A detection height correction method, device, equipment and medium of a visual detector

Abstract

The application relates to the technical field of target detection, and discloses a detection height correction method, device, equipment and medium of a visual detector, which comprises the following steps: calculating a false detection probability when the visual detector detects an obstacle height; determining a point cloud correction parameter for the obstacle height through point cloud data; determining a correction weight corresponding to the point cloud data based on the false detection probability and a current environment; and correcting a first obstacle height output by the visual detector by using the correction weight and the point cloud correction parameter to obtain a second obstacle height after correction. The application improves the detection accuracy of the obstacle height.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and specifically to a method, apparatus, device, and medium for correcting the detection height of a visual detector. Background Technology

[0002] Estimating the height of obstacles is a key technology in intelligent parking systems. Currently, obstacle height detection based on visual detectors (such as visual occupancy) is a common approach. However, this approach still has some shortcomings. For example, visual occupancy height estimation is easily affected by uneven ground, shadows, lighting intensity, data diversity, and data annotation quality, leading to errors in obstacle height estimation. This, in turn, causes the intelligent parking system to incorrectly estimate the vehicle's path planning, thus affecting the comfort and safety of autonomous driving vehicles. Therefore, improving the accuracy of obstacle height detection by visual detectors is a problem worthy of research. Summary of the Invention

[0003] This invention provides a method, apparatus, device, and medium for correcting the detection height of a visual detector, in order to solve the problem of inaccurate obstacle height detection by visual detectors.

[0004] In a first aspect, the present invention provides a method for correcting the detection height of a visual detector, the method comprising: calculating the false detection probability when the visual detector detects the height of an obstacle; determining point cloud correction parameters for the obstacle height using point cloud data; determining correction weights corresponding to the point cloud data based on the false detection probability and the current environment; and correcting the first obstacle height output by the visual detector using the correction weights and the point cloud correction parameters to obtain a corrected second obstacle height.

[0005] Based on the aforementioned technical methods, by quantifying the inherent false detection probability of visual detectors in height detection tasks, crucial prior information is provided for subsequent fusion correction. Instead of blindly trusting any single sensor, the correction weights of the laser point cloud data are dynamically calculated based on the reliability of visual detection and current environmental conditions. Through a weighted fusion strategy, high-precision, but potentially sparse, physical measurement data can be intelligently utilized to correct semantically rich visual data that is susceptible to environmental interference. This overcomes the shortcomings of visual height estimation, which is easily affected by lighting, ground conditions, and data quality, organically combining the perceptual bias of two-dimensional images with accurate measurements in three-dimensional space, ultimately outputting a robust and accurate corrected height.

[0006] In some optional implementations, calculating the false detection probability when the visual detector detects the height of an obstacle includes: collecting height set data of different obstacles using the visual detector; calculating the corresponding visual height mean and visual height standard deviation based on the height set data; and inputting the visual height mean, visual height standard deviation, and the height of the first obstacle into an improved Gaussian distribution statistical model to calculate the false detection probability.

[0007] Based on the aforementioned technical methods, massive amounts of height data output by the visual detector in actual or test scenarios are collected and statistically analyzed to calculate the average error and dispersion of its height estimation. Subsequently, the corresponding statistical features, along with the currently observed obstacle height, are input into an improved Gaussian distribution model, concretizing the abstract detector performance into calculable probabilities. Compared to simple fixed thresholds or linear corrections, calculations based on statistical models can more scientifically and objectively reflect the confidence level of the visual detector under different inputs.

[0008] In some optional implementations, the mean visual height, standard deviation of visual height, and height of the first obstacle are input into an improved Gaussian distribution statistical model to calculate the false detection probability, including: The false detection probability is calculated using the following improved Gaussian distribution statistical model. :

[0009] In the formula, Indicates the mean visual height. Indicates the standard deviation of visual height. Indicates the height of the first obstacle. This indicates the maximum false positive rate for the first obstacle. Sensitivity used to control the false detection probability offset.

[0010] Based on the aforementioned technical methods, compared to the standard Gaussian distribution, this model introduces two key parameters: the maximum suspicion rate A and the sensitivity factor B. Parameter A defines the upper limit of the probability of generating false detection suspicion for visual observation results, which conforms to the actual physical meaning. Parameter B provides flexibility in adjusting the model's sensitivity to data offset; by adjusting the value of B, the rate at which the error probability decays as the deviation increases can be controlled. Engineers can optimize the calculation logic of the false detection probability according to the performance characteristics of different visual inspection networks, different sensor configurations, or the security requirements of different application scenarios. This allows for a more accurate fit to the true output characteristics of various visual detectors, providing a more reliable quantitative basis for the entire fusion correction system.

[0011] In some optional implementations, point cloud correction parameters for obstacle height are determined using point cloud data, including: acquiring point cloud data detected by radar equipment within its optimal observation range; calculating the standard deviation of point cloud height based on the height values ​​of different obstacles scanned using the point cloud data; and using the standard deviation of point cloud height as a point cloud correction parameter.

[0012] Based on the aforementioned technical methods, although lidar offers extremely high ranging accuracy within a specific range, the elevation values ​​of its output point cloud can still fluctuate due to factors such as sparse scan lines, point cloud noise, or differences in the reflectivity of obstacle surface materials. Therefore, using the standard deviation of point cloud height values ​​is an excellent measure of uncertainty. Furthermore, by limiting the use of point cloud data to the "optimal observation range," such as data within a certain distance ahead, the problem of decreased measurement accuracy of lidar in near blind spots or distant areas with sparse point clouds can be effectively avoided.

[0013] In some optional implementations, acquiring point cloud data detected by the radar device within its optimal observation range includes: recording point cloud sub-data collected by the radar device from its optimal observation range at different times; and stitching the point cloud data together according to the positions of the point cloud sub-data recorded at each time to obtain the point cloud data.

[0014] Based on the aforementioned technical means, this invention further solves the technical problems of sparsity in single-frame laser point cloud data and the difficulty in continuously covering the optimal observation area due to vehicle movement. In actual driving, vehicles are constantly moving, and the optimal observation area is passed over time. By recording and saving point cloud sub-data within the optimal observation area at different historical times, and then using odometer information for coordinate transformation and stitching, a "spatiotemporally cumulative" local point cloud map is constructed. Calculating the height standard deviation using the aforementioned point cloud data with higher information density yields more accurate and stable results, significantly reducing misjudgments caused by missing single-frame data or noise points.

