A method for simulating and testing laser radar point cloud position and intensity

Abstract

The application relates to a laser radar point cloud position and intensity simulation simulation and test method, which comprises the following steps: selecting a target plate of interest; selecting a laser radar to be modeled; collecting real point cloud data of the laser radar under different pose conditions by using a target plate point cloud statistical data collection rack; extracting position information and intensity information of the collected point cloud data; establishing and training a simulation point cloud single-point calculation model comprising an intensity model, a position model and a detection probability model according to the extracted information; performing machine learning on the single-point calculation model of the simulation point cloud and evaluating the curve fitting accuracy, and performing corresponding optimization; evaluating the calculation speed and the simulation point cloud accuracy of the simulation point cloud calculation model, and performing corresponding optimization to obtain a laser radar simulation model. Compared with the prior art, the application has the advantages of high consistency of point cloud intensity and point cloud ranging error simulation, high real-time performance and strong universality.

Description

Technical Field

[0001] This invention relates to the field of lidar point cloud data simulation technology, and in particular to a method for simulating and testing the position and intensity of lidar point clouds. Background Technology

[0002] With the development of intelligent connected vehicle technology, lidar, as a new type of sensor, is being used more and more widely in vehicle and roadside systems.

[0003] Currently, the main purpose of LiDAR applications is to detect, track, and predict the trajectories of road traffic participants, with the mainstream approach being based on deep learning theory. Therefore, training such methods requires a sufficient amount of high-quality datasets with good generalization performance. However, whether for vehicle-side or roadside applications, constructing datasets requires complex data acquisition systems, careful planning of data acquisition cycles, post-processing, and labeling. These tedious tasks make it difficult to guarantee the quantity and quality of the datasets, while also incurring significant costs.

[0004] Therefore, more and more experts and scholars in the industry are considering using simulation to expand LiDAR point cloud data and broaden the datasets that can be used for pre-training deep learning algorithms. Currently, there are three main methods for LiDAR modeling and simulation.

[0005] The first type of method is a simple model based on ray tracing algorithms. This type of model only simulates basic parameters such as the scanning beams, maximum detection distance, resolution, and scanning frequency of different LiDAR models, and outputs a 3D point cloud by combining it with a 3D scene model. Such models are only suitable for input to components such as autonomous driving planning and decision-making, and have the lowest simulation consistency.

[0006] The second type of method is a lidar model based on the principle of light attenuation. Building upon the first type of model, this type studies the transmission process of light in the medium, considering the beam attenuation patterns caused by atmospheric phenomena such as rain or fog, and then establishes point cloud missed detection and maximum detection distance. While this model is theoretically sound, the inevitable simplifications due to numerous environmental influencing factors result in low universality and complex configuration.

[0007] The third type of method is based on data-driven LiDAR models. These models directly construct point cloud maps of the real traffic environment using real-time localization and mapping methods. Based on this, voxel downsampling and normal estimation algorithms are used to reconstruct 3D meshes and 3D traffic participant models, and then point clouds are reconstructed through resampling. Because this type of model generates simulated point clouds based on real-world point cloud data, it naturally possesses high consistency. However, this method relies on the 3D reconstruction of the real scene and cannot fully leverage the flexibility of simulated point clouds. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a method for simulating and testing the position and intensity of LiDAR point clouds. It fully combines the advantages and disadvantages of physical modeling and data-driven modeling, constructing ideal and controllable data acquisition conditions to collect data on the real LiDAR and target board materials. Combined with the constructed simulated point cloud computing model, it focuses on reconstructing the noise characteristics of the real point cloud. Furthermore, by integrating a ray tracing-based simulation model, it enables flexible 3D traffic scene modeling and convenient simulation of various LiDAR models, providing a foundation for testing autonomous driving algorithms and constructing synthetic data for depth perception models. Synthetic data refers to simulated data with high consistency constructed using an autonomous driving simulation toolchain.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] A method for simulating and testing the position and intensity of lidar point clouds includes the following steps:

[0011] Select the target board of interest;

[0012] Select the lidar to be modeled, and use the target board point cloud data acquisition table to collect the real point cloud data of the lidar under different pose conditions. The pose conditions include the incident angle and the relative distance between the lidar and the target board.

