LiDAR 3D target detection optimization method for small target

Abstract

The invention discloses a LiDAR 3D target detection optimization method for a small target, and the method comprises the steps: collecting original point cloud data through a laser radar, carrying out the voxelization processing, and generating a three-dimensional voxel feature tensor; inputting the three-dimensional voxel feature tensor into a point cloud target detection model to obtain a 3D target detection frame parameter and a target category; positive and negative samples are generated based on a dynamic sample distribution strategy, and the model is trained; calculating a total loss function according to a detection result and a manually labeled label, and further optimizing the model; and finally, reasoning the to-be-detected point cloud subjected to voxelization processing by using the optimized model to realize accurate detection of a small target. According to the method, through dynamic sample distribution and loss function optimization, the detection precision and robustness of the small target in the LiDAR point cloud are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and 3D target detection, and in particular to an optimized LiDAR 3D target detection method for small targets. Background Technology

[0002] 3D target detection plays a crucial role in environmental perception systems. LiDAR (Light Laser Detection and Ranging), as a key sensor, is widely used in various environmental perception tasks due to its high-precision 3D point cloud data. However, the sparsity of LiDAR point cloud data itself presents significant challenges in target detection. Especially when detecting small targets, distant targets, and obstacles in complex environments, the accuracy and robustness of existing technologies often fall short of practical requirements. With increasing demands for accuracy and reliability, improving the processing performance of LiDAR data, particularly in dynamic and complex scenarios, has become a critical issue that urgently needs to be addressed.

[0003] Traditional voxel-based methods achieve target detection by dividing point cloud data into uniform 3D grids and then processing the voxel data. However, these methods have several shortcomings: First, existing techniques usually rely on coarse and static supervision signals, which makes the learning effect poor when dealing with small or distant targets that are more difficult to handle. Second, in the process of processing sparse point cloud data, due to the lack of effective multi-scale feature fusion, existing methods often cannot make full use of spatial information at different scales, which limits the performance of target detection in complex environments. Summary of the Invention

[0004] This invention provides an optimized LiDAR 3D target detection method for small targets, which overcomes the problems of poor learning effect when dealing with small or distant targets that are difficult to handle, and the limited target detection performance due to the inability to fully utilize spatial information at different scales.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: An optimization method for LiDAR 3D target detection for small targets, comprising: S1. Collect raw point cloud data using lidar and perform voxelization processing to generate a three-dimensional voxel feature tensor. S2. Input the 3D voxel feature tensor into the point cloud target detection model, and process it sequentially through the 3D feature extraction BEV generation module, the sparse bidirectional feature fusion module, and the multi-task learning target detection head network to obtain the parameters of the 3D target detection box and identify the category of the target. S3. Based on the dynamic sample allocation strategy and the fused sparse BEV feature map output by the sparse bidirectional feature fusion module, generate positive and negative samples, and train the point cloud target detection model to obtain the trained point cloud target detection model. S4. Based on the parameters of the 3D target detection box and the category of the identified target and the manually labeled point cloud in the original point cloud data, calculate the total loss function, and then optimize the trained point cloud target detection model to obtain the optimized point cloud target detection model. S5. The raw point cloud data to be processed is voxelized, and the detection results of small targets are finally obtained through the optimized point cloud target detection model.

[0006] Furthermore, the point cloud target detection model includes a 3D feature extraction BEV generation module, a sparse bidirectional feature fusion module, and a multi-task learning target detection head network; The 3D feature extraction BEV generation module is used to extract features from the three-dimensional voxel feature tensor through the 3D feature extraction skeleton to obtain a multi-scale feature map; and to project the multi-scale feature map onto the BEV space to generate a multi-scale sparse BEV feature map. The sparse bidirectional feature fusion module is used to perform sparse weighted fusion of multi-scale sparse BEV feature maps to obtain a fused sparse BEV feature map. The multi-task learning target detection head network is used to perform classification operations on the fused sparse BEV feature map through the multi-task detection head to obtain the category of the target; at the same time, it obtains the parameters of the 3D target detection box by performing regression processing on the fused sparse BEV feature map.