[0015] In some optional implementations, the standard deviation of point cloud height is calculated by using the height values ​​of different obstacles scanned from point cloud data, including: dividing the ground into a grid map according to the detection range of the radar equipment; filling the grid map with point cloud data and calculating the median height of the point cloud in each cell of the grid map to obtain the height values ​​for different obstacles scanned; and calculating the standard deviation of the height values ​​for different obstacles scanned to obtain the standard deviation of point cloud height.

[0016] Based on the aforementioned technical means, this invention constructs a raster map to discretize and spatially structure continuous, unordered point cloud data. Secondly, within each raster, the "median" for calculating the point cloud height is selected, enabling robust estimation of the true height value within each raster without the need for complex filtering algorithms. By traversing all relevant raster cells, a series of robust single-point height values ​​are obtained. Finally, the standard deviation calculated based on these values ​​accurately reflects the spatial stability and consistency of the point cloud data.

[0017] In some optional implementations, the correction weights corresponding to the point cloud data are determined based on the false detection probability and the current environment, including: querying a trust coefficient used to characterize the degree of trust in visual detection through the current environment; and determining the correction weights by the ratio of the false detection probability to a first parameter, wherein the first parameter is obtained by summing the trust coefficient and the false detection probability.

[0018] Based on the aforementioned technical means, this invention integrates two dimensions of information to determine the correction weights for the laser point cloud. The first is the false detection probability from the visual detector itself, and the second is a trust coefficient representing the degree of influence of the current environment on visual detection. When visual detection is highly reliable, the correction weights approach 0, at which point the correction algorithm almost completely trusts the visual detection, avoiding interference from sparse measurements by the LiDAR. Conversely, when visual detection is unreliable under poor lighting or complex terrain conditions, the correction weights approach 1, and the fusion algorithm relies more heavily on the correction of the laser point cloud. More importantly, the introduction of the trust coefficient allows the system to proactively adapt to different weather conditions (such as rain and snow) and lighting conditions (such as night and backlight), ensuring that the fusion strategy maintains optimal performance under various operating conditions, achieving true adaptive fusion.

[0019] In some optional implementations, a trust coefficient used to characterize the degree of trust in visual detection is queried through the current environment, including: acquiring historical obstacle height data collected by the visual detector under different environmental conditions, and calculating the average height detection value and standard deviation of height detection for each environment based on the historical obstacle height data; setting a trust coefficient corresponding to each environment based on the average height detection value and standard deviation of height detection, and forming an environment-trust coefficient mapping table; acquiring current environment information, and searching for a trust coefficient that matches the current environment information in the environment-trust coefficient mapping table.

[0020] Based on the aforementioned technical methods, unlike complex theoretical models, this approach directly reflects the degree of performance degradation of visual detectors in the real world by collecting historical data under different environments (rain, snow, backlight, underground parking, etc.) and statistically analyzing the visual detection characteristics (mean, standard deviation) under each environment. By setting the confidence coefficient based on this real statistical data, the confidence coefficient has a clear physical meaning and quantitative basis, rather than being a vague parameter set based on experience. This further improves the accuracy of subsequent high-level corrections and significantly reduces computational costs, thereby increasing computational efficiency.

[0021] In some optional implementations, the first obstacle height output by the visual detector is corrected using correction weights and point cloud correction parameters to obtain a corrected second obstacle height. This includes: calculating the first height by multiplying the first obstacle height by the second parameter, where the second parameter is obtained by summing a confidence coefficient with a preset scalar value; calculating the corrected height by multiplying the point cloud correction parameters and correction weights; and determining the second obstacle height using the sum of the first height and the corrected height.

[0022] This invention proposes a specific fusion formula for final height correction. This formula utilizes an environmental coefficient to provide a basic positive compensation for the height of the first obstacle. The second part introduces a dynamic correction term driven by laser point cloud data. It doesn't simply add the point cloud height, but multiplies the standard deviation of the point cloud height by a weighting coefficient. This provides a "fine-tuning term" based on the stability of the laser point cloud data itself. Through the composite correction model of this invention, both statistically based prior compensation and dynamic fine-tuning based on real-time data are integrated. These two complement each other, ultimately enabling accurate and robust correction of the underestimated visual height to near the true value.

[0023] Secondly, the present invention provides a detection height correction device for a visual detector, the device comprising: a false detection probability identification module for calculating the false detection probability when the visual detector detects the height of an obstacle; a correction amount determination module for determining point cloud correction parameters for the obstacle height using point cloud data; a correction weight module for determining the correction weight corresponding to the point cloud data based on the false detection probability and the current environment; and a height correction module for correcting the first obstacle height output by the visual detector using the correction weight and the point cloud correction parameters to obtain a corrected second obstacle height.

[0024] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0025] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method described in the first aspect or any corresponding embodiment thereof.

[0026] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to perform the method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a schematic flowchart of a first method for correcting the detection height of a visual detector according to an embodiment of the present invention; Figure 2 This is a second flowchart illustrating a method for correcting the detection height of a visual detector according to an embodiment of the present invention. Figure 3 This is a third flowchart of a detection height correction device for a visual detector according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0031] 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 technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] According to an embodiment of the present invention, a method for correcting the detection height of a visual detector is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] This embodiment provides a method for correcting the detection height of a visual detector. Figure 1 This is a flowchart of a detection height correction method for a visual detector according to an embodiment of the present invention. The process includes the following steps: Step S101: Calculate the false detection probability when the visual detector detects the height of an obstacle; Step S102: Determine the point cloud correction parameters for obstacle height using point cloud data; Step S103: Determine the correction weights corresponding to the point cloud data based on the false detection probability and the current environment; Step S104: Correct the first obstacle height output by the visual detector using the correction weight and point cloud correction parameters to obtain the corrected second obstacle height.

