Method, device, storage medium, electronic device and computer program product for optimizing a target detection model of a lidar

Abstract

A method, device, storage medium, electronic device and computer program product for optimizing a target detection model of a laser radar. The method comprises: performing domain bias processing on first point cloud data of a source domain to obtain second point cloud data satisfying a preset target domain line number condition; inputting third point cloud data of a target domain to a pre-trained target detection model to obtain initial pseudo labels of the third point cloud data; screening the initial pseudo labels to obtain screened pseudo labels satisfying a preset screening condition; and iteratively training the pre-trained target detection model according to the second point cloud data and true labels corresponding to the second point cloud data, the third point cloud data and the screened pseudo labels corresponding to the third point cloud data.

Description

Technical Field

[0001] This application relates to the field of target detection model technology, and more specifically, to a method, apparatus, storage medium, electronic device, and computer program product for optimizing a target detection model of a lidar. Background Technology

[0002] 3D object detection algorithms based on LiDAR point clouds can be divided into point-based methods and mesh-based methods. Point-based methods, such as PointNet or PointNet++, directly extract features from unordered raw data to generate 3D candidate boxes, preserving the original geometric features of the point cloud, but are very time-consuming when processing large-scale outdoor scene data. Mesh-based methods, such as SECOND and PointPillars, use structured representations to quantize LiDAR data into fixed-size voxel or cylindrical meshes, and then use this representation to further extract semantic features for 3D object detection. They are typically used in scenarios with high real-time requirements, such as autonomous driving.

[0003] However, in practical engineering applications, due to cost considerations, training is often performed on public datasets of high-line-count LiDARs, and deployment is carried out on low-line-count LiDARs. Different LiDAR models generate point cloud densities and distributions with varying line counts, leading to a decrease in detection performance when the model is transferred between different data domains. Re-collecting and labeling data for each LiDAR model on different unmanned systems to address the domain shift problem during training would increase annotation costs and time, and lacks scalability. Alibaba DAMO Academy proposed using low-line-count LiDAR data to simulate and reconstruct high-line-count LiDAR data, applicable to scenarios such as 3D environment reconstruction, but not suitable for real-time high-performance point cloud 3D target detection tasks. More research approaches are based on unsupervised domain adaptation machine learning theory, transferring the trained model from the source domain data to the unlabeled target domain.

[0004] Unsupervised domain adaptation is a machine learning algorithm that addresses the distributional offset between the source and target domains. By learning domain-invariant features of both the source and target domains, it applies a classifier learned from the source domain to the target domain, even when the target domain has few or no labels. In recent years, research on unsupervised adaptive algorithms has largely focused on 2D images; less attention has been paid to 3D object detection, particularly unsupervised adaptive training algorithms for LiDAR. ST3D redesigns its self-training algorithm framework for 3D object detection tasks based on a 2D unsupervised adaptive algorithm. While ST3D uses data augmentation techniques such as random scaling to reduce the domain bias between the source and target domains, it does not fully consider the differences in scanning methods, density distribution, and background noise in LiDAR point clouds.

[0005] Existing unsupervised adaptive training methods for 3D object detection aim to improve the performance of 3D object detection algorithms in domain adaptive learning from different dimensions such as raw data processing, domain feature alignment, and pseudo-label denoising. This provides inspiration and technical support for optimizing the cross-domain object detection performance of LiDAR. However, these methods do not fully incorporate the imaging characteristics of LiDAR to design targeted and easy-to-operate parameter domain alignment methods.

[0006] The content of the background section is merely technology known to the public and does not necessarily represent existing technology in the field. Summary of the Invention

[0007] According to one aspect of this application, this application provides a method for optimizing a target detection model of a lidar radar. The method includes: performing domain bias processing on first point cloud data in the source domain to obtain second point cloud data that meets a preset target domain line count condition; inputting third point cloud data in the target domain into a pre-trained target detection model to obtain initial pseudo-labels for the third point cloud data; filtering the initial pseudo-labels to obtain filtered pseudo-labels that meet preset filtering conditions; and iteratively training the pre-trained target detection model based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo-label.

[0008] According to some embodiments of this application, performing domain bias processing on the first point cloud data of the source domain to obtain second point cloud data that meets the preset target domain line number conditions includes: filtering the first point cloud data to obtain filtered point cloud data that meets the preset target domain lidar vertical field of view; determining the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions based on the scanning resolution of the target domain lidar and the filtered point cloud data; and determining the second point cloud data based on the equivalent distribution point cloud data.

[0009] According to some embodiments of this application, initial pseudo-labels are filtered to obtain filtered pseudo-labels that meet preset filtering conditions, including: determining the scoring result of the initial pseudo-labels according to a preset scoring algorithm; deleting the corresponding initial pseudo-labels if the scoring result is less than a preset low scoring threshold; or retaining the corresponding initial pseudo-labels if the scoring result is greater than a preset high scoring threshold, thereby obtaining filtered pseudo-labels; or determining the clustering result of the third point cloud data according to a ground point filtering algorithm and a target clustering algorithm if the scoring result is less than or equal to a preset high scoring threshold and greater than or equal to a preset low scoring threshold; matching the corresponding initial pseudo-labels with the clustering results; and retaining the successfully matched initial pseudo-labels, thereby obtaining filtered pseudo-labels.

[0010] According to some embodiments of this application, determining the clustering result of the third point cloud data based on the ground point filtering algorithm and the target clustering algorithm includes: filtering the third point cloud data to obtain preliminary screened point cloud data within a preset selected range; sorting the preliminary screened point cloud data in order of height values ​​to obtain a preset number of seed point data; determining the current candidate ground plane step; determining the ground plane step for the next iteration; executing the step of determining the ground plane step for the next iteration according to a preset number of iterations to obtain a set of non-ground points of the point cloud data within the selected range; and performing clustering processing on the set of non-ground points based on a deep clustering algorithm to obtain the clustering result. The step of determining the current candidate ground plane includes: determining the difference between the height values ​​of the preliminary screened point cloud data and the average height of the preset number of seed point data; determining points with differences less than a preset height threshold as current candidate ground points; traversing all the preliminary screened point cloud data to obtain all current candidate ground points; and determining the current ground plane corresponding to the current candidate ground point based on all current candidate ground points. The steps for determining the ground plane for the next iteration include: determining the distance between the initial screening point cloud data and the current ground plane; identifying points whose distance is less than a preset distance threshold as candidate ground points for the next iteration; traversing all the initial screening point cloud data to determine all candidate ground points for the next iteration; and determining the ground plane for the next iteration based on all candidate ground points for the next iteration.

