A dual-mode point supplement enhancement method for a radar point cloud and a computer device

Abstract

The application discloses a dual-mode point supplement enhancement method for a radar point cloud and a computer device, wherein the method is suitable for multiple mainstream 4D millimeter wave radar target detection data sets, general pretreatment such as rotation and scaling is performed on the radar point cloud, and then the points are generated through 3D radial distribution and a BEV plane fixed z-axis mode, the physical characteristics of the radar point cloud such as radial velocity and radar cross section are strictly reserved, then local diffusion enhancement is performed by simulating a multipath scattering effect, the point cloud scale is uniformly output through fixed-point sampling, and finally the enhanced point cloud is used for 3D target detection model training, and only the sampling operation is reserved in the test stage. While solving the problem of the sparsity of the radar point cloud, the physical authenticity of the point cloud is ensured, multiple radar data sets and mainstream 3D target detection frameworks are adapted, and the detection precision, generalization ability and robustness of the model are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of radar point cloud data processing and target detection technology, specifically to a dual-modal point enhancement method and computer equipment for radar point clouds, which is particularly suitable for improving the sparsity of 4D millimeter-wave radar point clouds and can be directly connected to target detection model training and application in scenarios such as autonomous driving and robot perception. Background Technology

[0002] In intelligent sensing fields such as autonomous driving and robotic perception, radar point clouds have become a core data source for 3D target detection tasks due to their core advantages of all-weather operation, resistance to light interference, and direct acquisition of target 3D spatial and motion information. Among them, 4D millimeter-wave radar can also simultaneously acquire physical characteristics such as target radial velocity and radar cross-section, further enhancing the comprehensiveness of perception. However, due to factors such as radar hardware detection principles, detection range, complex environmental obstructions, and differences in target reflection characteristics, radar point clouds generally suffer from sparsity, manifested as insufficient effective point cloud quantity and uneven point cloud distribution in the target area. This sparsity problem directly leads to low target localization accuracy of target detection models, easy classification errors, and a high rate of missed detection for small and distant targets, severely restricting the practical application effect of radar perception systems.

[0003] In existing technologies, point cloud enhancement methods mostly focus on processing lidar point clouds, with few dedicated point cloud augmentation strategies for imaging radar, especially 4D millimeter-wave radar point clouds. Some general point cloud augmentation methods do not consider the physical characteristics of radar point clouds, easily losing crucial information such as radial velocity and radar cross-section during the augmentation process, or generating point cloud spatial distributions that do not conform to radar detection patterns, resulting in insufficient realism of the augmented point cloud and failing to effectively improve the performance of the detection model. Furthermore, existing augmentation methods lack adaptability to different radar datasets; single parameter configurations are difficult to match the point cloud distribution characteristics of different datasets, leading to a significant dataset dependency in the augmentation effect. Summary of the Invention

[0004] This invention provides a dual-modal point augmentation method and computer equipment for radar point clouds that efficiently completes sparse point clouds while preserving their physical characteristics, adapts to various radar datasets, and can directly interface with mainstream 3D target detection frameworks. This invention can at least solve one of the aforementioned technical problems.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A dual-modal point cloud augmentation method includes the following steps: S1. Prepare radar point cloud data for training the 3D target detection model. The basic dimensions of the point cloud include three-dimensional spatial coordinates of x, y, and z. The physical features include radial velocity vr and radar cross-section (RCS). Complete data cleaning and format standardization. S2. Automatically load and adapt the basic parameter configuration according to the mainstream radar target detection dataset type to match the enhancement strategy with the point cloud distribution characteristics of the dataset; S3. Perform general enhancement preprocessing on the radar point cloud after parameter configuration, including rotation around the z-axis, proportional scaling of the center, random feature jittering, and random discarding, to improve the generalization ability of the 3D target detection model. S4. Perform three-dimensional spatial clustering on the radar point cloud after general enhancement preprocessing, extract the effective target area, generate supplementary points in the target area based on 3D radial distribution rules, constrain the position of the supplementary points to the neighborhood of the original points, and perform feature assignment on the z coordinate, radar cross-section (RCS) and radial velocity of the supplementary points to preserve physical properties. S5. Project the radar point cloud after general enhancement preprocessing onto the BEV (Bird's Eye View) top-down plane, extract the effective target area, generate supplementary points in the plane and fix the z coordinate, maintain the consistency in the height direction, and assign the radar cross-section (RCS) and radial velocity feature of the supplementary points in the same way as the 3D radial supplementary points. S6. Select some points of the point cloud after the point cloud is supplemented according to the preset probability as diffusion source points, generate diffusion points in the neighborhood of the diffusion source points to simulate the radar multipath scattering effect, and merge the diffusion points with the original point cloud. S7. Perform fixed-point sampling on the radar point cloud after local diffusion enhancement, and unify the number of output point clouds to the preset number of target points that meet the input format requirements of the 3D target detection model; S8. Input the radar point cloud with a fixed number of points into the 3D target detection model for training. At the same time, set up an independent testing phase. In the testing phase, turn off all enhancement operations and only perform fixed-number sampling on the original point cloud to output the detection results.

[0006] Furthermore, in S1, the radar point cloud data format is compatible with multiple mainstream radar target detection datasets. The three-dimensional spatial coordinates of x, y, and z represent the horizontal, vertical, and height positions of the radar point in three-dimensional space, respectively. The radial velocity vr represents the target's speed relative to the radar, and the radar cross-section RCS represents the target's ability to reflect radar electromagnetic waves.