[0034] Specifically, in intelligent driving scenarios (especially automatic parking), the obstacle height output by visual detectors (such as visual occupancy networks) is easily affected by environmental interference such as uneven lighting, ground undulations, and shadow occlusion, often resulting in underestimation and directly impacting vehicle accessibility judgment and path planning safety. The height correction method proposed in this invention addresses the technical problem of insufficient robustness in visual height detection by quantifying the reliability of visual detection and fusing precise laser point cloud measurements.

[0035] First, the false detection probability of the visual detector when detecting obstacle height is determined. The false detection probability specifically refers to the quantified probability that the obstacle height output by the visual detector deviates from the true value, leading to an incorrect judgment of height attribute (crossable / insurmountable). It is used to characterize the reliability of the visual detection results. In practical applications, a large amount of obstacle sample data can be collected in parking scenarios (such as curbs around horizontal and perpendicular parking spaces, low bollards, flower bed edges, etc.). Using the manually calibrated true height as a benchmark, the overall distribution characteristics of the visual detector's output height are statistically analyzed. For example, in a low-light scenario in an urban underground parking garage, 1000 sets of curb samples were collected, with a true average height of 15cm, a visual detector output height of 11cm, and a standard deviation σ of 2.1cm. In a daytime backlit outdoor scenario, 800 sets of bollard samples were collected, with a true average height of 30cm, a visual output height of 24cm, and a standard deviation σ of 2.8cm. Based on the mean μ and standard deviation σ obtained from the above statistics, the false detection probability of the current detection result is quantified using a probabilistic statistical model. The closer the value is to 1, the lower the reliability of the visual height result, and vice versa. This step transforms the abstract visual detection error into a quantifiable probabilistic indicator, providing a core prior basis for subsequent fusion correction.

[0036] After calculating the false detection probability, point cloud correction parameters for obstacle height are determined using point cloud data. Point cloud data is generated by the laser beam emitted by the vehicle-mounted LiDAR scanning the environment. It is a discrete set of points containing information such as three-dimensional spatial coordinates and reflection intensity, possessing advantages such as high accuracy in three-dimensional spatial measurement and insensitivity to illumination. The point cloud correction parameters are quantitative indicators characterizing the stability of the laser point cloud in obstacle height measurement, used to measure the reliable correction capability of the laser measurement.

[0037] Based on the false detection probability and point cloud correction parameters, the correction weights corresponding to the point cloud data are determined in conjunction with the current environment. The correction weights are coefficients between 0 and 1, used to dynamically allocate the contribution ratio of visual detection results and laser point cloud data in height correction. The closer the weight is to 1, the more reliant it is on the laser point cloud correction results; conversely, the closer the weight is to 1, the more trust is placed in the visual detection results. The current environment encompasses external conditions affecting visual detection accuracy, such as lighting (strong light / weak light / backlight), weather (sunny / rainy / cloudy), and location (basement / outdoor). Different correction weights are assigned to different environments based on historical data statistics.

[0038] Finally, the height of the first obstacle output by the visual detector is corrected using the correction weights and point cloud correction parameters to obtain the corrected height of the second obstacle. The first obstacle height is the original height value directly output by the visual detector, and the second obstacle height is the final height value used for vehicle decision-making after multi-source data fusion correction. In a real parking scenario, assuming the calculation is performed using the values ​​from the previous example, in an underground parking garage scenario, the visual output curb height is 11cm, the point cloud correction parameter (standard deviation) is 0.8cm, and the laser correction weight is 0.71. After fusion calculation, the second obstacle height is approximately 11.57cm, which is closer to the actual 12cm judgment benchmark and can avoid misjudging a 12cm curb as a low obstacle that can be crossed.

[0039] Through the aforementioned technical means, this invention quantifies the reliability of visual detection by the probability of false detection, abandoning the crude mode of traditional fixed threshold correction and achieving a refined assessment of visual errors through prior probability. Secondly, it integrates the precise 3D measurement capabilities of laser point clouds to compensate for the perception shortcomings of visual sensors in complex environments such as lighting and shadows, breaking through the performance limits of a single sensor. Furthermore, it calculates dynamically corrected weights to adapt to different scenarios, significantly improving the accuracy of the output obstacle height, providing reliable perception input for intelligent parking systems, effectively reducing path planning errors and collision risks caused by height misjudgments, and greatly improving the environmental adaptability and operational safety of intelligent driving systems. It possesses strong engineering application value in various parking scenarios such as urban underground parking garages and open-air parking lots.

[0040] In some alternative implementations, step S101 includes: Step a1: Collect height data of different obstacles using a visual detector; Step a2: Calculate the mean visual height and standard deviation of visual height based on the height set data; Step a3: Input the mean visual height, standard deviation of visual height, and height of the first obstacle into the improved Gaussian distribution statistical model to calculate the false detection probability.

[0041] Specifically, the detection error of a visual detector in determining the height of an obstacle is not randomly discretely distributed, but generally exhibits a normal distribution. The fluctuation pattern of this error can be accurately characterized by a statistical model. This invention constructs statistical characteristics of visual detection errors using massive amounts of real-world data. Based on an improved Gaussian distribution model, it achieves precise quantification of the false detection probability, providing a reliable probabilistic basis for subsequent multi-source data fusion.

[0042] The first step aims to build a visual height sample library covering diverse scenarios and various obstacle types to ensure that subsequent statistical results have scenario universality and data representativeness. The height set data is a collection of raw height values ​​output by the visual detector under different scenarios and obstacle categories.

[0043] For example, in actual data collection, typical intelligent driving scenarios such as urban underground parking garages, open-air parking lots, and park roads can be selected. Obstacle categories can cover common targets in parking and low-speed driving scenarios, including curbs, bollards, flower bed edges, low guardrails, and speed bumps. Data collection should simultaneously cover different lighting conditions (strong midday light, weak evening light, no light in underground parking garages, backlighting) and different ground conditions (flat concrete, slightly undulating asphalt roads, and paved surfaces) to avoid data bias caused by a single environment. For instance, in the weak light scenario of an underground parking garage, samples of obstacles around marked parking spaces such as horizontal, vertical, and angled parking spaces are continuously collected; in the backlighting scenario of an open-air parking garage, 400 sets of bollard samples with a height of 25–40 cm are collected; and in park roads, 300 sets of speed bump samples with a height of 5–15 cm are collected. Finally, a high-resolution dataset is compiled to ensure that the sample size is sufficient to support the validity of subsequent statistical analysis.