[0011] According to some embodiments of this application, a clustering process is performed on a set of non-ground points based on a deep clustering algorithm to obtain clustering results. This includes: determining an initial depth map matrix, initial horizontal and vertical angle matrices, and an initial clustering label matrix for the non-ground point set based on the number of lines and horizontal resolution of the LiDAR corresponding to the target domain; determining the depth map matrix; determining the horizontal and vertical angle matrices; determining the clustering label matrix based on the four-neighborhood algorithm, using the depth map matrix and the horizontal and vertical angle matrices; and determining the clustering results based on the clustering label matrix. The step of determining the depth map matrix includes: determining the index position of the non-ground point in the initial depth map matrix based on the channel information and horizontal angle of the non-ground point in the set; determining the pixel value corresponding to the index position based on the distance between the non-ground point and the LiDAR corresponding to the target domain; and traversing all non-ground points in the set to determine the depth map matrix. The step of determining the horizontal and vertical angle matrices includes: determining the horizontal and vertical angles corresponding to points in the depth map matrix based on a preset horizontal and vertical angle algorithm; and traversing all points in the depth map matrix to obtain all horizontal and vertical angles, thereby obtaining the horizontal and vertical angle matrix.

[0012] According to some embodiments of this application, the pre-trained target detection model is iteratively trained based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo label, including: according to a preset number of training rounds, the pre-trained target detection model is cross-iteratively trained based on the second point cloud data and the corresponding true label, the third point cloud data and the pseudo label corresponding to the third point cloud data that meets the preset filtering conditions.

[0013] According to one aspect of this application, an apparatus for optimizing a target detection model of a lidar radar is provided. The apparatus may include a domain bias processing module, an initial pseudo-label acquisition module, a pseudo-label filtering module, and a joint optimization module. The domain bias processing module performs domain bias processing on first point cloud data in the source domain to obtain second point cloud data that meets a preset target domain line count condition. The initial pseudo-label acquisition module inputs third point cloud data in the target domain into a pre-trained target detection model to obtain initial pseudo-labels for the third point cloud data. The pseudo-label filtering module filters the initial pseudo-labels to obtain filtered pseudo-labels that meet preset filtering conditions. The joint optimization module iteratively trains the pre-trained target detection model based on the second point cloud data, the corresponding true labels, the third point cloud data, and the filtered pseudo-labels.

[0014] According to another aspect of this application, this application also provides a non-volatile computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is capable of implementing the method for optimizing the target detection model of lidar as described above.

[0015] According to another aspect of this application, this application also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement the method for optimizing the target detection model of the lidar as described above.

[0016] According to another aspect of this application, this application also provides a computer program product, comprising: a computer program stored on a computer-readable storage medium; the computer program includes program instructions that, when executed by a computer, cause the computer to perform the method for optimizing the target detection model of a lidar as described above.

[0017] The technical solution of this application can obtain second point cloud data that meets the preset target domain line count condition by performing domain bias processing on the first point cloud data of the source domain. The technical solution of this application can input the third point cloud data of the target domain into a pre-trained object detection model to obtain initial pseudo-labels for the third point cloud data. The technical solution of this application can filter the initial pseudo-labels to obtain filtered pseudo-labels that meet preset filtering conditions. The technical solution of this application can iteratively train the pre-trained object detection model using the second point cloud data and its corresponding true label, the third point cloud data and its corresponding filtered pseudo-labels.

[0018] The technical solution of this application can construct source domain point cloud data with equivalent line count to the target domain's LiDAR by performing domain bias processing on the source domain point cloud data, thereby achieving domain alignment between the source domain point cloud data and the target domain point cloud data in terms of line count and distribution. The technical solution of this application uses source domain data that has undergone domain bias processing, which can better assist in the joint optimization of the target domain's target detection model performance and improve the unsupervised training effect of the target domain. Attached Figure Description

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

[0020] Figure 1 A flowchart illustrating method 1000 according to an embodiment of this application is shown; Figure 2 A flowchart illustrating step S100 according to an embodiment of this application is shown; Figure 3 A flowchart illustrating step S300 according to an embodiment of this application is shown; Figure 4 A flowchart illustrating step S340 according to an embodiment of this application is shown; Figure 5 A flowchart illustrating step S343 according to an embodiment of this application is shown; Figure 6 A flowchart illustrating step S344 according to an embodiment of this application is shown; Figure 7 A flowchart illustrating step S346 according to an embodiment of this application is shown; Figure 8 A flowchart illustrating step S3462 according to an embodiment of this application is shown; Figure 9 A flowchart illustrating step S3463 according to an embodiment of this application is shown; Figure 10 A schematic diagram showing the horizontal and vertical angles according to an embodiment of this application; Figure 11 A schematic diagram of the structure of an apparatus according to an embodiment of this application is shown.

[0021] Explanation of reference numerals in the attached figures: Device 20; Domain bias processing module 21; Initial pseudo-label acquisition module 22; Pseudo-label filtering module 23; Joint optimization module 24. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this application will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0023] The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of these specific details, or other methods, components, materials, devices, etc. In these cases, well-known structures, methods, devices, implementations, materials, or operations will not be shown or described in detail.

[0024] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0025] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, rather than to describe a specific order.

[0026] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] The English terms used in this application, their full English names, and their corresponding Chinese definitions are as follows: 3D, Three-Dimensional.

[0028] PointNet is a point cloud processing network.