[0007] Furthermore, in step S2, the basic parameter configuration for adaptation is automatically loaded, wherein the parameters include: Effective threshold of radar cross-section This is used to filter out invalid radar points with low signal-to-noise ratio and retain radar points with radar cross-section (RCS) values ​​greater than this threshold for subsequent processing. DBSCAN cluster radius , used to control the target region range of unsupervised clustering in density-based clustering algorithms, characterizing the maximum distance between two radar points in a cluster that can be classified into the same cluster; Minimum number of cluster points , used to remove isolated noise points, characterizes the minimum number of radar points required for an effective cluster; Radial compensation distance range This is used to limit the radial distance between the complement point and the cluster center, and the radial distance of the complement point is randomly selected within this interval; The amplitude of the z-coordinate perturbation of the supplementary point This is used to control the magnitude of the disturbance in the height direction of the supplementary point; BEV planar point-filling disturbance amplitude This is used to control the intensity of positional jitter at points within the top-view plane of the BEV. Radar cross-section RCS attenuation range This is used to limit the attenuation range of the supplementary radar cross-section. The RCS value of the supplementary radar cross-section is the RCS value of the original radar cross-section minus the random value within this range. Compensation point radial velocity noise amplitude This is used to control the amount of noise added to the radial velocity at the compensation point; Maximum number of complements in a single cluster This is used to control the maximum number of supplementary points for a single effective cluster, preventing excessive supplementary points from causing distortion in the point cloud distribution.

[0008] Furthermore, the preprocessing operation in S3 includes: S3.1 Rotation operation: Perform a rotation transformation around the z-axis, randomly generating a rotation angle within the range of angles suitable for the dataset. The rotation matrix R is used to perform a rotation transformation on the x and y three-dimensional spatial coordinates of all radar points, while keeping the z coordinate unchanged. The rotation matrix R is defined as follows:

[0009] in, Rotation angle cosine value, Rotation angle The sine value; S3.2 Scaling operation: Scaling is performed proportionally based on the center of the point cloud. First, the center coordinates c of the entire radar point cloud are calculated. Then, a scaling factor s is randomly generated within a preset scaling range. The x, y, and z coordinates of all radar points are scaled according to the center alignment rule. The scaling formula is:

[0010] in, The spatial coordinates of radar point i before scaling. The scaled spatial coordinates of the radar point; S3.3 Random jitter operation: Gaussian noise is added to the position, radar cross-section (RCS), and radial velocity characteristics of the radar point to achieve slight perturbation of the characteristics. Among them, position jitter is based on the preset position noise amplitude by adding Gaussian noise to the x, y and z coordinates; radar cross-section RCS jitter is based on the preset radar cross-section RCS noise amplitude by adding Gaussian noise to the radar cross-section RCS feature; and radial velocity jitter is based on the preset velocity noise amplitude by adding Gaussian noise to the radial velocity vr feature, simulating the environmental noise and hardware measurement error in the actual detection process. S3.4 Random discard operation: randomly retain a portion of the radar point cloud according to a preset discard ratio range, and discard the remaining point cloud to simulate the scenario of point cloud loss caused by target obstruction and electromagnetic wave attenuation during radar detection.

[0011] Furthermore, S4 further includes: S4.1. Perform 3D spatial clustering on the radar point cloud that has undergone general enhancement preprocessing based on the DBSCAN unsupervised clustering algorithm, combined with the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out noisy clusters and false clusters with excessively large velocity variances, and retain valid target clusters; S4.2 For each valid target cluster, calculate the coordinates of its cluster center. ,in, , and These are the coordinates of the cluster centers in the x, y, and z directions, respectively. S4.3, Using the cluster center as the pole, the radial distance range of the supplementary points is determined based on the parameters configured in S2. Randomly select a radial distance r, and randomly select a polar angle within the range [0, 2π]. The x and y coordinates of the supplementary points are generated according to the 3D radial distribution rule. The calculation formula is as follows:

[0012]

[0013] in, , These are the coordinates of the j-th supplementary point in the x and y directions, respectively; S4.4, The z-coordinate of the supplementary point corresponds to the z-coordinate of the original point. Based on this, add the z-coordinate perturbation amplitude of the parameter supplementation point based on the S2 configuration. Gaussian noise with standard deviation To achieve a slight perturbation in the height direction, the calculation formula is:

[0014] in, Let be the coordinate value of the j-th supplementary point in the z-direction. Gaussian noise Follows a mean of 0 and a variance of The normal distribution; S4.5, The radar cross-section (RCS) of the supplementary radar is attenuated according to a preset range, expressed as:

[0015] in, Let RCS be the radar cross-section value of the j-th supplementary point. The eigenvalues ​​corresponding to the original points. To randomly select attenuation values; The radial velocity at the supplementary point inherits from the original point and a small amount of noise is added, expressed as:

[0016] in, Let J be the radial velocity value of the j-th supplementary point. Gaussian noise Follows a mean of 0 and a variance of It follows a normal distribution.

[0017] Furthermore, S5 further includes: S5.1 Project the radar point cloud that has completed general enhancement preprocessing onto the BEV top view plane, retaining the x and y two-dimensional coordinates and ignoring the z coordinate; S5.2. Cluster the point cloud in the BEV plane based on the DBSCAN unsupervised clustering algorithm, and combine the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out invalid noise clusters and retain valid target clusters; S5.3 Within the effective target cluster, for the x and y coordinates of each original point, add the BEV plane perturbation amplitude based on the parameter configured in S2. Gaussian noise with standard deviation and The formula for generating the x and y coordinates of the supplementary point is:

[0018]

[0019] in, and Let x and y be the coordinates of the m-th BEV supplementary point, respectively. and These are the x-coordinate and y-coordinate of the corresponding original point, respectively. , Gaussian noise and All follow a mean of 0 and a variance of . The normal distribution; S5.4 The z-coordinate of the fixed complement point is the z-coordinate of the original point within its respective cluster. To ensure that all complementary points within the same cluster maintain a high degree of consistency with the original points, the calculation formula is as follows:

[0020] in, Let be the coordinates of the m-th BEV supplementary point in the z-direction. This is the z-coordinate value of the nth original point within the cluster to which the supplementary point belongs; S5.5 The processing rules for the radar cross-section (RCS) features and radial velocity features of the supplemented points are consistent with those of the 3D radial supplementation in S4. That is, the radar cross-section (RCS) features are attenuated within a preset range, and the radial velocity features inherit the original points and add small Gaussian noise, ensuring that the physical characteristics of the two supplementation strategies are preserved in a unified manner.