[0044] Subsequently, in this embodiment of the invention, the mean visual height and standard deviation of visual height corresponding to the visual detector are calculated based on the height set data.

[0045] After completing sample data collection, the core distribution parameters of the visual detector's height detection error—namely, the visual height mean and the visual height standard deviation—are extracted using mathematical statistics. The visual height mean characterizes the overall average level of the visual detector's output height, reflecting the overall deviation trend of the detection results. The visual height standard deviation characterizes the dispersion of the output height relative to the mean, quantifying the stability of the detection results' fluctuations.

[0046] In the specific calculation, the collected height set data is first preprocessed to remove outliers caused by sensor malfunctions or extreme occlusion, retaining only valid sample data. Let N be the total number of valid samples output by the visual occupancy after preprocessing. To randomly select a certain number of samples, the visual output height of the i-th sample is... Mean visual height The calculation formula is:

[0047] Visual height standard deviation The calculation formula is:

[0048] The above parameters accurately characterize the distribution features of visual detection height in different scenarios, providing basic input for improving the Gaussian model.

[0049] The mean visual height, standard deviation of visual height, and height of the first obstacle are then input into the improved Gaussian distribution statistical model to calculate the false detection probability.

[0050] This invention relies on an improved Gaussian distribution statistical model to correlate static statistical parameters with real-time detection results, quantifying the false detection probability of the current obstacle height. The first obstacle height is the original height value of a single obstacle output by the visual detector in real time. The improved Gaussian distribution statistical model is a statistical model that optimizes probability calculation logic by introducing adjustment parameters adapted to the characteristics of visual detection based on the standard Gaussian distribution, accurately matching the actual distribution pattern of visual detection errors. The false detection probability specifically refers to the probability that the current visual output height leads to an incorrect judgment of the obstacle's height attribute (crossable / non-crossable), with a value range of [0,1].

[0051] In this invention, the improved Gaussian distribution statistical model expression is as follows:

[0052] In the formula, Indicates the probability of false positives. Indicates the mean visual height. Indicates the standard deviation of visual height. Indicates the height of the first obstacle. This represents the maximum suspicion rate for false detection of the height of the first obstacle, with a value range of (0,1]. It represents the highest degree of suspicion for the visual detection result, and is usually taken as 0.5 in engineering. The sensitivity used to control the false detection probability shift is greater than 0. It is used to control the decay rate of the false detection probability with the detection deviation and is adapted to visual detectors with different performance. It is usually set to 2.

[0053] For example, combining the aforementioned examples with numerical calculations: In a low-light scenario in a parking garage, the height of the first obstacle on a certain curb is detected in real time. Substitute =11.2、 =2.05、 =0.5、 =2, calculated to .

[0054] The false detection rate for the curb height detection is approximately 49.4%, indicating that the reliability of the visual results is low.

[0055] Through the aforementioned technical means, this embodiment of the invention constructs a visual height sample library for different scenarios, ensuring that statistical parameters closely match actual working conditions. It utilizes the mean and standard deviation to accurately characterize the error distribution, abandoning the crude method of assigning probabilities based on experience. Furthermore, by improving the Gaussian model, it dynamically correlates static statistical features with real-time detection data, achieving accurate and quantitative output of false detection probabilities. This method is fully applicable to the error characteristics of visual detectors in complex environments, and the output false detection probabilities possess scene adaptability and numerical reliability. It provides a solid statistical foundation for subsequent dynamic allocation of correction weights and accurate correction of obstacle height, effectively avoiding perception biases of a single visual sensor under complex conditions and significantly improving the robustness of the intelligent driving perception system.

[0056] In some optional implementations, step S102 above includes: Step b1: Acquire point cloud data detected by the radar equipment within its optimal observation range; Step b2: Calculate the standard deviation of point cloud height by scanning the height values ​​of different obstacles using point cloud data, and use the standard deviation of point cloud height as a point cloud correction parameter.

[0057] Specifically, as an active three-dimensional sensor, lidar has significantly better spatial measurement accuracy than visual perception. However, the effective measurement quality is significantly affected by the observation distance, point cloud density, and installation obstruction. Thus, the lidar acquires point cloud data detected by the radar equipment within its optimal observation range.

[0058] Radar equipment includes, but is not limited to, vehicle-mounted lidar. By emitting and receiving laser pulses, it outputs a discrete point cloud containing information such as three-dimensional coordinates, reflection intensity, and echo count, providing all-weather, high-precision ranging capabilities. The optimal observation range is a pre-defined distance interval based on the radar's installation location, horizontal and vertical fields of view, and the point cloud density attenuation characteristics with distance. Within this range, the point cloud distribution is uniform, noise levels are low, and ranging errors are stable, making it the optimal area for obstacle height measurement. In engineering, for forward-facing main radars, a rectangular area of ​​6 to 10 meters in front of the vehicle and 7.5 meters to the left and right is typically set as the optimal observation range. This range avoids angular distortion caused by near-field blind spots and measurement instability caused by sparse far-field point clouds and decreased reflection signal-to-noise ratio.

[0059] Meanwhile, the filtered data is time-series smoothed, and point clouds within the optimal observation range are continuously collected and cached for multiple frames. Vehicle odometer information is used to complete coordinate transformation and spatial stitching between adjacent frames to construct a local high-density point cloud set.

[0060] After obtaining the effective point cloud within the optimal observation range, obstacle heights are extracted and dispersion indices are calculated using rasterized statistics and robust estimation methods to form point cloud correction parameters. Point cloud height values ​​refer to the spatial coordinates of obstacles in the vertical direction (Z-axis). The standard deviation of point cloud height characterizes the degree of fluctuation in laser measurement heights within the same obstacle area; the smaller the value, the better the measurement consistency and the higher the reliability. In this embodiment of the invention, the point cloud correction parameter, i.e., the aforementioned standard deviation, is used to quantify the reliability level of laser measurements and participates in subsequent fusion weight and correction calculations.