[0007] Furthermore, the expression for projecting the multi-scale feature map into the BEV space is:

[0008]

[0009] In the formula, For batch indexing after projection and planar coordinates BEV feature vector at location; Located in four-dimensional coordinates The input feature vector of the active voxel at the location, wherein the active voxel is a voxel unit containing at least one original point cloud data point in a three-dimensional voxel grid; The set of coordinates of the active BEVs after projection; This function performs a deduplication operation on the projected coordinate indices to obtain unique sparse BEV spatial indices.

[0010] Furthermore, the sparse bidirectional feature fusion module includes a top-down path, a bottom-up path, and an output module; The top-down path is fused using the following formula:

[0011] In the formula, This is for sparse upsampling and feature refinement operations; This is a third-scale top-down intermediate feature that incorporates high-level semantic information; The third-scale sparse BEV feature is a multi-scale sparse BEV feature map. The fourth-scale sparse BEV feature is a multi-scale sparse BEV feature map. This is the top-down path output feature for the second scale; The second-scale sparse BEV feature is a multi-scale sparse BEV feature map. For sparse feature fusion operators; The bottom-up path is fused using the following formula:

[0012] In the formula, This indicates a sparse downsampling operation; This represents the final output feature at the third scale after bidirectional fusion. This is the final output feature of the fourth scale after bidirectional fusion; The output module is used to output the final features at the fourth scale. Convolution, batch normalization, and ReLU activation are performed to generate a fused sparse BEV feature map.

[0013] Furthermore, the specific execution steps of the dynamic sample allocation strategy are as follows: S31. Use all non-empty voxels in the fused sparse BEV feature map output by the sparse bidirectional feature fusion module as candidate anchor boxes. S32. Based on the manually annotated ground truth (GT) boxes, calculate the two-dimensional Euclidean distance between the center point of the GT box and the center point of the candidate anchor boxes; sort the two-dimensional Euclidean distances in ascending order and select the top K candidate anchor boxes as the initial set of positive candidate boxes. S33. Using each positive candidate box in the initial positive candidate box set as the center, and through the size of the corresponding GT box, and uniformly rotating it around the z-axis by 12 rotation angles, generate a set of hypothesis boxes; the hypothesis boxes are:

[0014] In the formula, It is the k-th hypothesis box; k is the index of the hypothesis box; It is the j-th candidate anchor box The center coordinates; It's a GT frame. The center height; It's a GT frame. Length, width, and height; It is the first Rotation angle in each sampling direction; S34. Calculate the 3D intersection-union ratio between each hypothetical box and its corresponding ground truth (GT) box to obtain a set of 3D intersection-union ratio values ​​between the hypothetical box and its corresponding GT box; based on the set of 3D intersection-union ratio values, obtain the maximum 3D intersection-union ratio value of the hypothetical box, and use it as the 3D intersection-union ratio value between the positive candidate box and the GT box. The expression for calculating the 3D intersection-union ratio of each hypothesis box and its corresponding ground truth box is as follows:

[0015] In the formula, It refers to the first Hypothesis boxes With GT frame IoU between; The maximum 3D intersection-union ratio of the hypothesis box is:

[0016] In the formula, This represents the maximum IoU value of the hypothesis box across all sampling directions; S35. Calculate the set of 3D intersection-union ratios (IURs) of positive candidate boxes and ground truth (GT) boxes, and use the median as the adaptive IUR threshold. The expression is:

[0017] In the formula, This represents the median; This represents the initial set of positive candidate boxes; This indicates the operation of retrieving the median of the values ​​in the set; S36. If the 3D intersection-union ratio (IU) of the positive candidate box and the ground truth box is greater than or equal to the adaptive IU threshold, then the positive candidate box is labeled as a positive sample. The expression is:

[0018] In the formula, This is the final set of positive samples; Otherwise, label the positive candidate box as a negative sample.