[0021] Furthermore, S6 further includes: S6.1, according to the preset diffusion probability For point clouds that have completed 3D radial point filling or BEV consistency point filling, local diffusion enhancement is performed, and some radar points are randomly selected as diffusion source points. S6.2 For each diffusion source point, randomly select a diffusion distance within a preset diffusion distance range. The diffusion angle is randomly selected within the range of [0, 2π]. The x and y coordinates of the diffusion point are generated by the following formula:

[0022]

[0023] in, and These are the coordinates of the diffusion point in the x and y directions, respectively. and These are the coordinates of the k-th diffusion source point in the x and y directions, respectively. For diffusion angle cosine value, For diffusion angle The sine value; S6.3 The z-coordinate of the diffusion point is based on the z-coordinate of the diffusion source point, with a small amount of Gaussian noise added to simulate slight height perturbation; S6.4. The radar cross-section (RCS) feature of the diffusion point is slightly attenuated, and a small amount of Gaussian noise is added to the radial velocity feature to ensure that the physical characteristics of the diffusion point are consistent with the original point cloud and to avoid feature distortion. S6.5. Merge all generated diffusion points with the point cloud after 3D radial point supplementation or BEV consistency point supplementation to form a radar point cloud with local diffusion enhancement.

[0024] Furthermore, S7 further includes: S7.1 Count the actual number of radar point clouds after local diffusion enhancement. Set the number of target points according to the input requirements of the 3D target detection model. ; S7.2, If the actual number of points >Target Points Then, a random sampling method without replacement is used to randomly select points from the radar point cloud after local diffusion enhancement. Each point ensures the randomness and representativeness of the point cloud; S7.3, If the actual number of points <Target Points A fixed number of point clouds is generated by repeated sampling and remainder padding, and the calculation formula is as follows:

[0025] in, This is the final point cloud after sampling. For the enhanced point cloud, Let d represent the number of repeated samples. The remainder To fill in the remaining points from the point cloud after local diffusion enhancement.

[0026] Furthermore, S8 further includes: S8.1 During the training phase, the enhanced fixed-point radar point cloud is used as the input to the 3D target detection model. Combined with the corresponding target annotation information, the model is trained according to the conventional training process, so that the model learns the point cloud features after the point enhancement and improves the model's ability to detect sparse radar point clouds. S8.2 During the testing phase, all point cloud enhancement operations are turned off, including the general enhancement preprocessing in S3, the 3D radial point filling in S4, the BEV consistency point filling in S5, and the local diffusion enhancement in S6. Only a fixed number of points are sampled on the original point cloud to ensure the originality of the test data. S8.3 Input the processed test point cloud into the trained 3D object detection model and output the final 3D object detection result.

[0027] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the above-described method for dual-modal point enhancement of radar point clouds.

[0028] The beneficial effects of this invention are reflected in: 1. The radar target detection method based on dual-modal point-filling enhancement proposed in this invention is particularly suitable for improving the sparsity of 4D millimeter-wave radar point clouds. Specifically, by employing two independently combinable strategies—3D radial point-filling and BEV-consistent point-filling—the sparsity problem of the radar point cloud is effectively solved while strictly preserving its physical characteristics. Simultaneously, through a complete process of adaptive parameter configuration of the dataset, local diffusion enhancement, and fixed-point sampling, the rationality of the point cloud distribution and the model's generalization ability are significantly improved.

[0029] 2. Compared with existing point cloud enhancement methods, this invention can avoid the loss of key physical information such as radial velocity and radar cross-section during the point supplementation process, ensure that the spatial distribution of the supplemented points conforms to the radar detection law, and can be adapted to various 4D millimeter-wave radar datasets and mainstream target detection frameworks. It can significantly improve target detection accuracy, reduce the false negative rate of small targets and long-distance targets, and at the same time maintain the algorithm's lightweight and high efficiency, which can meet the engineering application needs of real-time perception scenarios such as autonomous driving and robot perception. Attached Figure Description

[0030] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0031] Figure 1 This is a schematic diagram of the overall process of the method according to an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the 3D radial point filling process according to an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the BEV consistency patching process according to an embodiment of the present invention. Detailed Implementation

[0034] 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 a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] It should be noted that the meaning of "and / or" throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, solution B, or a solution that simultaneously satisfies A and B. Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0036] See Figure 1 This invention provides a dual-modal point cloud augmentation method, comprising the following steps: S1. Prepare radar point cloud data for training the 3D target detection model. The radar point cloud data format is compatible with various mainstream radar target detection datasets. The basic dimensions of the point cloud include three-dimensional spatial coordinates of x, y, and z. The three-dimensional spatial coordinates of x, y, and z represent the horizontal, vertical, and height positions of the radar point in three-dimensional space, respectively. The physical features include radial velocity vr and radar cross-section RCS. Radial velocity vr represents the target's speed relative to the radar, and radar cross-section RCS represents the target's ability to reflect radar electromagnetic waves. Complete data cleaning and format standardization to provide a basis for preserving physical characteristics during subsequent point addition.

[0037] S2. Automatically load and adapt the basic parameter configuration according to the mainstream radar target detection dataset type to match the enhancement strategy with the point cloud distribution characteristics of the dataset.

[0038] In this embodiment, step S2 involves automatically loading the adapted basic parameter configuration, wherein the parameter definitions and corresponding adjustment rules include: 1. Effective threshold of radar cross-section This is used to filter out invalid radar points with low signal-to-noise ratio and retain radar points with radar cross-section (RCS) values ​​greater than this threshold for subsequent processing. Adjustment rules: For datasets with low overall radar cross-section (RCS) values ​​in point clouds, the threshold can be appropriately lowered to reduce the loss of valid points; for datasets with many noisy points, the threshold can be appropriately increased to improve the purity of point clouds.