[0061] This invention, through defining the optimal observation range of the radar, suppresses near-field distortion and far-field noise at the source, ensuring the quality of input data. Using the height standard deviation as a correction parameter, the uncertainty of laser measurement is quantified into a computationally achievable value, enabling quantitative and adaptive multi-sensor fusion. This method fully leverages the advantages of lidar's three-dimensional measurement, avoids the sparsity defects of single-frame data, provides a reliable benchmark for subsequent weight allocation and height correction, significantly improves the stability and accuracy of obstacle height estimation in complex scenarios, and enhances the robustness and engineering practicality of intelligent driving perception systems.

[0062] In some alternative implementations, step b1 above includes: Step c1: Record the point cloud sub-data collected by the radar equipment from the optimal observation range at different times; Step c2: stitch the point cloud data together according to the positions of the point cloud sub-data recorded at each time point to obtain the point cloud data.

[0063] Specifically, during vehicle movement, the effective point cloud coverage of a single frame is limited due to the field of view limitation of vehicle-mounted LiDAR, which is prone to target occlusion or data sparsity, resulting in insufficient obstacle height sampling and decreased measurement stability. This implementation uses temporal memory and odometry-driven spatial coordinate transformation to dynamically stitch together point cloud sub-data from multiple moments within the optimal observation area, forming a high-density, continuously covered point cloud set, effectively solving the problems of single-frame data sparsity and target "out of view" in moving scenes.

[0064] The optimal observation range is determined through prior statistical analysis of point cloud distribution density, ranging error, and signal-to-noise ratio. Typically, a rectangular area is selected, extending 6.0 to 8.8 meters in front of the vehicle and 7.5 meters laterally to the left and right. This range falls within the radar's effective detection range of 10 meters, avoiding measurement deviations caused by near-field angular distortion and installation obstructions, while also preventing accuracy degradation due to far-field point cloud density attenuation and weakened reflected signals. It represents the optimal range that balances measurement stability and spatial coverage.

[0065] During low-speed driving or parking, this embodiment continuously acquires raw radar point clouds at fixed time intervals (e.g., 100 milliseconds), filters and retains local point clouds falling within the optimal observation range frame by frame, forming time-ordered point cloud sub-data. The point cloud sub-data only contains information such as the three-dimensional coordinates and reflection intensity within the current time and optimal observation range. Each segment of sub-data covers a local area of ​​the target, resulting in small data volume and high real-time performance. For example, when detecting static bollards in an open-air parking lot, point cloud sub-data is continuously recorded for five time points, with each segment containing approximately 40–60 valid points.

[0066] During vehicle movement, the optimal observation area dynamically shifts with changes in vehicle attitude, and previously recorded historical point cloud sub-data gradually moves out of the radar field of view (FOV) and cannot be directly observed in the current frame. This step utilizes the pose information output by the onboard odometer to construct the coordinate transformation relationship between adjacent time points, uniformly mapping point cloud sub-data from different time points within the optimal observation range to the current time coordinate system, thereby achieving spatial alignment and fusion stitching of time-series data.

[0067] Specifically, the odometer outputs the vehicle's attitude transformation matrix R (rotation matrix) and translation vector T at adjacent time points in real time, representing the vehicle's rotation and displacement in three-dimensional space. Let the spatial coordinates of a point in the optimal observation area at time t be... The transformation matrix can be used to map it to the coordinate system at time t+1. The transformation relationship is as follows:

[0068] By traversing all historical point cloud sub-data points and performing coordinate transformation point by point, the point clouds scattered across different times and local areas are unified and integrated into the current spatial coordinate system. This eliminates spatial misalignment caused by temporal offsets, resulting in stitched point cloud data that covers the complete outline of the target and has significantly improved density. For example, after coordinate transformation and stitching, the number of effective points in the aforementioned 5 time-time isolation stake data increases to more than 200, completely covering the side of the isolation stake column.

[0069] This invention ensures the measurement accuracy of raw sub-data from the source by precisely defining the optimal observation range. It employs time-series recording to continuously accumulate local target observation information. Through odometer-driven coordinate transformation, it achieves dynamic shifting and spatial stitching of historical point clouds, effectively overcoming data fragmentation defects caused by radar field-of-view limitations and vehicle movement. The stitched point cloud data possesses high density, high integrity, and high stability, providing sufficient and reliable sample support for subsequent obstacle height extraction and standard deviation calculation. This significantly improves the robustness of laser point cloud measurements, lays a solid data foundation for multi-sensor fusion correction of visual height deviations, and adapts to the perception needs of various dynamic, low-speed intelligent driving scenarios.

[0070] In some alternative implementations, step b2 above includes: Step d1: Divide the ground into a grid map according to the detection range of the radar equipment; Step d2: Fill the point cloud data into the raster map and calculate the median height of the point cloud in each cell of the raster map to obtain the height value for scanning different obstacles; Step d3: Calculate the standard deviation of the height values ​​of different obstacles to obtain the standard deviation of point cloud height.

[0071] Specifically, the raw point cloud data from lidar is large in volume, discretely distributed, and contains ground noise and reflection anomalies. Directly using it for height statistics is susceptible to interference from random outliers and consumes a significant amount of chip computing resources. This invention reduces the computational load while ensuring the stability and reliability of obstacle height measurement through point cloud sparsification.

[0072] A grid map is a two-dimensional planar representation of the effective detection area of ​​a radar, discretized into a regular grid. It is used to map disordered point clouds to a structured space, achieving data dimensionality reduction and sparsity processing. Based on a typical vehicle radar configuration, a body coordinate system is constructed with the rear axle center as the origin. For example, a rectangular area with a horizontal detection range of 15 meters in front and 15 meters to the left and right is defined, completely covering the radar's horizontal field of view (FOV) boundary. Within this area, square grids of uniform size (0.1m x 0.1m) are divided, with each grid corresponding to a 0.01 square meter spatial unit on the actual ground. After division, a structured grid map is formed, providing a regular spatial carrier for subsequent point cloud projection and statistical analysis.