[0013] The collected real point cloud data is preprocessed; plane fitting is performed based on the preprocessed real point cloud data to extract the relative incident angle between the laser beam and the target plate and the laser beam ranging. Based on the preset intensity modeling and ranging modeling targets, simulation is performed to obtain relevant data of different incident angles and intensities, as well as relevant data of different distances and ranging errors.

[0014] A simulation calculation is performed on a point on a target board generated by a single laser beam to generate a simulated point cloud single-point calculation model, which includes an intensity model, a position model, and a detection probability model. The machine learning parameters and curve fitting parameters in the simulated point cloud single-point calculation model are trained using relevant data obtained from real point cloud data.

[0015] The machine learning and curve fitting accuracy of the trained single-point computing model of the simulated point cloud is evaluated. If the fitting accuracy meets the preset accuracy requirements, the simulated point cloud computing model is obtained; otherwise, the fitting parameters are optimized.

[0016] The computation speed and accuracy of the simulated point cloud computing model are evaluated. If both meet the preset simulation requirements, it is used as the lidar simulation model. If the computation speed does not meet the simulation requirements, the simulated point cloud computing model is optimized. If the accuracy of the simulated point cloud does not meet the simulation requirements, the simulated point cloud computing model or the single-point computing model is optimized.

[0017] Furthermore, the material of the target board of interest can be different colored vehicle paint, different colored car cover material, clothing fabric, asphalt road surface, or cement road surface material.

[0018] Furthermore, the different pose conditions correspond to multiple different incident angles and different relative distances between the lidar and the target board. The multiple different incident angles include relative incident angles between the laser beam and the target board of 0°, 1°, 2°, 3°, 4°, 5°, 6°, 7°, 8°, 9°, 10°, 20°, 30°, 40°, 50°, 60°, 70°, 80°, 81°, 82°, 83°, 84°, 85°, 86°, 87°, 88°, and 89°; the different relative distances between the lidar and the target board include relative distances between the lidar and the target board of 10 meters, 20 meters, 30 meters, 40 meters, 50 meters, 60 meters, and 70 meters.

[0019] Furthermore, the preprocessing of the collected real point cloud data includes removing non-target board point clouds, removing abnormal points within the target board, and accumulating point clouds from multiple frames. It also determines whether the point cloud within the target board is too sparse or has point cloud gaps caused by occlusion. If so, the point cloud frame is deemed invalid, and data collection is restarted.

[0020] Furthermore, the expression for calculating the relative incident angle between the extracted laser beam and the target plate is as follows:

[0021]

[0022] In the formula, θ is the relative incident angle between the laser beam and the target plate, and the vector (x) T ,y T ,z T ) T Let T be the normal to the fitted plane, and let T be the fitted plane. The vector (x...) L ,y L ,z L ) T The direction of the laser beam;

[0023] The calculation expression for the distance measurement of the extracted laser beam is as follows:

[0024] d=||(x L ,y L ,z L )T ||

[0025] In the formula, d represents the distance measured by the laser beam.

[0026] Furthermore, the expression for the intensity model is:

[0027]

[0028]

[0029] In the formula, I represents the final calculated strength. B Based on the Disney bidirectional reflectance distribution function, the input parameters R, G, and B are the color values ​​of the target board measured by a photometer. Ro, Me, and Sp are the roughness, metallicity, and specularity parameters of the target board for BRDF, which are the parameters that need to be trained for subsequent machine learning and curve fitting. C The compensation term consists of the fitted Gaussian kernel function, N. i σ i For the parameters that need to be trained later, I N The noise term consists of random numbers that follow a fitted Gaussian distribution, and σ represents the parameters that need to be trained later.