[0039] 2. DBSCAN cluster radius , used to control the target region range of unsupervised clustering in density-based clustering algorithms, characterizing the maximum distance between two radar points in a cluster that can be classified into the same cluster; Adjustment rules: For datasets with dense point cloud distribution, the radius can be appropriately reduced to improve the precision of clustering; for datasets with sparse point cloud distribution, the radius can be appropriately increased to ensure complete clustering of the target region.

[0040] 3. Minimum number of nodes in a cluster , used to remove isolated noise points, characterizes the minimum number of radar points required for an effective cluster; Adjustment rules: For datasets with a high proportion of small targets, this value can be appropriately reduced to prevent small target point clouds from being judged as noise due to insufficient point count; for datasets with complex environmental noise, this value can be appropriately increased to improve the noise filtering effect.

[0041] 4. Radial compensation distance range This is used to limit the radial distance between the complement point and the cluster center, and the radial distance of the complement point is randomly selected within this interval; Adjustment rules: For point clouds of nearby targets, the interval range can be appropriately reduced to ensure the spatial rationality of point supplementation; for point clouds of distant targets, the interval range can be appropriately increased to improve the coverage of point supplementation.

[0042] 5. Amplitude of z-coordinate disturbance at the supplementary point This is used to control the magnitude of the disturbance in the height direction of the supplementary point; Adjustment rules: For datasets with small differences in height features, the amplitude can be appropriately reduced to avoid excessive perturbation in the height direction; for datasets with large differences in height features, the amplitude can be appropriately increased to match the actual height distribution characteristics of point clouds.

[0043] 6. BEV planar compensation point disturbance amplitude This is used to control the intensity of positional jitter at points within the top-view plane of the BEV. Adjustment rules: For datasets with relatively concentrated BEV plane point cloud distribution, the amplitude can be appropriately reduced to ensure the positional accuracy of the supplemented points; for datasets with relatively dispersed BEV plane point cloud distribution, the amplitude can be appropriately increased to improve the realism of the distribution of the supplemented points.

[0044] 7. Radar Cross Section (RCS) Attenuation Range This is used to limit the attenuation range of the supplementary radar cross-section. The RCS value of the supplementary radar cross-section is the RCS value of the original radar cross-section minus the random value within this range. Adjustment rules: For datasets with relatively stable target reflection characteristics, the interval range can be appropriately reduced to ensure the consistency of radar cross-section (RCS) features; for datasets with large differences in target reflection characteristics, the interval range can be appropriately increased to simulate the actual radar reflection intensity attenuation law.

[0045] 8. Add radial velocity noise amplitude This is used to control the amount of noise added to the radial velocity at the compensation point; Adjustment rules: For datasets with small radial velocity fluctuations, the amplitude can be appropriately reduced to ensure the stability of the velocity characteristics; for datasets with large radial velocity fluctuations, the amplitude can be appropriately increased to better reflect the actual target motion characteristics.

[0046] 9. Maximum number of complements in a single cluster This is used to control the maximum number of supplementary points for a single effective cluster, so as to avoid excessive supplementary points leading to distortion of the point cloud distribution. Adjustment rules: For datasets with extremely sparse original point clouds, this value can be increased appropriately to improve the point filling effect; for datasets with low original point cloud sparsity, this value can be decreased appropriately to ensure the realism of the point cloud.

[0047] All of the above parameters are adjustable and can be fine-tuned according to the point cloud features and detection scenarios of the radar dataset used. The parameter values ​​can also be recalibrated based on a new radar dataset, so that the method can be flexibly adapted to different radar datasets.

[0048] S3. Perform general enhancement preprocessing on the radar point cloud after parameter configuration, including rotation around the z-axis, proportional scaling of the center, random jittering of features, and random discarding. Through diverse geometric enhancement and noise enhancement, improve the generalization ability and robustness of the 3D target detection model to different viewpoints, scales, and noise scenes.

[0049] In this embodiment, the preprocessing operation in S3 includes: S3.1 Rotation Operation: Perform a rotation transformation around the z-axis, randomly generating a rotation angle within the angle range suitable for the dataset. The rotation matrix R is used to perform a rotation transformation on the x and y three-dimensional spatial coordinates of all radar points, while keeping the z coordinate unchanged. The rotation matrix R is defined as follows:

[0050] in, Rotation angle cosine value, Rotation angle The sine value; This matrix can rotate the radar point cloud at any angle in the horizontal plane to simulate different relative perspectives between the radar and the target. S3.2 Scaling Operation: Scale proportionally using the point cloud center as the reference. First, calculate the center coordinates c of the entire radar point cloud. Then, randomly generate a scaling factor s within a preset scaling range. Perform scaling transformations on the x, y, and z coordinates of all radar points according to the center alignment rules. The scaling formula is:

[0051] in, The spatial coordinates of radar point i before scaling. The scaled spatial coordinates of the radar point; This operation can simulate different detection distances between the radar and the target, improving the model's ability to detect targets of different scales; S3.3 Random jitter operation: Gaussian noise is added to the position, radar cross-section (RCS), and radial velocity characteristics of the radar point to achieve slight perturbation of the characteristics; Among them, position jitter is based on the preset position noise amplitude by adding Gaussian noise to the x, y and z coordinates; radar cross-section RCS jitter is based on the preset radar cross-section RCS noise amplitude by adding Gaussian noise to the radar cross-section RCS feature; and radial velocity jitter is based on the preset velocity noise amplitude by adding Gaussian noise to the radial velocity vr feature, simulating the environmental noise and hardware measurement error in the actual detection process. S3.4 Random Discard Operation: Randomly retain a portion of the radar point cloud according to a preset discard ratio range, and discard the remaining point cloud to simulate the scenario of point cloud loss caused by target obstruction and electromagnetic wave attenuation during radar detection.