[0073] The time-series stitched high-density point cloud data is projected onto the corresponding grid according to its horizontal coordinates (X,Y). Each grid collects the vertical height (Z value) of all point clouds falling within its spatial range. Due to interference from ground clutter, specular reflection, and airborne suspended particles, a small number of abnormally high or low noise points often exist within a single grid. Directly using the mean value is easily skewed by noise. Therefore, this embodiment of the invention uses the median as the grid representative height, which has natural robustness to outliers.

[0074] For example, for the Z-axis height value within each non-empty grid cell, the maximum and minimum values ​​are first removed to filter out extreme noise. The remaining valid height values ​​are then sorted in ascending order, and the value at the middle of the sorted sequence is selected as the median point cloud height of that grid cell. This value is the scan height value of the obstacle at the corresponding location. By traversing all grid cells containing obstacles, a set of discretely distributed, highly reliable obstacle scan height values ​​can be obtained.

[0075] Based on multiple sets of obstacle scanning height values ​​extracted from the raster, the dispersion of the point cloud height is calculated using mathematical statistics to obtain the standard deviation of the point cloud height. This is used to quantify the stability of the lidar's measurement of obstacle height, serving as a core parameter for subsequent fusion correction.

[0076] This invention achieves structured sparsity of massive point clouds through large field-of-view grid division, significantly reducing chip computational pressure and adapting to the computing power constraints of automotive embedded platforms. By employing a statistical strategy of eliminating extreme values ​​and taking the median, ground noise and reflection anomalies are effectively filtered out, accurately restoring the true height distribution of obstacles and providing reliable quantitative data for multi-sensor fusion.

[0077] In some optional implementations, step S103 above includes: Step e1: Query the trust coefficient used to characterize the degree of trust in visual detection through the current environment; Step e2: Determine the correction weight by the ratio of the false detection probability to the first parameter. The first parameter is obtained by summing the confidence coefficient and the false detection probability.

[0078] Specifically, in the high-precision correction process of multi-sensor fusion, the reliability of visual detection dynamically changes with environmental conditions such as lighting, weather, and location. Relying solely on the probability of false detection to allocate fusion weights is insufficient to fully reflect the systematic impact of the environment on visual perception capabilities. This implementation introduces a trust coefficient strongly correlated with the environment to establish a quantitative mapping of "environment-reliability-weight," enabling adaptive and interpretable calculation of laser point cloud correction weights and improving the rationality and robustness of the fusion strategy under complex conditions.

[0079] The current environment refers to external conditions that affect visual imaging quality, typically including: low light in a parking garage, strong light during the day, backlighting, rain, cloudy days, nighttime, smooth road surfaces, and uneven road surfaces. Trust coefficient. The value is a pre-calibrated dimensionless parameter, ranging from (0,1). The larger the value, the higher the reliability of visual detection in the current environment and the stronger the trust in it. Conversely, the lower the trust and the more dependent it is on laser measurement.

[0080] The trust coefficient is generated by statistical calibration of historical data. After obtaining the trust coefficient, the corrected weight is calculated.

[0081] Adjusted weights The weight represents the contribution of the laser point cloud to the height correction, ranging from (0,1). A weight closer to 1 indicates a greater reliance on the laser point cloud correction result. The first parameter represents the intermediate computational cost, determined by the false detection probability. With trust coefficient The summation is used for normalized weight calculation to ensure that the weight values ​​are stable and the physical meaning is clear.

[0082] The calculation logic for the corrected weights is as follows:

[0083] This formula reflects the dynamic balance relationship of the fusion weights, when the false detection probability... High (visual unreliability), trust level At lower (harsh environment) weight When the weight approaches 1, laser correction dominates; when the false detection probability is low and the confidence coefficient is high, the weight approaches 0, and visual results are given priority to avoid unnecessary biases introduced by laser sparse noise.

[0084] Example using actual numerical data: Trust coefficient in a low-light environment of a basement. =0.25, the false detection probability of visual detection of curbs. =0.72, substitute into the formula to calculate, and get =0.74. That is, the laser correction weight is 0.74, and the height correction is mainly based on laser measurement.

[0085] This invention, through mapping environment and trust coefficients, quantifies abstract environmental influences into computable parameters, overcoming the limitations of relying solely on false detection probabilities. It employs normalized ratio calculations to adjust weights, resulting in concise logic and clear physical meaning, avoiding the subjectivity of empirically based weight settings. The adjusted weights dynamically adapt to both environmental and visual reliability, implementing an intelligent fusion strategy of "relying on vision when the environment is good, and relying on laser when vision is poor," effectively balancing the advantages and disadvantages of both sensors. The overall solution features low computational overhead, strong real-time performance, and high engineering feasibility, significantly improving the accuracy and stability of obstacle height correction under different operating conditions, providing reliable perception and decision support for intelligent parking and low-speed driving scenarios.

[0086] In some optional implementations, step e1 above includes: Step f1: Obtain historical obstacle height data collected by the visual detector under different environmental conditions, and calculate the average height detection value and standard deviation of height detection under each environment based on the historical obstacle height data; Step f2: Set the trust coefficient for each environment based on the average height detection value and the standard deviation of height detection, and form an environment-trust coefficient mapping table; Step f3: Obtain the current environment information and search for the trust coefficient that matches the current environment information in the environment-trust coefficient mapping table.

[0087] Specifically, the accuracy of visual detection varies systematically with external conditions such as lighting, weather, and location. Assigning values ​​based solely on experience is insufficient to accurately depict the impact of the environment on perception performance. To ensure the statistical results have universal applicability, a historical dataset covering all scenarios and multiple obstacle categories needs to be constructed. During actual data collection, typical intelligent driving application scenarios were selected, including urban underground parking garages, open-air parking lots, urban roads, and park roads. The environmental dimension covered strong midday light, weak evening light, backlight, no lighting at night, rainy days, cloudy days, snowy days, smooth paved roads, and slightly undulating roads, fully reproducing various operating conditions that vehicles may encounter during driving.

[0088] The obstacle categories cover common targets in low-speed scenarios such as curbs, bollards, flower bed edges, low guardrails, speed bumps, and stones. During the data collection process, the actual height of the obstacle output by a high-precision laser rangefinder was used as a benchmark, and the corresponding height value output by the visual detector was recorded simultaneously.