[0030] The expression for the location model is:

[0031]

[0032]

[0033] In the formula, [α0β0D0] T These are the theoretical polar coordinates of the point generated by the reflection of a single laser beam in the simulation environment. [γ] N γ N D N ] T For noise term, γ N The noise is angular, consisting of random numbers that follow a Gaussian distribution with empirically set parameters, D. N The distance noise is represented by random numbers that follow a Gaussian distribution based on the fitted parameters; [α β D] T These are the polar coordinates of the points generated in the simulation environment after adding noise, which are then converted to Cartesian coordinates [xyz]. T

[0034] The detection probability model is the probability that the currently generated point is missing. It is obtained based on the intensity model through a mapping function, which is obtained by fitting.

[0035] Furthermore, the optimization objective during the training process of the simulated point cloud single-point calculation model is:

[0036]

[0037] In the formula, param represents all parameters to be fitted or trained, and D... f The feasible region is Color, which is the set of colors of all the target boards being tested. S(R,G,B,param) refers to the relevant parameter vector of the point sequence generated in the virtual environment at different incident angles and distances. G(R,G,B,param) refers to the relevant parameter vector of the point sequence at different incident angles and distances corresponding to S(R,G,B,param) in the actual collected data.

[0038] Furthermore, the evaluation metrics for the fitting accuracy include histogram distribution similarity and / or point cloud cluster-level accuracy with histogram distribution similarity as the core.

[0039] Furthermore, in the process of evaluating the computation speed and accuracy of the simulated point cloud computing model, for computation speed, the same resolution as the target lidar is configured, the simulated point cloud is encoded in pointcloud2 format, and output to the simulated point cloud computing model under test using shared memory, and its highest output frequency is counted; for output accuracy, a three-stage point cloud accuracy evaluation standard is adopted to evaluate the simulation accuracy of the simulated point cloud generated by the simulated point cloud computing model at the point level, target level, and twin scene level.

[0040] Furthermore, the lidar includes mechanically rotating lidar, MEMS scanning lidar, or non-repetitive scanning lidar.

[0041] Furthermore, the RANSAC algorithm is used to perform plane fitting on the preprocessed real point cloud data.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] (1) High consistency of point cloud intensity simulation. This method uses real-world data collected under standard conditions as the original data basis for point cloud modeling, ensuring data authenticity and reliability. BRDF (Bidirectional Reflectance Distribution Function) and Gaussian kernel function are used to iteratively fit the point cloud intensity variation and noise level under different pose conditions, achieving high fitting accuracy. The point cloud simulation based on the above-mentioned simulated point cloud model achieves high consistency with real data.

[0044] (2) High consistency of point cloud ranging error simulation. This method uses real-world data collected under standard conditions as the original data basis for point cloud modeling, ensuring data authenticity and reliability. Based on the assumption that ranging noise conforms to Gaussian white noise, the standard deviation of ranging error under different real distances is estimated, resulting in high fitting accuracy. Point cloud simulation based on the above model can achieve high consistency with real data.

[0045] (3) High real-time performance of point cloud simulation. The computational pipeline of this method is based on the BRDF model of Disney principle, with the addition of accurate models for intensity and ranging error. The newly introduced computational load is at the constant level. With the aid of a graphics card, it can realize real-time simulation of high-resolution mainstream lidar point clouds.

[0046] (4) The point cloud modeling method is highly versatile and scalable. The core intensity and ranging models of this method adopt a data-driven approach, with high standardization, automation, and efficiency in data acquisition. Furthermore, it is not limited by the type of lidar used, and the model construction method is universal. In point cloud simulation, the ray tracing-based simulation pipeline is applicable to various 3D rendering components and has strong scalability. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a method for simulating and testing the position and intensity of a lidar point cloud, as provided in an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the processing results of a lidar point cloud position and intensity simulation and testing method provided in an embodiment of the present invention. Detailed Implementation

[0049] 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, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0050] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0051] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0052] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed during use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0053] It should be noted that 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. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0054] Furthermore, terms such as "horizontal" and "vertical" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0055] Example 1

[0056] This embodiment provides a method for simulating and testing the position and intensity of lidar point clouds, including the following steps:

[0057] Select the target board of interest;

[0058] Select the lidar to be modeled, and use the target board point cloud data acquisition table to collect the real point cloud data of the lidar under different pose conditions. The pose conditions include the incident angle and the relative distance between the lidar and the target board.