[0052] S4, such as Figure 2 As shown, step S4 is one of the core point-filling strategies of the present invention. It performs three-dimensional spatial clustering on the radar point cloud after general enhancement preprocessing, extracts the effective target area through unsupervised clustering, generates points in the target area based on 3D radial distribution rules, and constrains the position of the points to the neighborhood of the original points to ensure the spatial rationality of the points. It also performs feature assignment on the z-coordinate, radar cross-section (RCS), and radial velocity of the points to preserve their physical properties.

[0053] In this embodiment, step S4 further includes: S4.1. Perform 3D spatial clustering on the radar point cloud that has undergone general enhancement preprocessing based on the DBSCAN unsupervised clustering algorithm, combined with the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out noisy clusters and false clusters with excessively large velocity variances, and retain valid target clusters; S4.2 For each valid target cluster, calculate the coordinates of its cluster center. ,in, , and These are the coordinates of the cluster centers in the x, y, and z directions, respectively. S4.3, Using the cluster center as the pole, the radial distance range of the supplementary points is determined based on the parameters configured in S2. Randomly select a radial distance r, and randomly select a polar angle within the range [0, 2π]. The x and y coordinates of the supplementary points are generated according to the 3D radial distribution rule. The calculation formula is as follows:

[0054]

[0055] in, , These are the coordinates of the j-th supplementary point in the x and y directions, respectively; S4.4, The z-coordinate of the supplementary point corresponds to the z-coordinate of the original point. Based on this, add the z-coordinate perturbation amplitude of the parameter supplementation point based on the S2 configuration. Gaussian noise with standard deviation To achieve a slight perturbation in the height direction, the calculation formula is:

[0056] in, Let be the coordinate value of the j-th supplementary point in the z-direction. Gaussian noise Follows a mean of 0 and a variance of The normal distribution; S4.5, The radar cross-section (RCS) of the supplementary radar is attenuated according to a preset range, expressed as:

[0057] in, Let RCS be the radar cross-section value of the j-th supplementary point. The eigenvalues ​​corresponding to the original points. To randomly select attenuation values; The radial velocity at the supplementary point inherits from the original point and a small amount of noise is added, expressed as:

[0058] in, Let J be the radial velocity value of the j-th supplementary point. Gaussian noise Follows a mean of 0 and a variance of It follows a normal distribution.

[0059] S5, such as Figure 3As shown, step S5 is another core point-filling strategy of the present invention, which is independent of and can be freely combined with the above-mentioned 3D radial point-filling. The radar point cloud after general enhancement preprocessing is projected onto the BEV top view plane (xy plane) to complete the target clustering and point-filling operation, extract the effective target area, generate points in the plane and fix the z coordinate, maintain the consistency in the height direction, avoid spatial distortion in the height direction, adapt to the 3D target detection framework based on BEV features, and the radar cross-section RCS and radial velocity feature assignment rules of the point-filling are consistent with the above-mentioned 3D radial point-filling.

[0060] In this embodiment, S5 further includes: S5.1 Project the radar point cloud that has completed general enhancement preprocessing onto the BEV top view plane, retaining the x and y two-dimensional coordinates and ignoring the z coordinate; S5.2. Cluster the point cloud in the BEV plane based on the DBSCAN unsupervised clustering algorithm, and combine the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out invalid noise clusters and retain valid target clusters; S5.3 Within the effective target cluster, for the x and y coordinates of each original point, add the BEV plane perturbation amplitude based on the parameter configured in S2. Gaussian noise with standard deviation and The formula for generating the x and y coordinates of the supplementary point is:

[0061]

[0062] in, and Let x and y be the coordinates of the m-th BEV supplementary point, respectively. and These are the x-coordinate and y-coordinate of the corresponding original point, respectively. , Gaussian noise and All follow a mean of 0 and a variance of . The normal distribution; S5.4 The z-coordinate of the fixed complement point is the z-coordinate of the original point within its respective cluster. To ensure that all complementary points within the same cluster maintain a high degree of consistency with the original points, the calculation formula is as follows:

[0063] in, Let be the coordinates of the m-th BEV supplementary point in the z-direction. The z-coordinate value of the nth original point within the cluster to which the supplementary point belongs is given, ensuring no distortion in the height direction within the BEV plane; S5.5 The processing rules for the radar cross-section (RCS) features and radial velocity features of the supplemented points are consistent with those of the 3D radial supplementation in S4. That is, the radar cross-section (RCS) features are attenuated within a preset range, and the radial velocity features inherit the original points and add small Gaussian noise, ensuring that the physical characteristics of the two supplementation strategies are preserved in a unified manner.

[0064] S6. Select some points of the point cloud after point supplementation as diffusion source points according to the preset probability, generate diffusion points in the neighborhood of diffusion source points to simulate radar multipath scattering effect, that is, the scattering point formed when radar electromagnetic waves are reflected by environmental objects and reach the target again, improve the distribution realism and scene fit of the point cloud after point supplementation, and merge diffusion points with the original point cloud.

[0065] In this embodiment, step S6 further includes: S6.1, according to the preset diffusion probability For point clouds that have undergone 3D radial point patching or BEV consistency point patching, local diffusion enhancement is performed, and a portion of radar points are randomly selected as diffusion source points, where the diffusion probability is... It is an adjustable parameter that can be fine-tuned according to the scattering characteristics of the dataset; S6.2 For each diffusion source point, randomly select a diffusion distance within a preset diffusion distance range. The diffusion angle is randomly selected within the range of [0, 2π]. The x and y coordinates of the diffusion point are generated by the following formula:

[0066]

[0067] in, and These are the coordinates of the diffusion point in the x and y directions, respectively. and These are the coordinates of the k-th diffusion source point in the x and y directions, respectively. For diffusion angle cosine value, For diffusion angle The sine value; S6.3 The z-coordinate of the diffusion point is based on the z-coordinate of the diffusion source point, with a small amount of Gaussian noise added to simulate slight height perturbation; S6.4. The radar cross-section (RCS) feature of the diffusion point is slightly attenuated, and a small amount of Gaussian noise is added to the radial velocity feature to ensure that the physical characteristics of the diffusion point are consistent with the original point cloud and to avoid feature distortion. S6.5. Merge all generated diffusion points with the point cloud after 3D radial point supplementation or BEV consistency point supplementation to form a radar point cloud with local diffusion enhancement.