[0029] SECOND stands for Sparsely Embedded Convolutional Detection.

[0030] Point Pillars.

[0031] IoU, Intersection over Union, is a term referring to the comparison of intersections over unions.

[0032] 2D, Two-Dimensional.

[0033] ST3D, Self-training for Unsupervised Domain Adaptation on 3D Object Detection.

[0034] See Figure 11 The apparatus 20 for optimizing the target detection model of lidar provided in this application may include a domain bias processing module 21, an initial pseudo-label acquisition module 22, a pseudo-label screening module 23, and a joint optimization module 24.

[0035] According to one aspect of this application, this application provides a method 1000 for optimizing a target detection model of a lidar.

[0036] See Figure 1 Method 1000 may include steps S100-S400.

[0037] In step S100, the first point cloud data of the source domain is subjected to domain deviation processing to obtain the second point cloud data that meets the preset target domain line number condition.

[0038] According to the example embodiment, the source domain can be a dataset acquired by a LiDAR for pre-training the target detection model. The first point cloud data can be a 3D point cloud dataset from the source domain. The first point cloud data can be labeled with true labels. Each point cloud data in the first point cloud data can include polar coordinate data such as the measured distance from the point to the LiDAR, the vertical angle of the LiDAR, the horizontal rotation angle of the LiDAR, and the horizontal offset angle of the channel.

[0039] The target domain can be the dataset acquired by LiDAR for unsupervised training of the target detection model. The number of lines of the LiDAR corresponding to the target domain can be different from the number of lines of the LiDAR corresponding to the source domain.

[0040] The preset target domain line count condition can be an equivalent line count condition to the line count of the LiDAR corresponding to the target domain. The second point cloud data can be a source domain 3D point cloud dataset with a line count equivalent to the LiDAR corresponding to the source domain and the LiDAR corresponding to the target domain.

[0041] For example, in step S100, the domain deviation processing module 21 performs domain deviation processing on the first point cloud data of the source domain to obtain the second point cloud data that meets the preset target domain line number condition.

[0042] The domain deviation processing module 21 filters the first point cloud data of the source domain to obtain filtered point cloud data that meets the preset target domain lidar vertical field of view. Based on the scanning resolution of the target domain lidar and the filtered point cloud data, the domain deviation processing module 21 determines the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions. Based on the equivalent distribution point cloud data, the domain deviation processing module 21 determines the second point cloud data.

[0043] In step S200, the third point cloud data of the target domain is input into the pre-trained target detection model to obtain the initial pseudo-label of the third point cloud data.

[0044] According to the example embodiment, the third point cloud data can be a 3D point cloud dataset of the target domain. The third point cloud data is unlabeled. The pre-trained object detection model can be a 3D point cloud detection model trained on source domain data after domain bias processing.

[0045] A pre-trained object detection model can perform inference and post-processing on the input third-point cloud data to obtain prediction results. The initial pseudo-labels can be the prediction results output by the pre-trained object detection model.

[0046] For example, in step S200, the initial pseudo-label acquisition module 22 inputs the third point cloud data of the target domain into the pre-trained target detection model to obtain the initial pseudo-label of the third point cloud data.

[0047] In step S300, the initial pseudo-labels are filtered to obtain filtered pseudo-labels that meet the preset filtering conditions.

[0048] According to the example embodiment, the preset filtering conditions can be a scoring threshold condition that satisfies the initial pseudo-label and a match with the clustering results obtained by the object detection algorithm. The filtered pseudo-labels can be pseudo-labels that meet the preset filtering conditions after filtering.

[0049] For example, in step S300, the pseudo-label filtering module 23 filters the initial pseudo-labels to obtain filtered pseudo-labels that meet the preset filtering conditions.

[0050] The pseudo-label filtering module 23 determines the initial pseudo-label score based on a preset scoring algorithm. If the score is less than a preset low threshold, the module deletes the corresponding initial pseudo-label. Alternatively, if the score is greater than a preset high threshold, the module retains the corresponding initial pseudo-label, thus obtaining pseudo-labels that meet the preset filtering conditions. Or, if the score is less than or equal to a preset high threshold and greater than or equal to a preset low threshold, the module determines the clustering result of the third point cloud data based on a ground point filtering algorithm and a target clustering algorithm; the module then matches the corresponding initial pseudo-label with the clustering result; and retains the successfully matched initial pseudo-labels, thus obtaining the filtered pseudo-labels.

[0051] In step S400, the pre-trained target detection model is iteratively trained based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo label.

[0052] For example, in step S400, the joint optimization module 24 can perform a certain number of rounds of iterative training on the pre-trained detection model based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo label.

[0053] Through the above embodiments, the technical solution of this application can obtain second point cloud data that meets the preset target domain line count condition by performing domain bias processing on the first point cloud data of the source domain. The technical solution of this application can input the third point cloud data of the target domain into a pre-trained object detection model to obtain initial pseudo-labels for the third point cloud data. The technical solution of this application can filter the initial pseudo-labels to obtain filtered pseudo-labels that meet preset filtering conditions. The technical solution of this application can iteratively train the pre-trained object detection model using the second point cloud data and its corresponding true label, the third point cloud data and its corresponding filtered pseudo-labels.

[0054] The technical solution of this application can construct source domain point cloud data with equivalent line count to the target domain's LiDAR by performing domain bias processing on the source domain point cloud data, thereby achieving domain alignment between the source domain point cloud data and the target domain point cloud data in terms of line count and distribution. The technical solution of this application uses source domain data that has undergone domain bias processing, which can better assist in the joint optimization of the target domain's target detection model performance and improve the unsupervised training effect of the target domain.

[0055] Optionally, see Figure 2 Step S100 may include steps S110-S130.

[0056] In step S110, the first point cloud data is filtered to obtain filtered point cloud data that meets the preset target domain lidar vertical field of view.