[0019] Furthermore, the expression for calculating the two-dimensional Euclidean distance between the center point of the ground truth bounding box and the center point of the candidate anchor box is as follows:

[0020] In the formula, The distance between the i-th ground truth bounding box and the j-th candidate anchor box; For the i-th ground truth (GT) frame; This is the j-th candidate anchor box.

[0021] Furthermore, after step S36, the positive candidate boxes labeled as positive samples are traversed. If a positive candidate box is labeled as a positive sample by multiple ground truth (GT) boxes, the positive candidate box is assigned to the GT box with the largest 3D intersection-union ratio, expressed as:

[0022] In the formula, Represented as anchor frame The index of the assigned GT frame; This indicates that the GT box with the largest IoU value is selected.

[0023] Beneficial effects: This invention provides an optimized LiDAR 3D target detection method for small targets. By employing a sparse bidirectional feature fusion module to perform weighted fusion of multi-scale sparse BEV feature maps, it can fully utilize the contextual information under different receptive fields. While preserving detailed features (which are beneficial for small target localization), it integrates semantic features (which are beneficial for target classification), thereby improving the feature map's ability to represent small targets and enhancing its adaptability to the overall scene. By employing a dynamic sample allocation strategy, the generation and allocation mechanism of positive and negative samples is optimized to address the issues of sparse distribution and insufficient features of small targets in point clouds. This allows the model training to focus more on small target regions, effectively enhancing the model's ability to identify and locate small-scale, low-density targets, and significantly improving detection accuracy and recall. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the target detection optimization method of the present invention; Figure 2 This is a schematic diagram of the 3D feature extraction skeleton structure in an embodiment of the present invention; Figure 3 This is a schematic diagram of the sparse bidirectional feature fusion module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dynamic sample allocation process in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This embodiment provides an optimized LiDAR 3D target detection method for small targets, such as... Figure 1 As shown, it includes: S1. Collect raw point cloud data using lidar and perform voxelization processing to generate a three-dimensional voxel feature tensor. S2. Input the 3D voxel feature tensor into the point cloud target detection model, and process it sequentially through the 3D feature extraction BEV generation module, the sparse bidirectional feature fusion module, and the multi-task learning target detection head network to obtain the parameters of the 3D target detection box and identify the category of the target. S3. Based on the dynamic sample allocation strategy and the fused sparse BEV feature map output by the sparse bidirectional feature fusion module, generate positive and negative samples, and train the point cloud target detection model to obtain the trained point cloud target detection model. S4. Based on the parameters of the 3D target detection box and the category of the identified target and the manually labeled point cloud in the original point cloud data, calculate the total loss function, and then optimize the trained point cloud target detection model to obtain the optimized point cloud target detection model. S5. The raw point cloud data to be processed is voxelized, and the detection results of small targets are finally obtained through the optimized point cloud target detection model.

[0028] In this embodiment, the small target is a target object in the original point cloud data whose number of points is less than a preset threshold (15 points).

[0029] Specifically, the voxelization process involves mapping the original point cloud data onto a regular three-dimensional grid of a preset size, and extracting features from non-empty voxels containing point cloud data to obtain a three-dimensional voxel feature tensor. Specifically, the grid coordinate index to which each point belongs is determined by point cloud coordinate quantization, and the geometric centroid and average reflection intensity of all points within each non-empty voxel are calculated to generate a mean feature vector. The three-dimensional voxel feature tensor structurally includes: a list of coordinate indices that identify the spatial location of non-empty voxels, and a list of mean feature vectors that correspond one-to-one with the coordinate indices.