[0068] S7. Perform fixed-point sampling on the radar point cloud after local diffusion enhancement, and unify the number of output point clouds to the preset number of target points that meet the input format requirements of the 3D target detection model, so as to avoid model training instability due to differences in the number of point clouds.

[0069] In this embodiment, S7 further includes: S7.1 Count the actual number of radar point clouds after local diffusion enhancement. Set the number of target points according to the input requirements of the 3D target detection model. Among them, the preset target number of points This is an adjustable parameter, typically ranging from [1024, 4096]. S7.2, If the actual number of points >Target Points Then, a random sampling method without replacement is used to randomly select points from the radar point cloud after local diffusion enhancement. Each point ensures the randomness and representativeness of the point cloud; S7.3, If the actual number of points <Target Points A fixed number of point clouds is generated by repeated sampling and remainder padding, and the calculation formula is as follows:

[0070] in, This is the final point cloud after sampling. For the enhanced point cloud, Let d represent the number of repeated samples. The remainder To fill in the remaining points from the point cloud after local diffusion enhancement.

[0071] S8. Input a fixed number of radar point clouds into the 3D target detection model for training. At the same time, set up an independent testing phase to ensure the fairness and accuracy of model evaluation. During the testing phase, turn off all augmentation operations and only perform fixed-number sampling on the original point cloud to output the detection results.

[0072] In this embodiment, S8 further includes: S8.1 During the training phase, the enhanced fixed-point radar point cloud is used as the input to the 3D target detection model. Combined with the corresponding target annotation information, the model is trained according to the conventional training process, so that the model learns the point cloud features after the point enhancement and improves the model's ability to detect sparse radar point clouds. S8.2 During the testing phase, all point cloud enhancement operations are turned off, including the general enhancement preprocessing in S3, the 3D radial point filling in S4, the BEV consistency point filling in S5, and the local diffusion enhancement in S6. Only a fixed number of points are sampled on the original point cloud to ensure the originality of the test data. S8.3 Input the processed test point cloud into the trained 3D object detection model and output the final 3D object detection results, including the object's location, category, confidence level, etc., to achieve an accurate evaluation of the model's actual detection performance.

[0073] To verify the effectiveness of the method of the present invention, a comparative experiment was conducted on the publicly available 4D millimeter-wave radar target detection dataset View-of-Delft. The mean accuracy (mAP), small target AP, and long-range target AP were used as evaluation indicators. The experimental results are shown in Tables 1 and 2 below.

[0074] Table 1. Experimental results within labeled regions based on the View-of-Delft dataset

[0075] Table 2. Experimental results within the driving lane based on the View-of-Delft dataset.

[0076] In Tables 1 and 2, Module 1 characterizes 3D radial patching enhancement, Module 2 characterizes BEV consistency patching enhancement, AP characterizes average accuracy, and mAP 3D The mean accuracy (mAP) of 3D detection is a measure of the average precision of the detection process. BEV The average accuracy of the bird's-eye view detection is represented by the labeled area, the labeled area represents the fully labeled detection area, and the driving channel represents the vehicle driving channel area.

[0077] As shown in Tables 1 and 2, on the View-of-Delft dataset, without the enhancement module described in this invention, the model's detection accuracy in both the labeled region and the driving channel is relatively low. However, after enabling Module 1 alone (3D radial augmentation), the AP values ​​and mAP of the three target classes (cars, pedestrians, and cyclists) are significantly improved. 3D mAP BEVAll showed significant improvements; activating module 2 alone, i.e., BEV consistency point enhancement, also effectively improved model detection accuracy. When both 3D radial point enhancement and BEV consistency point enhancement modules were activated simultaneously, the dual-level point enhancement effect of this invention was fully demonstrated, further improving the detection accuracy of cars, pedestrians, and cyclists within the driving lane, and increasing mAP. 3D With mAP BEV The achievement of optimal performance indicates that the dual-level point-filling enhancement strategy proposed in this invention can synergistically improve the overall performance of the 4D millimeter-wave radar target detection model and effectively improve the problem of insufficient detection accuracy caused by sparse point clouds.

[0078] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the dual-modal point enhancement method for radar point clouds described above.

[0079] This invention also provides a computer device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the dual-modal point enhancement method for radar point clouds described above.

[0080] This invention also provides a computer program product containing instructions that, when run on a computer, causes the computer to perform the steps of the dual-modal point enhancement method for radar point clouds described above.

[0081] It is understood that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above-described dual-modal point enhancement method for radar point clouds.

[0082] It should be noted that those skilled in the art will understand that all or part of the steps implemented in the embodiments of the present invention can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. When implemented in hardware, it can be implemented entirely or partially by purchasing standard parts or modifications. When implemented in software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0083] In summary, to address the shortcomings of existing point cloud augmentation methods, such as insufficient realism and inability to effectively improve the performance of detection models, this invention proposes a dual-modal point cloud augmentation method suitable for radar point clouds, particularly for improving the sparsity of 4D millimeter-wave radar point clouds. Specifically, this method innovatively designs two independent and freely combinable point augmentation strategies: 3D radial point augmentation and BEV-consistent point augmentation. It also includes a supporting design for adaptive dataset parameter configuration, local diffusion enhancement, and fixed-point sampling. While strictly preserving the physical characteristics of radar point clouds, such as radial velocity and radar cross-section, it efficiently augments sparse point clouds and adapts to various radar datasets. This allows for direct integration with mainstream 3D target detection frameworks, such as those used in autonomous driving and robot perception scenarios, for target detection model training and application, effectively improving the accuracy, generalization ability, and robustness of the detection model.