[0059] The collected real point cloud data is preprocessed; plane fitting is performed based on the preprocessed real point cloud data to extract the relative incident angle between the laser beam and the target plate and the laser beam ranging. Based on the preset intensity modeling and ranging modeling targets, simulation is performed to obtain relevant data of different incident angles and intensities, as well as relevant data of different distances and ranging errors.

[0060] A simulation calculation is performed on a point on a target board generated by a single laser beam to generate a simulated point cloud single-point calculation model, which includes an intensity model, a position model, and a detection probability model. The machine learning parameters and curve fitting parameters in the simulated point cloud single-point calculation model are trained using relevant data obtained from real point cloud data.

[0061] The machine learning and curve fitting accuracy of the trained simulation point cloud single-point computing model is evaluated. If the fitting accuracy meets the preset accuracy requirements, the simulation point cloud computing model is obtained; otherwise, the fitting parameters are optimized.

[0062] The computation speed and accuracy of the simulated point cloud computing model are evaluated. If both meet the preset simulation requirements, it is used as the LiDAR simulation model. If the computation speed does not meet the simulation requirements, the simulated point cloud computing model is optimized. If the accuracy of the simulated point cloud does not meet the simulation requirements, the simulated point cloud computing model or the single-point computing model is optimized.

[0063] This method first establishes ideal and controllable data acquisition conditions to collect data on real LiDAR and target board materials. Then, using BRDF and Gaussian kernel functions, it focuses on reconstructing the noise characteristics of the real point cloud. Further, by combining this with a ray tracing simulation model, it enables flexible 3D traffic scene modeling and convenient simulation of various LiDAR models, providing a foundation for testing autonomous driving algorithms and constructing synthetic data for depth perception models.

[0064] like Figure 1 As shown, in specific implementation, the above method includes the following steps:

[0065] (100) Target selection: Select a specific target of interest and collect real point cloud data of the target. The target can be made of materials that are of interest in the application, such as different colors of vehicle paint, different colors of car cover, common clothing fabrics, asphalt or cement road surfaces, etc.

[0066] (200) Real Point Cloud Acquisition of Target Board: Select a specific type of LiDAR to be modeled, and use a target board point cloud statistical data acquisition platform to collect real point cloud data of the LiDAR under different pose conditions. The different pose conditions include 27 incident angles between the laser beam and the target board: 0°, 1°, 2°, 3°, 4°, 5°, 6°, 7°, 8°, 9°, 10°, 20°, 30°, 40°, 50°, 60°, 70°, 80°, 81°, 82°, 83°, 84°, 85°, 86°, 87°, 88°, and 89°. The relative distances between the LiDAR and the target board are 7: 10 meters, 20 meters, 30 meters, 40 meters, 50 meters, 60 meters, and 70 meters. For each specific pose condition, the LiDAR operating frequency is set to 10Hz, and 300 frames of point cloud data are acquired.

[0067] Optionally, the lidar includes, but is not limited to, mechanical rotating type, MEMS scanning type, and non-repetitive scanning type, and the lidar operating wavelength includes 905 nanometers and 1550 nanometers.

[0068] The list of different relative poses includes, but is not limited to, the incident angles and distances explicitly listed above.

[0069] (300) After acquiring the collected real lidar point cloud data, preprocessing is performed. Preprocessing operations include point cloud removal from non-target boards, removal of outliers within the target board, accumulation of point clouds across multiple frames, and plane fitting based on the RANSAC algorithm. Let the fitting plane be T, and the laser beam be L, then the normal to the plane can be represented as a vector (x... T ,y T ,z T ) T The direction of the laser beam can be represented as a vector (x). L ,y L ,z L ) T Therefore, the relative angle of incidence between the laser beam and the target plate is:

[0070]

[0071] The range of the laser beam is

[0072] D0=||(x L ,y L ,z L ) T ||

[0073] Using the above formulas, the point clouds on all target boards are processed to extract their incident angles and distances. For both intensity modeling and ranging modeling targets, relevant data on different incident angles and intensities, as well as relevant data on different distances and ranging errors, are compiled.