[0030] Preferably, the point cloud target detection model includes a 3D feature extraction BEV generation module, a sparse bidirectional feature fusion module, and a multi-task learning target detection head network; The 3D feature extraction BEV generation module is used to extract features from the three-dimensional voxel feature tensor through the 3D feature extraction skeleton to obtain a multi-scale feature map; and to project the multi-scale feature map onto the BEV space to generate a multi-scale sparse BEV feature map. The sparse bidirectional feature fusion module is used to perform sparse weighted fusion of multi-scale sparse BEV feature maps to obtain a fused sparse BEV feature map. The multi-task learning target detection head network is used to perform classification operations on the fused sparse BEV feature map through the multi-task detection head to obtain the category of the target; at the same time, it obtains the parameters of the 3D target detection box by performing regression processing on the fused sparse BEV feature map.

[0031] Specifically, the 3D feature extraction skeleton is as follows: Figure 2 As shown, through a top-down unidirectional information flow and multiple downsampling operations, the multi-scale features of the three-dimensional voxel feature tensor are obtained. The specific steps are existing technologies available to those skilled in the art and will not be elaborated here. The multi-task learning object detection head network includes a classification task branch and a regression task branch. The classification task branch identifies the target category in the fused BEV feature map through multiple convolutional layers and a classification convolutional layer. The regression task branch predicts various parameters of the detection box in the fused BEV feature map through multiple convolutional layers and a regression layer to obtain the parameters of the 3D target detection box. The specific steps of the multi-task learning object detection head network are existing technologies available to those skilled in the art and will not be elaborated here.

[0032] Preferably, the expression for projecting the multi-scale feature map into the BEV space is:

[0033]

[0034] In the formula, For batch indexing after projection and planar coordinates BEV feature vector at location; Located in four-dimensional coordinates The input feature vector of the active voxel at the location, wherein the active voxel is a voxel unit containing at least one original point cloud data point in a three-dimensional voxel grid; The set of coordinates of the active BEVs after projection; This function performs a deduplication operation on the projected coordinate indices to obtain unique sparse BEV spatial indices.

[0035] In this embodiment, the BEV space is a sparse 2D feature representation plane that carries height information; the projected active BEV coordinate set is a set of non-empty positions of the multi-scale sparse BEV feature map projected into the BEV space.

[0036] Specifically, such as Figure 3 As shown, the sparse bidirectional feature fusion module is used to receive two or more projected sparse tensors as input and output a sparse tensor. The active coordinates of the output sparse tensor are the union of the coordinates of each input tensor. For regions with overlapping coordinates, weighted fusion is performed using learnable weights. For non-overlapping coordinates, they are directly retained to achieve context-aware feature fusion only in non-empty regions.

[0037] Preferably, the sparse bidirectional feature fusion module includes a top-down path, a bottom-up path, and an output module; The top-down path is fused using the following formula:

[0038] In the formula, This is for sparse upsampling and feature refinement operations; This is a third-scale top-down intermediate feature that incorporates high-level semantic information; The third-scale sparse BEV feature is a multi-scale sparse BEV feature map. The fourth-scale sparse BEV feature is a multi-scale sparse BEV feature map. This is the top-down path output feature for the second scale; The second-scale sparse BEV feature is a multi-scale sparse BEV feature map. For sparse feature fusion operators; The bottom-up path is fused using the following formula:

[0039] In the formula, This indicates a sparse downsampling operation; This represents the final output feature at the third scale after bidirectional fusion. This is the final output feature of the fourth scale after bidirectional fusion; The output module is used to output the final features at the fourth scale. Convolution, batch normalization, and ReLU activation are performed to generate a fused sparse BEV feature map.

