjacent light-emitting units πu and π(u+1), first determine the set of detector channels that they jointly influence, i.e., those in the crosstalk matrix that simultaneously satisfy a coupling strength greater than 0.1 for unit πu and a coupling strength greater than 0.1 for the crosstalk component. All detector channels with a coupling strength greater than 0.1 in unit π(u+1) are identified. Then, the time-domain waveforms of these channels are extracted from the crosstalk component matrix output in step S2. For each channel, the time required for the crosstalk component to decay from its peak to half its original value is found; this is the half-life. The maximum value among the half-lives of all commonly affected channels is taken as the minimum time interval required for theoretical crosstalk attenuation. This value is typically between 2 and 8 nanoseconds. The inflection point kappa of the Huber loss function is preset to 1 nanosecond. When the absolute value of the deviation is less than or equal to 1 nanosecond, squared loss is used; when it is greater than 1 nanosecond, linear loss is used. The first weighting coefficient is preset to 0.2.

[0074] The scan integrity factor is obtained by multiplying the scan ratio of each sub-region of the current frame by the KL divergence of the ideal uniform distribution, and a preset second weighting coefficient. The sub-regions are divided as follows: the 32-row, 16-column VCSEL array is uniformly divided into 64 sub-regions, each containing 8 light-emitting units in 4 rows and 2 columns. The scan ratio p_g of each sub-region of the current frame is equal to the number of lit light-emitting units in that sub-region divided by the total number of lit units L in the current frame. The ideal uniform distribution ratio q_g is always equal to 1 / 64. The KL divergence is calculated as follows: for each sub-region g, p_g is multiplied by the logarithm of p_g divided by q_g with base e, and then the results of all sub-regions are summed. The second weighting coefficient is preset to 0.1.

[0075] The power consumption balancing term is obtained by multiplying the sum of the squares of all quiet periods by a preset third weighting coefficient; the four terms are added together to obtain the global timing cost function; the third weighting coefficient is preset to 0.005; the unit of quiet period δ_u is nanoseconds, and its sum of squares is used to penalize excessively long quiet periods to prevent excessive frame rate drop; the final global timing cost function is the sum of the above four terms, which serves as the optimization target for greedy initialization combined with simulated annealing fine-tuning in step S4.

[0076] In this embodiment, it is specifically necessary to explain the method by which the main control unit uses a hybrid strategy of greedy initialization combined with simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame in step S4:

[0077] First, using the crosstalk risk tensor as the distance matrix, the initial lighting order is constructed using the nearest neighbor heuristic. That is, a starting point is randomly selected from all VCSEL light-emitting units to be lit in the next frame. Then, the unselected VCSEL light-emitting unit with the smallest crosstalk risk tensor element value with the current last VCSEL light-emitting unit is selected as the next lighting unit. All silence periods are initialized to the minimum time interval required for theoretical crosstalk attenuation plus a preset offset. The preset offset is 1 nanosecond to ensure that the initial silence period is slightly larger than the theoretical minimum value, leaving a safety margin. The minimum time interval required for theoretical crosstalk attenuation is determined by step S3, and the value range is usually between 2 nanoseconds and 8 nanoseconds. Therefore, the initial silence period is between 3 nanoseconds and 9 nanoseconds.