[0084] It should be understood that the examples and embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various modifications or changes based on them. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A dual-modal point cloud augmentation method for radar point clouds, characterized in that, Includes the following steps: S1. Prepare radar point cloud data for training the 3D target detection model. The basic dimensions of the point cloud include three-dimensional spatial coordinates of x, y, and z. The physical features include radial velocity vr and radar cross-section (RCS). Complete data cleaning and format standardization. S2. Automatically load and adapt the basic parameter configuration according to the mainstream radar target detection dataset type to match the enhancement strategy with the point cloud distribution characteristics of the dataset; S3. Perform general enhancement preprocessing on the radar point cloud after parameter configuration, including rotation around the z-axis, proportional scaling of the center, random feature jittering, and random discarding, to improve the generalization ability of the 3D target detection model. S4. Perform three-dimensional spatial clustering on the radar point cloud after general enhancement preprocessing, extract the effective target area, generate supplementary points in the target area based on 3D radial distribution rules, constrain the position of the supplementary points to the neighborhood of the original points, and perform feature assignment on the z coordinate, radar cross-section (RCS) and radial velocity of the supplementary points to preserve physical properties. S5. Project the radar point cloud after general enhancement preprocessing onto the BEV top view plane, extract the effective target area, generate supplementary points in the plane and fix the z coordinate, maintain the consistency in the height direction, and assign the radar cross-section (RCS) and radial velocity feature of the supplementary points in the same way as the 3D radial supplementary points. S6. Select some points of the point cloud after the point cloud is supplemented according to the preset probability as diffusion source points, generate diffusion points in the neighborhood of the diffusion source points to simulate the radar multipath scattering effect, and merge the diffusion points with the original point cloud. S7. Perform fixed-point sampling on the radar point cloud after local diffusion enhancement, and unify the number of output point clouds to the preset number of target points that meet the input format requirements of the 3D target detection model; S8. Input the radar point cloud with a fixed number of points into the 3D target detection model for training. At the same time, set up an independent testing phase. In the testing phase, turn off all enhancement operations and only perform fixed-number sampling on the original point cloud to output the detection results.

2. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, In S1, the radar point cloud data format is compatible with multiple mainstream radar target detection datasets. The three-dimensional spatial coordinates of x, y, and z represent the horizontal, vertical, and height positions of the radar point in three-dimensional space, respectively. The radial velocity vr represents the target's speed relative to the radar, and the radar cross-section RCS represents the target's ability to reflect radar electromagnetic waves.

3. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, In step S2, the basic parameter configuration for adaptation is automatically loaded, and the parameters include: Effective threshold of radar cross-section This is used to filter out invalid radar points with low signal-to-noise ratio and retain radar points with radar cross-section (RCS) values ​​greater than this threshold for subsequent processing. DBSCAN cluster radius , used to control the target region range of unsupervised clustering in density-based clustering algorithms, characterizing the maximum distance between two radar points in a cluster that can be classified into the same cluster; Minimum number of cluster points , used to remove isolated noise points, characterizes the minimum number of radar points required for an effective cluster; Radial compensation distance range This is used to limit the radial distance between the complement point and the cluster center, and the radial distance of the complement point is randomly selected within this interval; The amplitude of the z-coordinate perturbation of the supplementary point This is used to control the magnitude of the disturbance in the height direction of the supplementary point; BEV planar point-filling disturbance amplitude This is used to control the intensity of positional jitter at points within the top-view plane of the BEV. Radar cross-section RCS attenuation range This is used to limit the attenuation range of the supplementary radar cross-section. The RCS value of the supplementary radar cross-section is the RCS value of the original radar cross-section minus the random value within this range. Compensation point radial velocity noise amplitude This is used to control the amount of noise added to the radial velocity at the compensation point; Maximum number of complements in a single cluster This is used to control the maximum number of supplementary points for a single effective cluster, preventing excessive supplementary points from causing distortion in the point cloud distribution.

4. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, The preprocessing operations in S3 include: S3.1 Rotation operation: Perform a rotation transformation around the z-axis, randomly generating a rotation angle within the range of angles suitable for the dataset. The rotation matrix R is used to perform a rotation transformation on the x and y three-dimensional spatial coordinates of all radar points, while keeping the z coordinate unchanged. The rotation matrix R is defined as follows: in, Rotation angle cosine value, Rotation angle The sine value; S3.2 Scaling operation: Scaling is performed proportionally based on the center of the point cloud. First, the center coordinates c of the entire radar point cloud are calculated. Then, a scaling factor s is randomly generated within a preset scaling range. The x, y, and z coordinates of all radar points are scaled according to the center alignment rule. The scaling formula is: in, The spatial coordinates of radar point i before scaling. The scaled spatial coordinates of the radar point; S3.3 Random jitter operation: Gaussian noise is added to the position, radar cross-section (RCS), and radial velocity characteristics of the radar point to achieve slight perturbation of the characteristics. Among them, position jitter is based on the preset position noise amplitude by adding Gaussian noise to the x, y and z coordinates; radar cross-section RCS jitter is based on the preset radar cross-section RCS noise amplitude by adding Gaussian noise to the radar cross-section RCS feature; and radial velocity jitter is based on the preset velocity noise amplitude by adding Gaussian noise to the radial velocity vr feature, simulating the environmental noise and hardware measurement error in the actual detection process. S3.4 Random discard operation: randomly retain a portion of the radar point cloud according to a preset discard ratio range, and discard the remaining point cloud to simulate the scenario of point cloud loss caused by target obstruction and electromagnetic wave attenuation during radar detection.

5. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, S4 further includes: S4.