[0074] (400) Determine if there is any unprocessable point cloud data. If there is a phenomenon such as the point cloud in the target board being too sparse or the point cloud being missing due to occlusion, then the point cloud frame is determined to be invalid and data acquisition needs to be carried out again. Go to (100).

[0075] (500) Simulate and calculate a single point on a target plate generated by a single laser beam, producing a simulated point cloud single-point calculation model including an intensity model, a position model, and a detection probability model. The intensity model is...

[0076]

[0077]

[0078] Where I is the final calculated strength, I BThe results show the calculation of a BRDF based on the Disney Bidirectional Reflection Distribution Function (BRDF) (Physically Based Shading at Disney, Brent Burley, Walt Disney Animation Studios). The improved BRDF differs from the original Disney BRDF in that it is simplified, reducing the number of required parameters. The input parameters R, G, and B are the color values ​​of the target board measured by a photometer. Ro, Me, and Sp are the target board roughness, metallicity, and specularity parameters for the BRDF, which are the parameters required for subsequent machine learning and curve fitting training. C The compensation term consists of the fitted Gaussian kernel function, N. i σ i These are the parameters that need to be trained later. N The noise term is a random number that follows a fitted Gaussian distribution, and σ is a parameter that needs to be trained later.

[0079] Location model is

[0080]

[0081] [α0 β0 D0] T These are the theoretical polar coordinates of the point generated by the reflection of a single laser beam in the simulation environment. [γ] N γ N D N ] T For noise term, γ N The noise is angular, consisting of random numbers that follow a Gaussian distribution with empirically set parameters, D. N The distance noise is represented by random numbers that follow a Gaussian distribution with the fitted parameters set. [α β D] T These are the polar coordinates of the points generated in the simulation environment after adding noise, which are then converted to Cartesian coordinates [xyz]. T .

[0082] The detection probability is the probability that the currently generated point is missing. This probability is related to the intensity value, and the correlation mapping function is obtained by fitting.

[0083] (600) Use the gradient descent algorithm and real point cloud data to train the parameters required for curve fitting and the neural network parameters. The training optimization objective of the cost function is...

[0084]

[0085] Where param represents all the parameters to be fitted or trained mentioned above, and Df The feasible region is Color, which is the set of colors of all the target boards being tested. S(R,G,B,param) refers to the relevant parameter vector of the point sequence generated in the virtual environment at different incident angles and distances. G(R,G,B,param) refers to the relevant parameter vector of the point sequence at different incident angles and distances corresponding to S(R,G,B,param) in the actual collected data.

[0086] (700) Evaluation of Machine Learning and Curve Fitting Accuracy: The ranging and intensity characteristics of point cloud simulations can be represented by histograms. Therefore, histogram distribution similarity is introduced as a core evaluation index. Under specific conditions, assuming the actual ranging sequence is r, the simulated ranging sequence is r′, and |r|=|r′|=N, then the point cloud distance simulation accuracy is: Under specific conditions, assuming the actual intensity sequence is i, the simulated ranging sequence is i′, and |i|=|i′|=N, then the point cloud distance simulation accuracy is:

[0087] Among them, the accuracy evaluation indicators for lidar point cloud simulation include, but are not limited to, histogram similarity and point cloud cluster-level accuracy evaluation indicators based on histogram similarity.

[0088] (800) Determine if the fitting accuracy meets the requirements. If the accuracy meets the requirements, proceed to (900). If the accuracy does not meet the requirements, proceed to (801).

[0089] (801) Determine whether the machine learning hyperparameter settings are reasonable. If they are not reasonable, proceed to (802). If they are reasonable, the single-point calculation model needs to be adjusted. Proceed to (803).

[0090] (802) Adjust the machine learning hyperparameters, then perform machine learning and curve fitting again, and go to (600).

[0091] (803) Optimize the single-point calculation model, improve the intensity model, location model and detection probability model to improve simulation accuracy or increase calculation speed, go to (500).