[0078] Then, in the simulated annealing fine-tuning stage, the initial temperature, termination temperature, and cooling rate are set. A fixed number of Markov chain iterations are performed at each temperature. In each iteration, either a 2-opt subsequence inversion operation is randomly selected for the lighting sequence, or a random replacement operation is performed for the silent period. Whether to accept a new solution is determined by the exponential function of the cost increment and the current temperature. This process is repeated after cooling until the termination temperature is reached. The specific parameters for simulated annealing fine-tuning are: initial temperature 100, termination temperature 0.1, cooling rate 0.95, and Markov chain iterations at each temperature 200. The specific process of the 2-opt subsequence inversion operation is as follows: two positions a and b are randomly selected, where a is less than b. The subsequence from position a to position b in the lighting sequence is... The columns are reversed, and the corresponding subsequences in the silent period vector are also reversed. The specific process of the silent period random replacement operation is as follows: a position u is randomly selected, and the silent period of the position is replaced with a random integer in the range of 1 nanosecond to 20 nanoseconds. After each iteration, the difference between the global temporal cost function of the new solution and the current solution is calculated. If the difference is less than zero, the new solution is accepted directly. If the difference is greater than or equal to zero, the negative difference with the natural constant base is divided by the exponential function value of the current temperature as the acceptance probability, and a random number between 0 and 1 is generated. If the random number is less than the acceptance probability, the new solution is accepted; otherwise, the original solution is retained. After 200 iterations at each temperature, the current temperature is multiplied by the cooling rate of 0.95 to obtain the new temperature. The above process is repeated until the current temperature is 0.1 lower than the termination temperature.

[0079] In step S4, the main control unit also performs hard constraint verification and penalty function backflow during the simulated annealing fine-tuning process:

[0080] After each new solution is received, the total frame time (the sum of all silent periods) is checked to see if it exceeds a preset upper limit. If it does, a frame rate penalty term proportional to the square of the excess is added to the global temporal cost function. The preset upper limit for the total frame time is 50 microseconds. The frame rate penalty term is calculated as follows: first, the difference between the total frame time and 50 microseconds is calculated. If the difference is greater than zero, the square of the difference is taken and then multiplied by the frame rate penalty coefficient of 10 to obtain the value of the frame rate penalty term. This value is then added to the global temporal cost function.

[0081] Simultaneously, it checks whether the scan ratio of each sub-region and the KL divergence of the ideal uniform distribution exceed a preset threshold. If they do, a coverage uniformity penalty term proportional to the square of the excess is added to the global temporal cost function. The KL divergence threshold is preset to 0.1. The coverage uniformity penalty term is calculated as follows: First, calculate the KL divergence of each sub-region's scan ratio and the ideal uniform distribution ratio of 1 / 64 under the current lighting sequence. If the KL divergence is greater than 0.1, take the square of the difference between the KL divergence and 0.1, and then multiply it by the coverage uniformity penalty coefficient of 5 to obtain the value of the coverage uniformity penalty term. This value is then added to the global temporal cost function.

[0082] The cost function with the added penalty term is used for the acceptance decision in subsequent simulated annealing. The global temporal cost function with the added penalty term is equal to the original global temporal cost function plus the frame rate penalty term plus the coverage uniformity penalty term. This modified cost function is used for the cost increment calculation and acceptance probability decision in subsequent simulated annealing iterations, thereby ensuring that the optimization result simultaneously satisfies the hard constraints of frame rate and coverage uniformity.

[0083] In step S4, the main control unit feeds back the deconvolution residual output in step S2 to step S1 for recursive least squares update of the crosstalk matrix. The specific method is as follows:

[0084] The product of the deconvolution residual matrix and the transpose of the clean echo signal is calculated, then divided by the square of the Frobenius norm of the clean echo signal plus a preset positive number to obtain the initial update amount. The preset positive number is 10 to the power of -8 to prevent denominator overflow when the clean echo signal energy is zero. The square of the Frobenius norm of the clean echo signal is equal to the sum of the squares of all elements of the clean echo signal matrix. The product of the deconvolution residual matrix and the transpose of the clean echo signal is a 512-row, 512-column matrix, where the element in the j-th row and i-th column is equal to the dot product of the j-th row of the deconvolution residual matrix and the i-th row of the clean echo signal. Each element of this product matrix is ​​divided by the square of the Frobenius norm plus 10 to the power of -8 to obtain the initial update amount matrix.