[0057] According to the example embodiment, the preset target domain lidar vertical field of view range can be the vertical field of view range of the lidar corresponding to the target domain. The vertical field of view range can be determined by the characteristics of the lidar sensor. The filtered point cloud data can be the first point cloud data whose vertical field of view angle satisfies the preset target domain lidar vertical field of view range.

[0058] For example, in step S110, the domain deviation processing module 21 filters the first point cloud data to obtain filtered point cloud data that meets the preset target domain lidar vertical field of view.

[0059] For example, the domain deviation processing module 21 determines the set of point cloud data whose vertical angle of the first point cloud data in the source domain falls within the preset target domain lidar vertical field of view range [target_ωmin, target_ωmax] as the filtered point cloud data. target_ωmin is the minimum value of the preset target domain lidar vertical field of view range, and target_ωmax is the maximum value of the preset target domain lidar vertical field of view range.

[0060] The Cartesian coordinates for filtering point cloud data can be represented as:

[0061] Where r is the measured distance from the point in the selected point cloud data to the lidar; ω is the vertical angle of the lidar laser, i.e., the angle between the laser and the XOY plane; α is the horizontal rotation angle of the lidar laser, i.e., the angle between the projection of the laser onto the XOY plane and the Y-direction coordinate; δ is the horizontal offset angle of the lidar channel. x, y, and z are the polar coordinates projected onto the X, Y, and Z axes, respectively.

[0062] The Cartesian coordinates of all point cloud data can be filtered as follows: [( , , ), ( , , ), ..., ( , , )],[( , , ), ( , , ), ..., ( , , )],...,[( , , ), ( , , ), ..., ( , , )).

[0063] In step S120, based on the scanning resolution of the target domain lidar and the filtered point cloud data, the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions is determined.

[0064] According to the example embodiment, the preset target domain line count distribution condition can be a distribution condition that is closer to the distribution form of the target domain. The equivalent distribution point cloud data can be source domain point cloud data whose distribution form is closer to the distribution form of the target domain.

[0065] For example, in step S120, the domain deviation processing module 21 determines the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions based on the scanning resolution of the target domain lidar and the filtered point cloud data.

[0066] The domain bias processing module 21 constructs an M-row × N-column grid based on the scanning resolution of the target domain lidar. Then, the module iterates through the filtered point cloud data, calculates the cell index (row, col) for each point cloud data point, and performs channel alignment. The channel-aligned point cloud data can be used as equivalent distribution data. The point cloud data of the equivalent distribution data can be distributed across an M × N grid.

[0067] For example, the domain offset processing module 21 can determine the channel alignment row according to the following formula:

[0068] For example, the domain offset processing module 21 can determine the channel alignment col (column) according to the following formula:

[0069] in,( , , ) represents the Cartesian coordinates of the point cloud data, target_vres represents the vertical angular resolution of the LiDAR corresponding to the target domain, and target_hres represents the horizontal angular resolution of the LiDAR corresponding to the target domain.

[0070] In step S130, the second point cloud data is determined based on the equivalent distribution point cloud data.

[0071] According to the example implementation, for a grid with multiple point cloud data, only the point cloud data that was first placed in the cell can be retained, or the point cloud data that is closest to the geometric center of all the point cloud data in the cell can be retained, thereby obtaining the second point cloud data.

[0072] For example, if a cell C contains a point cloud dataset as follows: S represents the number of points in the point cloud data contained within cell C. Point cloud data set. The geometric center c (i.e., the centroid) can be represented as:

[0073] The point cloud data representing the points stored in cell C It can be represented as: .

[0074] For grids without point cloud data distribution, which are considered empty grids, the domain deviation processing module 21 can perform corresponding operations based on whether the number of points within a window grid centered on that grid and with a column length of 2K+1 exceeds a quantity threshold. If the number of points within the window is less than the quantity threshold, no interpolation operation is performed; if the number of points within the window is greater than or equal to the quantity threshold, interpolation is performed to obtain the second point cloud data. Interpolation calculation methods can include, but are not limited to, linear interpolation, polynomial interpolation, spline curve interpolation, etc. For example, interpolation calculation can be expressed by the following formula:

[0075] in, Points after interpolation from an empty grid. The row index position for an empty grid. The column index position of the empty grid, K is the window radius, and interp() is the interpolation function; The row index is k, and the column index is The coordinates of points within a non-empty grid.

[0076] Through the above embodiments, the technical solution of this application can obtain filtered point cloud data that meets the vertical field of view of a preset target domain lidar by filtering the first point cloud data. The technical solution of this application can determine the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions by using the scanning resolution of the target domain lidar and the filtered point cloud data. The technical solution of this application can determine the second point cloud data using the equivalent distribution point cloud data.

[0077] The technical solution of this application can construct point cloud data with an equivalent number of lines to the target domain by analyzing the scanning imaging principle and point cloud characteristics of LiDAR, using the method of "interval mapping + channel alignment" on the original point cloud of the source domain. Compared with the existing general 3D target detection methods based on domain bias processing such as scale scaling, the technical solution of this application is more in line with the characteristics of LiDAR point cloud scanning imaging and the sparsity of point clouds, and can better mitigate the detection performance loss caused by domain bias in model transfer.

[0078] Optionally, see Figure 3 Step S300 may include steps S310-S360.

[0079] In step S310, the scoring result of the initial pseudo-label is determined according to the preset scoring algorithm.

[0080] According to the example embodiment, the preset scoring algorithm can be a scoring algorithm for the initial pseudo-labels. The scoring result can be the quality score result of the initial pseudo-labels. For example, the preset scoring algorithm may include the scoring algorithm of ST3D, etc., and the scoring result can be expressed as the following formula:

[0081] Where score is the initial pseudo-label score; cls_conf is the class confidence obtained by the object detection branch inference of the pre-trained object detection model; iou_conf is the IoU prediction value output by the IoU prediction branch inference of the pre-trained object detection model; and a is the weight coefficient for adjusting the ratio of cls_conf to iou_conf.

[0082] In step S320, if the score result is determined to be less than the preset low score threshold, the corresponding initial pseudo-label is deleted.