1. Perform 3D spatial clustering on the radar point cloud that has undergone general enhancement preprocessing based on the DBSCAN unsupervised clustering algorithm, combined with the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out noisy clusters and false clusters with excessively large velocity variances, and retain valid target clusters; S4.2 For each valid target cluster, calculate the coordinates of its cluster center. ,in, , and These are the coordinates of the cluster centers in the x, y, and z directions, respectively. S4.3, Using the cluster center as the pole, the radial distance range of the supplementary points is determined based on the parameters configured in S2. Randomly select a radial distance r, and randomly select a polar angle within the range [0, 2π]. The x and y coordinates of the supplementary points are generated according to the 3D radial distribution rule. The calculation formula is as follows: in, , These are the coordinates of the j-th supplementary point in the x and y directions, respectively; S4.4, The z-coordinate of the supplementary point corresponds to the z-coordinate of the original point. Based on this, add the z-coordinate perturbation amplitude of the parameter supplementation point based on the S2 configuration. Gaussian noise with standard deviation To achieve a slight perturbation in the height direction, the calculation formula is: in, Let be the coordinate value of the j-th supplementary point in the z-direction. Gaussian noise Follows a mean of 0 and a variance of The normal distribution; S4.5, The radar cross-section (RCS) of the supplementary radar is attenuated according to a preset range, expressed as: in, Let RCS be the radar cross-section value of the j-th supplementary point. The eigenvalues ​​corresponding to the original points. To randomly select attenuation values; The radial velocity at the supplementary point inherits from the original point and a small amount of noise is added, expressed as: in, Let J be the radial velocity value of the j-th supplementary point. Gaussian noise Follows a mean of 0 and a variance of It follows a normal distribution.

6. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, S5 further includes: S5.1 Project the radar point cloud that has completed general enhancement preprocessing onto the BEV top view plane, retaining the x and y two-dimensional coordinates and ignoring the z coordinate; S5.

2. Cluster the point cloud in the BEV plane based on the DBSCAN unsupervised clustering algorithm, and combine the DBSCAN clustering radius parameter configured in S2. Minimum number of cluster points and radar cross-section effective threshold Filter out invalid noise clusters and retain valid target clusters; S5.3 Within the effective target cluster, for the x and y coordinates of each original point, add the BEV plane perturbation amplitude based on the parameter configured in S2. Gaussian noise with standard deviation and The formula for generating the x and y coordinates of the supplementary point is: in, and Let x and y be the coordinates of the m-th BEV supplementary point, respectively. and These are the x-coordinate and y-coordinate of the corresponding original point, respectively. , Gaussian noise and All follow a mean of 0 and a variance of . The normal distribution; S5.4 The z-coordinate of the fixed complement point is the z-coordinate of the original point within its respective cluster. To ensure that all complementary points within the same cluster maintain a high degree of consistency with the original points, the calculation formula is as follows: in, Let be the coordinates of the m-th BEV supplementary point in the z-direction. This is the z-coordinate value of the nth original point within the cluster to which the supplementary point belongs; S5.5 The processing rules for the radar cross-section (RCS) features and radial velocity features of the supplemented points are consistent with those of the 3D radial supplementation in S4. That is, the radar cross-section (RCS) features are attenuated within a preset range, and the radial velocity features inherit the original points and add small Gaussian noise, ensuring that the physical characteristics of the two supplementation strategies are preserved in a unified manner.

7. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, S6 further includes: S6.1, according to the preset diffusion probability For point clouds that have completed 3D radial point filling or BEV consistency point filling, local diffusion enhancement is performed, and some radar points are randomly selected as diffusion source points. S6.2 For each diffusion source point, randomly select a diffusion distance within a preset diffusion distance range. The diffusion angle is randomly selected within the range of [0, 2π]. The x and y coordinates of the diffusion point are generated by the following formula: in, and These are the coordinates of the diffusion point in the x and y directions, respectively. and These are the coordinates of the k-th diffusion source point in the x and y directions, respectively. For diffusion angle cosine value, For diffusion angle The sine value; S6.3 The z-coordinate of the diffusion point is based on the z-coordinate of the diffusion source point, with a small amount of Gaussian noise added to simulate slight height perturbation; S6.

4. The radar cross-section (RCS) feature of the diffusion point is slightly attenuated, and a small amount of Gaussian noise is added to the radial velocity feature to ensure that the physical characteristics of the diffusion point are consistent with the original point cloud and to avoid feature distortion. S6.

5. Merge all generated diffusion points with the point cloud after 3D radial point supplementation or BEV consistency point supplementation to form a radar point cloud with local diffusion enhancement.

8. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, S7 further includes: S7.1 Count the actual number of radar point clouds after local diffusion enhancement. Set the number of target points according to the input requirements of the 3D target detection model. ; S7.2, If the actual number of points >Target Points Then, a random sampling method without replacement is used to randomly select points from the radar point cloud after local diffusion enhancement. Each point ensures the randomness and representativeness of the point cloud; S7.3, If the actual number of points <Target Points A fixed number of point clouds is generated by repeated sampling and remainder padding, and the calculation formula is as follows: in, This is the final point cloud after sampling. For the enhanced point cloud, Let d represent the number of repeated samples. The remainder To fill in the remaining points from the point cloud after local diffusion enhancement.

9. The dual-modal point cloud augmentation method for radar point clouds as described in claim 1, characterized in that, S8 further includes: S8.1 During the training phase, the enhanced fixed-point radar point cloud is used as the input to the 3D target detection model. Combined with the corresponding target annotation information, the model is trained according to the conventional training process, so that the model learns the point cloud features after the point enhancement and improves the model's ability to detect sparse radar point clouds. S8.2 During the testing phase, all point cloud enhancement operations are turned off, including the general enhancement preprocessing in S3, the 3D radial point filling in S4, the BEV consistency point filling in S5, and the local diffusion enhancement in S6. Only a fixed number of points are sampled on the original point cloud to ensure the originality of the test data. S8.3 Input the processed test point cloud into the trained 3D object detection model and output the final 3D object detection result.

10. A computer device, characterized in that, The system includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the dual-modal point enhancement method for radar point clouds as described in any one of claims 1-9.

Images