[0085] The initial update is multiplied by a preset learning rate, and then multiplied element-wise by a confidence weighting matrix to obtain the final update. The preset learning rate is 0.01, used to control the step size of each update. The confidence weighting matrix is ​​a 512-row, 512-column matrix, where the element in the j-th row and i-th column is equal to 1 divided by 1 plus a scaling factor of 0.5 multiplied by the value of the element in the i-th row and j-th column of the crosstalk risk tensor. The crosstalk risk tensor element values ​​are constructed in step S3 and range from 0 to 1, so the confidence weighting matrix element values ​​are between 0.67 and 1. Units with higher crosstalk risk have smaller corresponding confidence weighting matrix element values, thereby reducing the correction magnitude of the coupling strength for that unit. Element-wise multiplication multiplies each element of the initial update matrix with the corresponding element of the confidence weighting matrix to obtain the final update matrix.

[0086] The final update amount is added to the current crosstalk matrix to obtain the updated crosstalk matrix. The current crosstalk matrix is ​​the crosstalk matrix constructed in step S1 and updated in the previous step, with a dimension of 512 rows and 512 columns. The final update amount matrix is ​​added to the corresponding position of the current crosstalk matrix to obtain the updated crosstalk matrix. This matrix is ​​used for steps S1, S2 and S3 in the next frame.

[0087] Each element of the confidence weighting matrix is ​​determined by the crosstalk risk tensor element values ​​of the two corresponding VCSEL emitting units in the crosstalk risk tensor. The element values ​​of the confidence weighting matrix are independent of the detector channel number; the higher the crosstalk risk of a VCSEL emitting unit, the smaller its corresponding weight. Specifically, for any two VCSEL emitting units i and k, the element in the j-th row and i-th column of the confidence weighting matrix (where j is any detector channel number) is equal to 1 divided by 1 plus 0.5 multiplied by the element value in the i-th row and k-th column of the crosstalk risk tensor. Since the crosstalk risk tensor element values ​​are between 0 and 1, and the confidence weighting matrix element values ​​are between 0.67 and 1, when the crosstalk risk tensor element value is 0, the weight is 1, indicating complete confidence in the update; when the crosstalk risk tensor element value is 1, the weight is 0.67, indicating a 33% suppression of the update.

[0088] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A short-pulse driving method for lidar based on matrix addressing, characterized in that, Specifically, the steps include the following: Step S1: The main control unit controls the horizontal drive module and the vertical drive module to light up each VCSEL light-emitting unit in the array one by one according to the predetermined test pattern, collects the original echo tensor of all detector channels, normalizes the response of each channel, and constructs an initial crosstalk matrix characterizing the optical and electromagnetic coupling strength between each unit. Step S2: The main control unit models the original echo tensor acquired in normal working mode as a superposition of clean echo signal, crosstalk component and noise, and uses the crosstalk matrix as physical constraint. It then uses a regularized non-negative matrix factorization algorithm to deconvolve the original echo tensor to separate the clean echo signal, crosstalk component, signal-to-noise ratio of each channel and deconvolution residual. Step S3: The main control unit constructs a crosstalk risk tensor based on the clean echo signal, crosstalk components and signal-to-noise ratio of each channel, and constructs a global time-series cost function based on the crosstalk risk tensor, which includes spatial crosstalk term, time decoupling term, scan integrity term and power consumption equalization term. Step S4: The main control unit uses the global temporal cost function as the objective and employs a hybrid strategy of greedy initialization combined with simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame. The optimized lighting order and silent period are then written into the horizontal driving module and the vertical driving module. At the same time, the deconvolution residual output in step S2 is fed back to step S1, which is used to recursively update the crosstalk matrix with least squares each time step S1 is executed, so as to obtain the crosstalk matrix with real-time calibration and form a closed-loop feedback.

2. The short-pulse driving method for lidar based on matrix addressing according to claim 1, characterized in that: In step S1, when constructing the initial crosstalk matrix, the main control unit uses the detector channel corresponding to the position of each VCSEL light-emitting unit as the main channel, and uses the echo signal of the main channel as the reference template to calculate the amplitude square coherence function of the echo signal of the other detector channels and the reference template within the effective frequency band preset by the laser pulse spectrum. Then, the amplitude squared coherence function is integrated over the effective frequency band, and the integration result is divided by the integral value of the self-coherence function of the main channel to obtain the off-diagonal elements of the initial crosstalk matrix. The integral value of the self-coherence function of the main channel is always equal to the effective frequency band width, thus ensuring that the diagonal elements of the initial crosstalk matrix are 1, and the off-diagonal elements represent the relative coherence coupling strength between each channel, and the values ​​are between 0 and 1.