[0083] According to the example embodiment, the preset low score threshold can be the minimum value of the quality score of the preset initial pseudo-label. If the score result is less than the preset low score threshold, it indicates that the quality score of the initial pseudo-label is low, the credibility of the initial pseudo-label is low, and the pseudo-label filtering module 23 can directly delete the initial pseudo-label.

[0084] In step S330, if the score result is determined to be greater than the preset high score threshold, the corresponding initial pseudo-label is retained, thereby obtaining the filtered pseudo-label.

[0085] According to the example embodiment, the preset high score threshold can be the highest value of the preset quality score of the initial pseudo-label. If the score result is greater than the preset high score threshold, it indicates that the quality score of the initial pseudo-label is high and the credibility of the initial pseudo-label is high. Then, the pseudo-label filtering module 23 can directly retain the initial pseudo-label to obtain the filtered pseudo-label.

[0086] In step S340, it is determined that the score result is less than or equal to the preset high score threshold and the score result is greater than or equal to the preset low score threshold. The clustering result of the third point cloud data is determined according to the ground point filtering algorithm and the target clustering algorithm.

[0087] According to the example embodiment, the ground filtering algorithm can be an algorithm that filters out ground points from the third point cloud data of the target domain. The pseudo-label filtering module 23 can use the ground filtering algorithm to filter out ground points in the second point cloud data and retain non-ground points.

[0088] Target clustering algorithms are algorithms that divide non-ground points in third-order point cloud data of a target domain into several independent point clusters. Each point cluster can correspond to an independent object in the physical world (such as a car, a pedestrian, a tree, etc.). The clustering result can be the point clusters obtained by the target clustering algorithm.

[0089] Optionally, the target clustering algorithm may include deep clustering algorithms, Euclidean clustering, scan-line based clustering, etc.

[0090] If the score result is less than or equal to the preset high score threshold and greater than or equal to the preset low score threshold, it indicates that the initial pseudo-label is in an uncertain state. Then, the pseudo-label filtering module 23 determines the clustering result of the third point cloud data according to the ground point filtering algorithm and the target clustering algorithm.

[0091] In step S350, the corresponding initial pseudo-labels are matched with the clustering results.

[0092] According to the example embodiment, the pseudo-label filtering module 23 can match the initial pseudo-label with the clustering results using criteria such as IoU, distance, and Hungarian matching.

[0093] In step S360, the initial pseudo-labels that match successfully are retained, thereby obtaining the filtered pseudo-labels.

[0094] According to the example embodiment, the pseudo-label filtering module 23 can retain the initial pseudo-labels that match successfully and delete the initial pseudo-labels that do not match successfully. The pseudo-label filtering module 23 can obtain filtered pseudo-labels (filtered pseudo-labels include initial pseudo-labels whose scores are greater than a preset high score threshold and initial pseudo-labels whose initial pseudo-labels match the clustering results successfully).

[0095] Through the above embodiments, the technical solution of this application can determine the scoring result of the initial pseudo-label using a preset scoring algorithm. If the scoring result is less than a preset low threshold, the corresponding initial pseudo-label is deleted. If the scoring result is greater than a preset high threshold, the corresponding initial pseudo-label is retained. If the scoring result is less than or equal to the preset high threshold and greater than or equal to the preset low threshold, the clustering result of the third point cloud data is determined based on the ground point filtering algorithm and the target clustering algorithm. The corresponding initial pseudo-label is then matched with the clustering result, and the successfully matched initial pseudo-labels are retained, thereby obtaining the filtered pseudo-labels.

[0096] The technical solution of this application can filter the initial pseudo-labels through a three-stage screening and matching screening method to obtain filtered pseudo-labels, thereby improving the quality of pseudo-labels used for subsequent unsupervised training of the pre-trained object detection model in the target domain.

[0097] Optionally, see Figure 4 Step S340 may include steps S341-S346.

[0098] In step S341, the third point cloud data is filtered to obtain preliminary point cloud data within a preset selected range.

[0099] According to the example embodiment, the preset selection range can be a pre-selected ground plane area. The initial screening point cloud data can be the point cloud data of the pre-selected ground plane. The pseudo-label screening module 23 can select a relatively wide and flat ground area to remove non-ground plane areas (such as flower beds, shrubs, low-height buildings, etc.).

[0100] For example, in step S341, the pseudo-label filtering module 23 filters the third point cloud data to obtain preliminary point cloud data within a preset selected range, thereby selecting a relatively broad ground area and reducing the difficulty of subsequent ground point filtering.

[0101] In step S342, the seed point data are sorted according to the height values ​​of the initial screening point cloud data to obtain a preset number of seed point data.

[0102] According to the example embodiment, the seed point data can be a small amount of initial ground point data.

[0103] For example, in step S342, the pseudo-label filtering module 23 sorts the initial screened point cloud data according to the height value (i.e., z value) (e.g., ascending or descending sort), and uses the preset number of point cloud data with the lowest height as the preset number of seed point data.

[0104] Step S343 is the step to determine the current candidate ground plane. See also... Figure 5 Step S343 may include steps S3431-S3432.

[0105] In step S3431, the difference between the height value of the initial screening point cloud data and the average height of the preset number of seed point data is determined.

[0106] According to the example embodiment, the height value of the initial screening point cloud data can be the z-value of the Cartesian coordinates of the initial screening point cloud data. The pseudo-label filtering module 23 can calculate the difference between the height value of the initial screening point cloud data and the average height of a preset number of seed point data.

[0107] In step S3432, points whose difference is less than a preset height threshold are identified as current candidate ground points.

[0108] According to the example embodiment, the preset height threshold can be a preset low height value. For example, the preset height threshold can be set to 0.2 meters, 0.1 meters, etc. If the difference is less than the preset height threshold, it can be said that the point cloud data point is at a low height above the ground, and the point cloud data point can be a current ground candidate point.

[0109] In step S3433, all the initial screening point cloud data are traversed to obtain all current candidate ground points.