[0089] After data collection, the datasets were grouped according to environmental labels, and outliers caused by sensor malfunctions or extreme occlusion were removed, retaining only valid samples. For each group of environmental data, the average height detection mean and standard deviation of the visual detection height were calculated. The average value reflects the overall deviation trend of the visual detection height in that environment, while the standard deviation quantifies the stability of the fluctuation in the detection results. For example, in a low-light environment in a parking garage, the average visual detection height of the curb was 11.3 cm with a standard deviation of 2.08 cm; in a sunny daytime environment, the average height of the bollard was 29.1 cm with a standard deviation of 1.52 cm; and in a backlit environment, the average height of the flower bed edge was 18.5 cm with a standard deviation of 2.75 cm. These statistical parameters accurately depict the error distribution patterns of visual detection under different environments.

[0090] The trust coefficient is a core parameter characterizing the reliability of visual inspection in the current environment. The larger the value, the higher the accuracy and stability of visual inspection in that environment, and the stronger the trust level. This step establishes objective and quantitative trust coefficient calibration rules based on the mean deviation and standard deviation of each environment: the smaller the mean deviation and the smaller the standard deviation, the higher the trust coefficient; conversely, the larger the deviation and the larger the standard deviation, the lower the trust coefficient.

[0091] When calibrating, a benchmark reference can be set first. For example, the optimal working condition is a sunny day with a smooth road surface, which has the smallest average deviation and the lowest standard deviation, with a corresponding upper limit of confidence coefficient set at 0.85. The worst working condition is a rainy day with no lighting and an uneven road surface, with a corresponding lower limit of confidence coefficient set at 0.15. For other environments, the confidence coefficient value is determined by linear interpolation based on the degree of deviation of their average deviation and standard deviation from the benchmark working condition.

[0092] Based on the above rules, the trust coefficients for all environments are calibrated, and an environment-trust coefficient mapping table is constructed to achieve a one-to-one correspondence between operating conditions and parameters. An example mapping relationship is as follows: Sunny daytime, smooth road surface: Trust factor 0.85; Cloudy weather, slightly uneven road surface: Trust level 0.65; Low-light underground parking garage with a smooth road surface: Trust factor 0.30; Backlighting, open-air parking lot: Trust factor 0.35; Rainy days, city roads: Trust factor 0.45; No lighting at night, park roads: Trust factor 0.20.

[0093] This mapping relationship transforms abstract environmental impacts into directly callable quantitative parameters, providing a standardized basis for quickly querying trust coefficients under real-time operating conditions.

[0094] During actual vehicle operation, it is necessary to perceive the external environmental conditions in real time, quickly match the corresponding trust coefficient, and ensure the real-time nature of fusion correction. Current environmental information is collected collaboratively by multiple onboard sensors, including a light sensor to detect ambient brightness and distinguish between strong light, weak light, and nighttime; a rain sensor to identify rainy and sunny days; an onboard camera to analyze backlight and shadow conditions; a positioning module to match site type with map data; and suspension sensors to perceive road surface smoothness, etc.

[0095] The signals output from various sensors are fused and analyzed to obtain standardized current environment labels, such as "low light in underground parking lot + smooth road surface". This label is used as an index for precise matching in a pre-built environment-trust coefficient mapping table, directly reading the corresponding trust coefficient value without the need for complex real-time calculations. The response latency is controlled within milliseconds, fully meeting the real-time requirements of in-vehicle systems.

[0096] This invention utilizes large-scale, multi-dimensional historical data statistics to objectively quantify the error distribution characteristics of visual detection under different environments, abandoning empirical and subjective parameter assignment methods. Based on error statistics, it establishes scientific trust coefficient calibration rules and constructs a standardized mapping table to achieve accurate and rapid matching of trust coefficients. Multi-sensor fusion sensing of the current environment ensures the accuracy and real-time performance of operational condition recognition. The overall solution provides strong data support and interpretability for trust coefficient generation, effectively improving the rationality of fusion weight allocation under complex operating conditions and providing a solid parameter foundation for highly accurate obstacle correction.

[0097] In some optional implementations, step S104 above includes: Step g1: Calculate the first height by multiplying the height of the first obstacle and the second parameter. The second parameter is obtained by summing the confidence coefficient with a preset scalar value. Step g2: Calculate the corrected height by multiplying the point cloud correction parameters and correction weights; Step g3: Determine the height of the second obstacle using the sum of the first height and the corrected height.

[0098] Specifically, such as Figure 2 As shown, the first obstacle height is the raw height value output in real time by the visual detector, which generally has a slight deviation. The second parameter is the amplification factor used for environmental compensation, which is obtained by summing the confidence coefficient and a preset scalar value. In this embodiment of the invention, the preset scalar value is fixed at 1 to ensure the basic amplification benchmark. The confidence coefficient β characterizes the reliability of visual detection in the current environment. The more severe the environment, the smaller the confidence coefficient, and the stronger the compensation of the laser point cloud, specifically offsetting the underestimation error of vision under complex working conditions.

[0099] In actual calculations, the second parameter = 1 + β, and the first height = H_occ × (1 + β). The point cloud correction parameter is the standard deviation of the laser point cloud height. The correction weight, γ, characterizes the stability of laser measurements. The correction weight reflects the current dependence on laser measurements, while the correction height is a fine-tuning amount provided by the laser point cloud to compensate for residual biases remaining after environmental compensation. The correction height is calculated as: Correction Height = γ × .

[0100] The second obstacle height is the final height value after integrating visual environment compensation and laser refinement correction, used for vehicle passability assessment and path planning. The calculation logic is as follows: Second obstacle height... =First altitude + Correction altitude, corresponding formula This achieves an organic combination of visual coarse compensation and laser fine correction.

[0101] This implementation method offsets fixed deviations caused by operating conditions through environmental compensation and achieves fine calibration using laser correction, forming a two-layer correction mechanism of "coarse compensation and fine correction." The fusion logic is simple and efficient, with clear physical meaning. It leverages the advantages of visual semantic perception while utilizing the accuracy of laser 3D measurement, dynamically adapting to different environments and visual reliability, resulting in a significant improvement in the final height accuracy. This effectively avoids collision risks caused by height underestimation, greatly enhancing the safety and reliability of intelligent parking and low-speed driving scenarios, and possesses extremely high engineering value and practical significance.