3. The short-pulse driving method for lidar based on matrix addressing according to claim 2, characterized in that: In step S1, when the main control unit acquires the original echo tensor, it applies a single short pulse drive to each VCSEL light-emitting unit in sequence, and synchronously records the time-domain echo signals of all M detector channels at T time sampling points to form a three-dimensional original echo tensor. Then, for each VCSEL light-emitting unit, the time-domain echo signal is extracted from the main channel as a reference template, and the cross power spectral density of the time-domain echo signal of each other channel with the reference template and their respective self-power spectral density are calculated, thereby obtaining the amplitude squared coherence function. The lower and upper limits of the effective frequency band are preset by the laser pulse spectrum to ensure that only the frequency components with concentrated signal energy are retained, and to eliminate the interference of out-of-band noise on the coupling strength estimation.

4. The short-pulse driving method for lidar based on matrix addressing according to claim 3, characterized in that: In step S2, the cost function optimized by the regularized nonnegative matrix factorization algorithm used by the main control unit consists of three terms: The first item is the data fidelity item, which measures the degree of difference between the original echo tensor and the reconstructed echo tensor. The reconstructed echo tensor is composed of the crosstalk matrix multiplied by the convolution result of the clean echo signal and the standard pulse waveform template, plus the crosstalk component. The second term is the total variation regularization term, which is used to constrain the sum of the absolute values ​​of the gradients of the clean echo signal in the time direction, so as to suppress the oscillation artifacts generated by deconvolution and preserve the rising edge of the echo. The third term is the timing overlap penalty term, which is used to square the energy of the VCSEL light-emitting unit corresponding to the main channel and its neighboring VCSEL light-emitting units determined by the crosstalk matrix within the theoretical echo arrival time window according to the lighting order of the current frame, so as to avoid crosstalk energy being incorrectly allocated to the clean echo signal.

5. The short-pulse driving method for lidar based on matrix addressing according to claim 4, characterized in that: In step S2, the main control unit uses the alternating direction multiplier method to iteratively solve the cost function. The iterative process includes: Initialize the clean echo signal and crosstalk components to be zero matrices, and initialize the Lagrange multipliers to be zero matrices; In each iteration, the crosstalk components and Lagrange multipliers are first fixed, and the clean echo signal is updated using the near-end gradient descent method, where the step size of the near-end gradient descent method is set to a preset value and the number of inner loop iterations is fixed; then the updated clean echo signal and Lagrange multipliers are fixed, and the crosstalk components are updated using the closed-form solution, which includes rectified linear unit functions to ensure the non-negativity of the crosstalk components; then the Lagrange multipliers are updated. Repeat the iteration until the relative change between two adjacent cost function values ​​is less than a preset threshold or the maximum number of iterations is reached. At this point, the final clean echo signal, crosstalk component, signal-to-noise ratio of each channel, and deconvolution residual are output.

6. The short-pulse driving method for lidar based on matrix addressing according to claim 5, characterized in that: In step S3, the main control unit constructs the crosstalk risk tensor in the following way: For any two VCSEL emitting units, the main control unit traverses all detector channels, multiplies the relative coupling strength of the first VCSEL emitting unit to the current channel in the crosstalk matrix with the relative coupling strength of the second VCSEL emitting unit to the current channel, divides by the signal-to-noise ratio of the current channel and adds a preset positive number to obtain the contribution value of the current channel. Then, the contribution values ​​of all channels are summed and multiplied by a spatial attenuation factor. The spatial attenuation factor is determined by the physical distance between the two VCSEL light-emitting units in the array and follows a Gaussian attenuation law. Finally, the crosstalk risk tensor element values ​​corresponding to the two VCSEL light-emitting units are obtained in the crosstalk risk tensor.