[0040] Preferred, such as Figure 4 As shown, the specific execution steps of the dynamic sample allocation strategy are as follows: S31. Use all non-empty voxels in the fused sparse BEV feature map output by the sparse bidirectional feature fusion module as candidate anchor boxes. S32. Based on the manually annotated ground truth (GT) boxes, calculate the two-dimensional Euclidean distance between the center point of the GT box and the center point of the candidate anchor boxes; sort the two-dimensional Euclidean distances in ascending order and select the top K candidate anchor boxes as the initial set of positive candidate boxes. S33. Using each positive candidate box in the initial positive candidate box set as the center, and through the size of the corresponding GT box, and uniformly rotating it around the z-axis by 12 rotation angles, generate a set of hypothesis boxes; the hypothesis boxes are:

[0041] In the formula, It is the k-th hypothesis box; k is the index of the hypothesis box; It is the j-th candidate anchor box The center coordinates; It's a GT frame. The center height; It's a GT frame. Length, width, and height; It is the first Rotation angle in each sampling direction; S34. Calculate the 3D intersection-union ratio between each hypothetical box and its corresponding ground truth (GT) box to obtain a set of 3D intersection-union ratio values ​​between the hypothetical box and its corresponding GT box; based on the set of 3D intersection-union ratio values, obtain the maximum 3D intersection-union ratio value of the hypothetical box, and use it as the 3D intersection-union ratio value between the positive candidate box and the GT box. The expression for calculating the 3D intersection-union ratio of each hypothesis box and its corresponding ground truth box is as follows:

[0042] In the formula, It refers to the first Hypothesis boxes With GT frame IoU between; The maximum 3D intersection-union ratio of the hypothesis box is:

[0043] In the formula, This represents the maximum IoU value of the hypothesis box across all sampling directions; S35. Calculate the set of 3D intersection-union ratios (IURs) of positive candidate boxes and ground truth (GT) boxes, and use the median as the adaptive IUR threshold. The expression is:

[0044] In the formula, This represents the median; This represents the initial set of positive candidate boxes; This indicates the operation of retrieving the median of the values ​​in the set; S36. If the 3D intersection-union ratio (IU) of the positive candidate box and the ground truth box is greater than or equal to the adaptive IU threshold, then the positive candidate box is labeled as a positive sample. The expression is:

[0045] In the formula, This is the final set of positive samples; Otherwise, label the positive candidate box as a negative sample.

[0046] Preferably, the expression for calculating the two-dimensional Euclidean distance between the center point of the ground truth bounding box and the center point of the candidate anchor box is:

[0047] In the formula, The distance between the i-th ground truth bounding box and the j-th candidate anchor box; For the i-th ground truth (GT) frame; This is the j-th candidate anchor box.

[0048] Preferably, after step S36, the positive candidate boxes labeled as positive samples are traversed. If a positive candidate box is labeled as a positive sample by multiple ground truth (GT) boxes, the positive candidate box is assigned to the GT box with the largest 3D intersection-union ratio, expressed as:

[0049] In the formula, Represented as anchor frame The index of the assigned GT frame; This indicates that the GT box with the largest IoU value is selected.

[0050] In a specific embodiment, a total loss function is calculated based on the parameters of the 3D target detection box and the category of the identified target, as well as the manually annotated point cloud labels in the original point cloud data. The total loss function includes classification loss and regression loss. The classification loss is calculated by the difference between the category of the identified target and the true category label in the manually annotated point cloud data. The regression loss is calculated by the difference between the parameters of the 3D target detection box and the parameters of the true annotation boxes in the manually annotated point cloud data. The gradient backpropagation algorithm is used to calculate the gradient of the total loss with respect to the training parameters in the trained point cloud object detection model. The optimizer is then used to update the training parameters of the trained point cloud object detection model based on the calculated gradient of the training parameters until the performance of the point cloud object detection model converges. The total loss function, gradient backpropagation algorithm, and optimizer mentioned are existing technologies and will not be described in detail here.