[0102] This embodiment also provides a detection height correction device for a visual detector, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0103] This embodiment provides a detection height correction device for a visual detector, such as... Figure 3 As shown, it includes: The false detection probability identification module 301 is used to calculate the false detection probability when the visual detector detects the height of an obstacle. The correction amount determination module 302 is used to determine the point cloud correction parameters for the obstacle height through point cloud data; The corrected weight module 303 is used to determine the corrected weights corresponding to the point cloud data based on the false detection probability and the current environment. The height correction module 304 is used to correct the first obstacle height output by the visual detector using correction weights and point cloud correction parameters to obtain the corrected second obstacle height.

[0104] In some optional implementations, the false detection probability identification module 301 includes: The height acquisition unit is used to collect height data of different obstacles through a visual detector. The indicator calculation unit is used to calculate the corresponding visual height mean and visual height standard deviation based on the height set data. The false detection probability calculation unit is used to input the mean visual height, standard deviation of visual height, and height of the first obstacle into an improved Gaussian distribution statistical model to calculate the false detection probability.

[0105] In some optional implementations, the correction amount determination module 302 includes: The point cloud acquisition unit is used to acquire point cloud data detected by radar equipment within its optimal observation range; The correction calculation unit is used to calculate the standard deviation of point cloud height by scanning the height values ​​of different obstacles using point cloud data, and uses the standard deviation of point cloud height as a point cloud correction parameter.

[0106] The detection height correction device for the visual detector provided in this embodiment of the invention can execute the detection height correction method for the visual detector provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0107] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0108] The following is a detailed reference. Figure 4This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0109] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0110] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the methods of the embodiments of the present invention.

[0111] Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0112] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0113] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0114] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method of correcting a detection height of a vision detector, characterized by, The method includes: The probability of false detection when a computer vision detector detects the height of an obstacle; Determine point cloud correction parameters for obstacle height using point cloud data; The correction weights corresponding to the point cloud data are determined based on the false detection probability and the current environment. The first obstacle height output by the visual detector is corrected using the correction weight and the point cloud correction parameter to obtain the corrected second obstacle height.

2. The method of claim 1, wherein, The false detection probability of the computational vision detector when detecting obstacle height includes: The visual detector collects height data of different obstacles. Calculate the corresponding visual height mean and visual height standard deviation based on the height set data; The mean visual height, the standard deviation of visual height, and the height of the first obstacle are input into an improved Gaussian distribution statistical model to calculate the false detection probability.

3. The method according to claim 2, characterized in that, The step of inputting the mean visual height, the standard deviation of visual height, and the height of the first obstacle into an improved Gaussian distribution statistical model to calculate the false detection probability includes: The false detection probability is calculated using the following improved Gaussian distribution statistical model. : In the formula, This represents the average visual height. Indicates the standard deviation of visual height. Indicates the height of the first obstacle. This indicates the maximum false positive rate for the height of the first obstacle. Sensitivity used to control the false detection probability offset.

4. The method according to claim 1, characterized in that, The step of determining point cloud correction parameters for obstacle height using point cloud data includes: Acquire the point cloud data detected by the radar equipment within its optimal observation range; The standard deviation of point cloud height is calculated by scanning the height values ​​of different obstacles using the point cloud data, and the standard deviation of point cloud height is used as the point cloud correction parameter.

5. The method according to claim 4, characterized in that, The acquisition of the point cloud data detected by the radar equipment within its optimal observation range includes: The point cloud sub-data collected by the radar equipment from the optimal observation range is recorded at different times. The point cloud data is obtained by stitching together the point cloud sub-data recorded at each time point.

6. The method according to claim 4, characterized in that, The step of calculating the standard deviation of point cloud height by scanning the height values ​​of different obstacles using the point cloud data includes: The ground is divided into grid maps based on the detection range of radar equipment; The point cloud data is filled into the grid map, and the median height of the point cloud in each cell of the grid map is calculated to obtain the height value for scanning different obstacles. The standard deviation of the height values ​​of different obstacles is calculated to obtain the standard deviation of the point cloud height.

7. The method according to claim 1, characterized in that, The step of determining the correction weights corresponding to the point cloud data based on the false detection probability and the current environment includes: The current environment query is used to characterize the level of trust in visual detection; The correction weight is determined by the ratio of the false detection probability to the first parameter, which is obtained by summing the confidence coefficient and the false detection probability.

8. The method according to claim 7, characterized in that, The step of querying a trust coefficient, which characterizes the level of trust in visual detection, through the current environment includes: The system acquires historical obstacle height data collected by the visual detector under different environmental conditions, and calculates the average height detection value and standard deviation of height detection under each environmental condition based on the historical obstacle height data. Based on the average height detection value and the standard deviation of height detection, a trust coefficient is set for each environment, and an environment-trust coefficient mapping table is formed; Obtain the current environment information and search for the trust coefficient that matches the current environment information in the environment-trust coefficient mapping table.

9. The method according to claim 7, characterized in that, The step of correcting the first obstacle height output by the visual detector using the correction weights and the point cloud correction parameters to obtain the corrected second obstacle height includes: The first height is calculated by multiplying the height of the first obstacle and the second parameter, where the second parameter is obtained by summing the confidence coefficient with a preset scalar value. The corrected height is calculated by multiplying the point cloud correction parameters and the correction weights; The height of the second obstacle is determined by the sum of the first height and the corrected height.

10. A detection height correction device for a visual detector, characterized in that, The device includes: The false detection probability identification module is used to calculate the false detection probability when the visual detector detects the height of an obstacle. The correction amount determination module is used to determine the point cloud correction parameters for obstacle height based on point cloud data; The corrected weight module is used to determine the corrected weights corresponding to the point cloud data based on the false detection probability and the current environment. The height correction module is used to correct the first obstacle height output by the visual detector using the correction weight and the point cloud correction parameter to obtain the corrected second obstacle height.

11. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the method of any one of claims 1 to 9.

Images