7. The short-pulse driving method for lidar based on matrix addressing according to claim 6, characterized in that: In step S3, the master control unit constructs the global time-series cost function in the following way: The spatial crosstalk term is obtained by summing the crosstalk risk tensor element values ​​of all adjacent unit pairs in the lighting sequence of the next frame; The time decoupling term is obtained by substituting the deviation between the actual silence period of each adjacent unit pair and the minimum time interval required for theoretical crosstalk attenuation into the Huber loss function and multiplying it by a preset first weighting coefficient, and then summing them up. The minimum time interval required for theoretical crosstalk attenuation is determined by the maximum half-life of the crosstalk component output in step S2 on the set of detector channels that are jointly affected. The scan integrity term is obtained by multiplying the scan ratio of each sub-region of the current frame by the KL divergence of an ideally uniform distribution by a preset second weighting coefficient; The power consumption equalization term is obtained by multiplying the sum of the squares of all quiet periods by a preset third weighting coefficient; the sum of the four terms yields the global timing cost function.

8. The short-pulse driving method for lidar based on matrix addressing according to claim 7, characterized in that: In step S4, the main control unit uses a hybrid strategy combining greedy initialization and simulated annealing fine-tuning to solve for the lighting order and silent period vector of the next frame. The specific method is as follows: First, using the crosstalk risk tensor as the distance matrix, the initial lighting order is constructed using the nearest neighbor heuristic. That is, a starting point is randomly selected from all VCSEL light-emitting units to be lit in the next frame. Then, the unselected VCSEL light-emitting unit with the smallest crosstalk risk tensor element value with the current last VCSEL light-emitting unit is selected as the next lighting unit. All silent periods are initialized to the minimum time interval required for theoretical crosstalk attenuation plus a preset offset. Then, in the simulated annealing fine-tuning stage, the initial temperature, the termination temperature, and the cooling rate are set. A fixed number of Markov chain iterations are performed at each temperature. In each iteration, a 2-opt subsequence reversal operation is randomly selected for the lighting order or a random replacement operation is performed for the silent period. The decision to accept a new solution is made based on the exponential function of the cost increment and the current temperature. The process is repeated after cooling until the termination temperature is reached.

9. A short-pulse driving method for lidar based on matrix addressing according to claim 8, characterized in that: In step S4, the main control unit also performs hard constraint verification and penalty function backflow during the simulated annealing fine-tuning process: After each new solution is received, the total frame time, which is the sum of all silent periods, is checked to see if it exceeds the preset upper limit. If it does, a frame rate penalty term proportional to the square of the excess is added to the global temporal cost function. At the same time, it checks whether the scan ratio of each sub-region and the KL divergence of the ideal uniform distribution exceed the preset threshold. If they exceed the threshold, a coverage uniformity penalty term proportional to the square of the excess is added to the global temporal cost function. The cost function with the added penalty term is used for acceptance judgment in subsequent simulated annealing.

10. A short-pulse driving method for lidar based on matrix addressing according to claim 9, characterized in that: In step S4, the main control unit feeds back the deconvolution residual output in step S2 to step S1 for recursive least squares update of the crosstalk matrix. The specific method is as follows: Calculate the product of the deconvolution residual matrix and the transpose matrix of the clean echo signal, then divide by the square of the Frobenius norm of the clean echo signal plus a preset positive number to obtain the initial update amount; The initial update amount is multiplied by a preset learning rate, and then multiplied element-wise by a confidence weighted matrix to obtain the final update amount. Add the final update value to the current crosstalk matrix to obtain the updated crosstalk matrix; Each element of the confidence weighting matrix is ​​determined by the crosstalk risk tensor element values ​​of the two corresponding VCSEL emitting units in the crosstalk risk tensor. The element values ​​of the confidence weighting matrix are independent of the detector channel number. The VCSEL emitting unit with higher crosstalk risk has a smaller weight.

Images