[0110] According to the example embodiment, the pseudo-label filtering module 23 traverses all the initial screening point cloud data to obtain all current ground candidate points.

[0111] In step S3434, the current ground plane corresponding to the current candidate ground point is determined based on all current candidate ground points.

[0112] According to the example embodiment, the current ground plane can be the fitted ground plane from the first iteration. The pseudo-label filtering module 23 can iteratively solve the covariance matrix of all current ground candidate points, and perform singular value decomposition on the covariance matrix to obtain the eigenvector corresponding to the smallest eigenvalue, which is the ground normal vector. The pseudo-label filtering module 23 obtains the fitted current ground plane based on the ground normal vector.

[0113] Step S344 is the step to determine the ground plane for the next iteration. See also... Figure 6 Step S344 may include steps S3441-S3444.

[0114] In step S3441, the distance between the initial screening point cloud data and the current ground plane is determined.

[0115] According to the example embodiment, the distance between the initial screening point cloud data and the current ground plane can be the vertical distance from the initial screening point cloud data to the current ground plane. The pseudo-label screening module 23 can calculate the distance between the initial screening point cloud data and the current ground plane.

[0116] In step S3442, points whose distance is less than a preset distance threshold are identified as candidate ground points for the next iteration.

[0117] According to the example embodiment, the preset distance threshold can be a preset low distance value. For example, the preset distance threshold can be set to a height of 0.2 meters, 0.1 meters, etc.

[0118] If the distance is less than the preset distance threshold, it can be indicated that the point in the point cloud data is relatively low from the current ground plane, and the pseudo-label filtering module 23 can use the point in the point cloud data as a ground candidate point for the next iteration.

[0119] In step S3443, all the initial screening point cloud data are traversed to determine all candidate ground points for the next iteration.

[0120] According to the example embodiment, the pseudo-label filtering module 23 can traverse all the initial screening point cloud data to determine all candidate ground points for the next iteration.

[0121] In step S3444, the ground plane for the next iteration is determined based on all the candidate ground points for the next iteration.

[0122] According to the example embodiment, the ground plane for the next iteration can be the ground plane that started the iteration from the second iteration. The pseudo-label filtering module 23 can iteratively solve the covariance matrix of all ground candidate points for the next iteration, perform singular value decomposition on the covariance matrix to obtain the eigenvector corresponding to the smallest eigenvalue, which is the normal vector of the ground. Then, the fitted ground plane for the next iteration is obtained based on the normal vector of the ground.

[0123] In step S345, the ground plane step for determining the next iteration is executed according to the preset number of iterations to obtain the set of non-ground points of the point cloud data within the selected range.

[0124] According to the example embodiment, the preset number of iterations can be the number of times the ground plane is fitted. The pseudo-label filtering module 23 performs the step of determining the ground plane for the next iteration according to the preset number of iterations, which can separate the set of ground points and the set of non-ground points within the selected range to obtain the set of non-ground points within the selected range.

[0125] In step S346, the non-ground point set is clustered based on the deep clustering algorithm to obtain the clustering results.

[0126] According to the example embodiment, the deep clustering algorithm can be a depth map-based clustering algorithm. Deep clustering algorithms project points from non-ground point sets onto a depth map, transforming disordered and unstructured point cloud data into organized data. This allows for the rapid and efficient calculation of target point clusters, thereby obtaining the clustering results.

[0127] Optionally, see Figure 7 Step S346 may include steps S3461-S3465.

[0128] In step S3461, the initial depth map matrix, initial horizontal and vertical angle matrices, and initial clustering label matrix of the non-ground point set are determined based on the number of lines and horizontal resolution of the lidar corresponding to the target domain.

[0129] According to the example embodiment, the depth map matrix can be a depth map showing the distance between each pixel and the LiDAR. The initial depth map matrix can be a depth map of a set of non-ground points initialized based on the number of lines and horizontal resolution of the LiDAR corresponding to the target domain. The number of rows in the initial depth map matrix is ​​determined by the number of laser beams in the vertical direction of the LiDAR, such as 16, 32, or 64; the number of columns in the initial depth map matrix is ​​determined by the horizontal resolution of the LiDAR. For example, if the LiDAR corresponding to the target domain is a 32-line mechanical rotating LiDAR with a horizontal angular resolution of 0.2°, then the size of the initial depth map matrix is ​​32 (number of rows) × 1800 (number of columns).

[0130] The horizontal and vertical angles β can be the angles between the laser beams of a point in the point cloud data and the straight lines connecting that point to its neighboring points. The neighboring points can be the point's nearest neighbors in the depth map matrix, located at the top, bottom, left, and right.

[0131] For example, see Figure 10 Points A and B are two points in the set of non-ground points. Point B is the adjacent point of point A in the depth map matrix. O is the scanning origin of the lidar (the position of the laser emitter). The horizontal and vertical angles β of point A can be the angle between lines OA and AB.

[0132] The initial horizontal and vertical angle matrices can be the horizontal and vertical angle matrices of the non-ground point set initialized according to the number of lines and horizontal resolution of the lidar corresponding to the target domain.

[0133] The clustering label matrix can be a matrix that stores the cluster labels of each pixel. The initial clustering label matrix can be a clustering label matrix of the non-ground point set initialized according to the number of lines and horizontal resolution of the LiDAR corresponding to the target domain.

[0134] For example, the pseudo-label filtering module 23 can initialize the depth map matrix, the horizontal and vertical angle matrices, and the cluster label matrix according to the size of the initial depth map matrix to obtain the initial depth map matrix, the initial horizontal and vertical angle matrices, and the initial cluster label matrix.

[0135] Step S3462 is the step for determining the depth map matrix. See also... Figure 8 Step S3462 may include steps S34621-S34623.

[0136] In step S34621, the index position of the non-ground points in the initial depth map matrix is ​​determined based on the channel information and horizontal angle of the non-ground points in the non-ground point set.