[0092] (900) A three-dimensional digital twin scene was constructed based on the actual data acquisition environment, and the target board for acquisition and the test vehicle were modeled. Five simulated laser beams centered at the angle of the real laser beam were used to simulate the divergence effect of the laser beam. The angles of all emitted beams of the lidar were set, and the intersection points and reflections of each emitted laser beam with the scene objects were calculated using a ray tracing algorithm to obtain an ideal point cloud. Then, the ideal point cloud was processed using the simulated point cloud single-point calculation model to generate the simulated point cloud of the scene.

[0093] (1000) To evaluate the computational speed and accuracy of the simulated point cloud, it is necessary to simultaneously ensure the output frequency and accuracy of the point cloud when conducting lidar point cloud simulation for testing in autonomous driving systems. The computational speed threshold is selected when the highest output frequency of the point cloud simulation system is equal to 10Hz. The distance and intensity accuracy of the point cloud simulation are evaluated using histogram distribution similarity. Under specific conditions, assuming the actual ranging sequence is r, the simulated ranging sequence is r′, and |r|=|r ′ If |=N, then the point cloud distance simulation accuracy is

[0094]

[0095] Under specific conditions, assume the actual mining intensity sequence is i, the simulated ranging sequence is i′, and |i|=|i ′ If |=N, then the point cloud intensity simulation accuracy is

[0096]

[0097] (1100) Determine if the calculation speed meets the requirements. If the calculation speed is higher than the threshold, it meets the simulation requirements. Go to (1200) otherwise go to (1101).

[0098] (1101) Optimize the overall calculation model and optimize the calculation method for generating simulated point clouds in the scene to improve the calculation speed or improve the calculation accuracy, and then repeat (900).

[0099] (1200) Determine if the accuracy of the simulated point cloud meets the requirements. If it does not meet the requirements, proceed to (1201). If it meets the requirements, end the overall process.

[0100] (1201) Determine if the calculation speed meets the requirements. If it does not meet the threshold, proceed to (1101). If it meets the threshold, proceed to (803).

[0101] The data processing flow implemented in this embodiment is as follows: Figure 2 As shown in the figure, the process is divided into data acquisition and processing, intensity and ranging error distribution fitting, 3D scene construction, lidar model construction and calculation, point cloud data post-processing and output, and point cloud simulation accuracy evaluation.

[0102] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for simulating and testing the position and intensity of lidar point clouds, characterized in that, Includes the following steps: Select the target board of interest; Select the lidar to be modeled, and use the target board point cloud data acquisition table to collect the real point cloud data of the lidar under different pose conditions. The pose conditions include the incident angle and the relative distance between the lidar and the target board. The collected real point cloud data is preprocessed; plane fitting is performed based on the preprocessed real point cloud data to extract the relative incident angle between the laser beam and the target plate and the laser beam ranging. Based on the preset intensity modeling and ranging modeling target, simulation is performed to obtain relevant data of different incident angles and intensities, as well as relevant data of different distances and ranging errors. A simulation calculation is performed on a point on a target board generated by a single laser beam to generate a simulated point cloud single-point calculation model, which includes an intensity model, a position model, and a detection probability model. The machine learning parameters and curve fitting parameters in the simulated point cloud single-point calculation model are trained using relevant data obtained from real point cloud data. The machine learning and curve fitting accuracy of the trained single-point computing model of the simulated point cloud is evaluated. If the fitting accuracy meets the preset accuracy requirements, the simulated point cloud computing model is obtained; otherwise, the fitting parameters are optimized. The computation speed and accuracy of the simulated point cloud computing model are evaluated. If both meet the preset simulation requirements, it is used as the lidar simulation model. If the computation speed does not meet the simulation requirements, the simulated point cloud computing model is optimized. If the accuracy of the simulated point cloud does not meet the simulation requirements, the simulated point cloud computing model or the single-point computing model is optimized.

2. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The target board of interest is made of different colored vehicle paint, different colored car cover material, clothing fabric, asphalt road surface, or cement road surface material.

3. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The different pose conditions correspond to multiple different incident angles and different relative distances between the lidar and the target board. The multiple different incident angles include the relative incident angles between the laser beam and the target board of 0°, 1°, 2°, 3°, 4°, 5°, 6°, 7°, 8°, 9°, 10°, 20°, 30°, 40°, 50°, 60°, 70°, 80°, 81°, 82°, 83°, 84°, 85°, 86°, 87°, 88°, and 89°. The different relative distances between the lidar and the target board include relative distances of 10 meters, 20 meters, 30 meters, 40 meters, 50 meters, 60 meters, and 70 meters.

4. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The preprocessing of the collected real point cloud data includes removing point clouds from non-target boards, removing abnormal points within the target board, and accumulating point clouds from multiple frames. It also determines whether the point cloud within the target board is too sparse or has point cloud gaps caused by occlusion. If so, the point cloud frame is deemed invalid, and data collection is restarted.

5. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The expression for calculating the relative incident angle between the extracted laser beam and the target plate is as follows: In the formula, θ is the relative incident angle between the laser beam and the target plate, and the vector (x) T ,y T ,z T ) T Let T be the normal to the fitted plane, and let T be the fitted plane. The vector (x...) L ,y L ,z L ) T The direction of the laser beam; The calculation expression for the distance measurement of the extracted laser beam is as follows: d = || (x L ,y L ,z L ) T || In the formula, d represents the distance measured by the laser beam.

6. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The expression for the intensity model is: In the formula, I represents the final calculated strength. B Based on the Disney bidirectional reflectance distribution function, the input parameters R, G, and B are the color values ​​of the target board measured by a photometer. Ro, Me, and Sp are the roughness, metallicity, and specularity parameters of the target board for BRDF, which are the parameters that need to be trained for subsequent machine learning and curve fitting. C The compensation term consists of the fitted Gaussian kernel function, N. i σ i For the parameters that need to be trained later, I N The noise term consists of random numbers that follow a fitted Gaussian distribution, and σ represents the parameters that need to be trained later. The expression for the location model is: In the formula, [α0 β0 D0] T These are the theoretical polar coordinates of the point generated by the reflection of a single laser beam in the simulation environment, [γ N γ N D N ] T For noise term, γ N The noise is angular, consisting of random numbers that follow a Gaussian distribution with empirically set parameters, D. N The distance noise is represented by random numbers that follow a Gaussian distribution based on the fitted parameters; [α β D] T These are the polar coordinates of the points generated in the simulation environment after adding noise, which are then converted to Cartesian coordinates [xyz]. T ; The detection probability model is the probability that the currently generated point is missing. It is obtained based on the intensity model through a mapping function, which is obtained by fitting.

7. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The optimization objective during the training process of the simulated point cloud single-point calculation model is: s.t.param∈D f In the formula, param represents all parameters to be fitted or trained, and D... f The feasible region is Color, which is the set of colors of all the target boards being tested. S(R,G,B,param) refers to the relevant parameter vector of the point sequence generated in the virtual environment at different incident angles and distances. G(R,G,B,param) refers to the relevant parameter vector of the point sequence at different incident angles and distances corresponding to S(R,G,B,param) in the actual collected data.

8. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The evaluation metrics for fitting accuracy include histogram distribution similarity and / or point cloud cluster-level accuracy with histogram distribution similarity as the core.

9. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, In evaluating the computation speed and accuracy of the simulated point cloud computing model, for computation speed, the same resolution as the target lidar is configured, the simulated point cloud is encoded in pointcloud2 format, and output to the simulated point cloud computing model under test using shared memory, and its highest output frequency is counted; for output accuracy, a three-stage point cloud accuracy evaluation standard is used to evaluate the simulation accuracy of the simulated point cloud generated by the simulated point cloud computing model at the point level, target level, and twin scene level.

10. The method for simulating and testing the position and intensity of lidar point clouds according to claim 1, characterized in that, The lidar includes mechanically rotating lidar, MEMS scanning lidar, or non-repetitive scanning lidar.

Images