[0137] According to the example embodiment, the channel information can be the channel of the LiDAR corresponding to the 3D point cloud data. The horizontal angle can be the horizontal angle value corresponding to the 3D point cloud data. The index position can be the index of the 3D point cloud data in the depth map matrix.

[0138] For example, the pseudo-label filtering module 23 determines the index position of the non-ground points in the initial depth map matrix based on the channel information and horizontal angle of the non-ground points in the non-ground point set.

[0139] In step S34622, the pixel value corresponding to the index position is determined based on the distance between the non-ground point and the lidar corresponding to the target domain.

[0140] According to the example embodiment, the pixel value corresponding to the index position can be the pixel value of the pixel point corresponding to the index position. The distance between a non-ground point and the lidar corresponding to the target domain (i.e., the distance from the point reached by the number of laser lines emitted by the laser emitter) can be used as the pixel value.

[0141] For example, the pseudo-label filtering module 23 determines the pixel value corresponding to the index position based on the distance between the non-ground point and the lidar corresponding to the target domain.

[0142] In step S34623, all non-ground points in the non-ground point set are traversed to determine the depth map matrix.

[0143] According to the example embodiment, the pseudo-label filtering module 23 traverses all non-ground points in the non-ground point set to determine the depth map matrix.

[0144] Step S3463 is the step of determining the horizontal and vertical angle matrices. See also... Figure 9 Step S3463 may include steps S34631-S134632.

[0145] In step S34631, the horizontal and vertical angles corresponding to the points in the depth map matrix are determined based on a preset horizontal and vertical angle algorithm.

[0146] According to the example embodiment, the preset horizontal and vertical angle algorithm can be a preset algorithm for determining the horizontal and vertical angles β.

[0147] For example, see Figure 10 Point H is the foot of the perpendicular from point B to line segment OA. λ is the angle between line segments OA and OB, which can be determined based on the horizontal alignment angles of point A and B.

[0148] For example, the pseudo-label filtering module 23 can determine the horizontal and vertical angles β according to the following formula:

[0149] Where d1 is the distance from the laser emitter to point A, which is the length of line segment OA, and d2 is the distance from the laser emitter to point B, which is the length of line segment OB.

[0150] In step S34632, all points in the depth map matrix are traversed to obtain all horizontal and vertical angles, thereby obtaining the horizontal and vertical angle matrix.

[0151] According to the example embodiment, the pseudo-label filtering module 23 can traverse all points of the depth map matrix to obtain all horizontal and vertical angles, thereby obtaining the horizontal and vertical angle matrix.

[0152] In step S3464, the clustering label matrix is ​​determined based on the four-neighborhood algorithm, according to the depth map matrix and the horizontal and vertical angle matrices.

[0153] According to the example implementation, the four-neighborhood algorithm can be an algorithm model based on a certain pixel point, which only considers the nearest pixels in the four directions of top, bottom, left and right.

[0154] For example, in step S3464, the pseudo-label filtering module starts from the top left corner of the depth map matrix and recursively determines whether the horizontal and vertical angles β at corresponding positions are less than the horizontal and vertical angle thresholds using the four-neighbor algorithm. If the horizontal and vertical angles β are greater than or equal to the horizontal and vertical angle thresholds, the pseudo-label filtering module 23 assigns the same clustering label to the point and its corresponding neighboring points that are greater than or equal to the horizontal and vertical angle thresholds. If both horizontal and vertical angles β are less than the horizontal and vertical angle thresholds, the pseudo-label filtering module 23 assigns a new clustering label to the point, thereby generating a clustering label matrix.

[0155] In step S3465, the clustering result is determined based on the clustering label matrix.

[0156] For example, the pseudo-label filtering module 23 extracts the minimum bounding box of each point cluster in the cluster label matrix as the target 3D envelope box, which is used as the clustering result.

[0157] Through the above embodiments, the technical solution of this application can cluster the set of non-ground plane points based on ground point filtering to obtain clustering results, which can better assist in the determination of uncertain pseudo-labels and improve the quality of pseudo-labels used for unsupervised training in the target domain. Compared with other pseudo-label screening methods, the technical solution of this application has no special requirements on the temporal relationship between data, does not require the introduction of other sensor modalities or perception models to assist in screening, and has high computational efficiency.

[0158] Optionally, step S400 may specifically be as follows: according to the preset number of training rounds, the pre-trained target detection model is cross-iteratively trained based on the second point cloud data and the true label corresponding to the second point cloud data, the third point cloud data and the pseudo label corresponding to the third point cloud data that meets the preset screening conditions.

[0159] According to the example embodiment, the second point cloud data is denoted as The third point is cloud data, denoted as , The number of true labels in the second point cloud data. This indicates the number of pseudo-labels in the third cloud data that meet the preset filtering conditions.

[0160] The joint optimization module 24 optimizes the pre-trained target detection model according to the preset number of training rounds. and Cross-training allows the overall loss function of the pre-trained object detection model to gradually decrease until it plateaus. The overall loss function of the pre-trained object detection model during training can be expressed as follows:

[0161] Where Ltotal is the total loss function during the training phase of the pre-trained object detection model, and Ltarget is the training loss of the pre-trained object detection model in the object domain. is the training loss of the pre-trained object detection model in the source domain after domain bias processing, and b is a weighting coefficient used to balance the model's contribution to loss between different domains.

[0162] Through the above embodiments, the technical solution of this application uses the source domain after domain bias processing to assist the target domain in unsupervised training, which can reduce the negative impact of pseudo-label noise in the target domain on the parameter adjustment direction of the target detection model, and at the same time make it easier to learn the common features between different domains, thereby improving the generalization ability of the target detection model.

[0163] Compared with training using the target domain alone, the technical solution of this application can avoid the negative impact on the optimization direction of the model caused by the accumulation and amplification of residual noise from the pseudo-labels in the target domain during iterative training.

[0164] According to another aspect of this application, this application also provides a non-volatile computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is capable of implementing the method for optimizing the target detection model of lidar as described above.

[0165] According to another aspect of this application, this application also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement the method for optimizing the target detection model of the lidar as described above.

[0166] According to another aspect of this application, this application also provides a computer program product, comprising: a computer program stored on a computer-readable storage medium; the computer program includes program instructions that, when executed by a computer, cause the computer to perform the method for optimizing the target detection model of a lidar as described above.

[0167] Finally, it should be noted that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions of the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for optimizing a target detection model for lidar, characterized in that, The method includes: The first point cloud data of the source domain is subjected to domain deviation processing to obtain the second point cloud data that meets the preset target domain line number condition; The third point cloud data of the target domain is input into the pre-trained target detection model to obtain the initial pseudo-label of the third point cloud data; The initial pseudo-labels are filtered to obtain filtered pseudo-labels that meet the preset filtering conditions; The pre-trained target detection model is iteratively trained based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo label.

2. The method according to claim 1, characterized in that, The process of performing domain offset processing on the first point cloud data of the source domain to obtain second point cloud data that meets the preset target domain line count condition includes: The first point cloud data is filtered to obtain filtered point cloud data that meets the vertical field of view range of the preset target domain lidar; Based on the scanning resolution of the target domain lidar and the filtered point cloud data, the equivalent distribution point cloud data of the source domain that meets the preset target domain line number distribution conditions is determined. The second point cloud data is determined based on the equivalent distribution point cloud data.

3. The method according to claim 1, characterized in that, The step of filtering the initial pseudo-tags to obtain filtered pseudo-tags that meet preset filtering conditions includes: The scoring result of the initial pseudo-label is determined according to the preset scoring algorithm; If the score is determined to be less than a preset low score threshold, the corresponding initial pseudo-label is deleted; or If the score result is greater than a preset high score threshold, the corresponding initial pseudo-label is retained, thus obtaining the filtered pseudo-label; or If the score result is determined to be less than or equal to the preset high score threshold and greater than or equal to the preset low score threshold, the clustering result of the third point cloud data is determined based on the ground point filtering algorithm and the target clustering algorithm. Match the corresponding initial pseudo-labels with the clustering results; The initial pseudo-labels that match successfully are retained, thus obtaining the filtered pseudo-labels.

4. The method according to claim 3, characterized in that, The clustering result of the third point cloud data determined based on the ground point filtering algorithm and the target clustering algorithm includes: The third point cloud data is filtered to obtain preliminary point cloud data within a preset selected range; Sort the initial screening point cloud data according to their height values ​​to obtain a preset number of seed point data. The steps to determine the current candidate ground plane include: Determine the difference between the height value of the initial screening point cloud data and the average height of the preset number of seed point data; Points whose difference is less than a preset height threshold are identified as current candidate ground points; Iterate through all the initial screening point cloud data to obtain all current candidate ground points; Based on all the current candidate ground points, determine the current ground plane corresponding to the current candidate ground points; Determining the ground plane step for the next iteration includes: Determine the distance between the initial point cloud data and the current ground plane; Points whose distance is less than a preset distance threshold are identified as candidate ground points for the next iteration; Traverse all the initial screening point cloud data to determine all candidate ground points for the next iteration; Based on all the candidate ground points for the next iteration, determine the ground plane for the next iteration; The ground plane step for determining the next iteration is executed according to a preset number of iterations to obtain the set of non-ground points of the point cloud data within the selected range; The non-ground point set is clustered based on a deep clustering algorithm to obtain the clustering result.

5. The method according to claim 4, characterized in that, The process of clustering the non-ground point set using a deep clustering algorithm to obtain the clustering results includes: The initial depth map matrix, initial horizontal and vertical angle matrices, and initial clustering label matrix of the non-ground point set are determined based on the number of lines and horizontal resolution of the lidar corresponding to the target domain. The steps for determining the depth map matrix include: Based on the channel information and horizontal angle of the non-ground points in the set of non-ground points, determine the index position of the non-ground points in the initial depth map matrix; The pixel value corresponding to the index position is determined based on the distance between the non-ground point and the lidar corresponding to the target domain. Traverse all non-ground points in the set of non-ground points to determine the depth map matrix; The steps to determine the horizontal and vertical angle matrices include: Based on a preset horizontal and vertical angle algorithm, the horizontal and vertical angles corresponding to the points in the depth map matrix are determined. Traverse all points of the depth map matrix to obtain all horizontal and vertical angles, thereby obtaining the horizontal and vertical angle matrix; Based on the four-neighbor algorithm, the clustering label matrix is ​​determined according to the depth map matrix and the horizontal and vertical angle matrices; The clustering result is determined based on the clustering label matrix.

6. The method according to claim 1, characterized in that, The step of iteratively training the pre-trained target detection model based on the second point cloud data and the corresponding true label, the third point cloud data and the corresponding filtered pseudo label includes: According to the preset number of training rounds, the pre-trained target detection model is cross-iteratively trained based on the second point cloud data and the true label corresponding to the second point cloud data, the third point cloud data and the pseudo label corresponding to the third point cloud data that meets the preset screening conditions.

7. An apparatus for optimizing a target detection model for lidar, characterized in that, The device includes: The domain deviation processing module performs domain deviation processing on the first point cloud data of the source domain to obtain the second point cloud data that meets the preset target domain line number condition. The initial pseudo-label acquisition module inputs the third point cloud data of the target domain into the pre-trained target detection model to obtain the initial pseudo-label of the third point cloud data; The pseudo-label filtering module filters the initial pseudo-labels to obtain filtered pseudo-labels that meet preset filtering conditions. The joint optimization module iteratively trains the pre-trained target detection model based on the second point cloud data and the corresponding true label, the third point cloud data, and the filtered pseudo label.

8. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for optimizing the target detection model of lidar as described in any one of claims 1-6.

9. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for optimizing the target detection model of lidar as described in any one of claims 1-6.

10. A computer program product, characterized in that, The method includes a computer program stored on a computer-readable storage medium, the computer program including program instructions that, when executed by a computer, cause the computer to perform the method for optimizing a target detection model for a lidar as described in any one of claims 1-6.

Images