[0051] The present invention has the following beneficial effects: This invention provides an optimized LiDAR 3D target detection method for small targets. By employing a sparse bidirectional feature fusion module to perform weighted fusion of multi-scale sparse BEV feature maps, it can fully utilize the contextual information under different receptive fields, retain detailed features (which are beneficial for small target localization) while integrating semantic features (which are beneficial for target classification), thereby improving the feature map's ability to represent small targets and enhancing its adaptability to the overall scene. By employing a dynamic sample allocation strategy, the generation and allocation mechanism of positive and negative samples is optimized to address the issues of sparse distribution and insufficient features of small targets in point clouds. This allows the model training to focus more on small target regions, effectively enhancing the model's ability to identify and locate small-scale, low-density targets, and significantly improving detection accuracy and recall.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An optimized LiDAR 3D target detection method for small targets, characterized in that, include: S1. Collect raw point cloud data using lidar and perform voxelization processing to generate a three-dimensional voxel feature tensor. S2. Input the 3D voxel feature tensor into the point cloud target detection model, and process it sequentially through the 3D feature extraction BEV generation module, the sparse bidirectional feature fusion module, and the multi-task learning target detection head network to obtain the parameters of the 3D target detection box and identify the category of the target. S3. Based on the dynamic sample allocation strategy and the fused sparse BEV feature map output by the sparse bidirectional feature fusion module, generate positive and negative samples, and train the point cloud target detection model to obtain the trained point cloud target detection model. S4. Based on the parameters of the 3D target detection box and the category of the identified target and the manually labeled point cloud in the original point cloud data, calculate the total loss function, and then optimize the trained point cloud target detection model to obtain the optimized point cloud target detection model. S5. The raw point cloud data to be processed is voxelized, and the detection results of small targets are finally obtained through the optimized point cloud target detection model.

2. The LiDAR 3D target detection optimization method for small targets according to claim 1, characterized in that, The point cloud target detection model includes a 3D feature extraction BEV generation module, a sparse bidirectional feature fusion module, and a multi-task learning target detection head network. The 3D feature extraction BEV generation module is used to extract features from the three-dimensional voxel feature tensor through the 3D feature extraction skeleton to obtain a multi-scale feature map; and to project the multi-scale feature map onto the BEV space to generate a multi-scale sparse BEV feature map. The sparse bidirectional feature fusion module is used to perform sparse weighted fusion of multi-scale sparse BEV feature maps to obtain a fused sparse BEV feature map. The multi-task learning target detection head network is used to perform classification operations on the fused sparse BEV feature map through the multi-task detection head to obtain the category of the target; at the same time, it obtains the parameters of the 3D target detection box by performing regression processing on the fused sparse BEV feature map.

3. The LiDAR 3D target detection optimization method for small targets according to claim 2, characterized in that, The expression for projecting multi-scale feature maps into the BEV space is: In the formula, For batch indexing after projection and planar coordinates BEV feature vector at location; Located in four-dimensional coordinates The input feature vector of the active voxel at the location, wherein the active voxel is a voxel unit containing at least one original point cloud data point in a three-dimensional voxel grid; The set of coordinates of the active BEVs after projection; This function performs a deduplication operation on the projected coordinate indices to obtain unique sparse BEV spatial indices.

4. The LiDAR 3D target detection optimization method for small targets according to claim 2, characterized in that, The sparse bidirectional feature fusion module includes a top-down path, a bottom-up path, and an output module; The top-down path is fused using the following formula: In the formula, This is for sparse upsampling and feature refinement operations; This is a third-scale top-down intermediate feature that incorporates high-level semantic information; The third-scale sparse BEV feature is a multi-scale sparse BEV feature map. The fourth-scale sparse BEV feature is a multi-scale sparse BEV feature map. This is the top-down path output feature for the second scale; The second-scale sparse BEV feature is a multi-scale sparse BEV feature map. For sparse feature fusion operators; The bottom-up path is fused using the following formula: In the formula, This indicates a sparse downsampling operation; This represents the final output feature at the third scale after bidirectional fusion. This is the final output feature of the fourth scale after bidirectional fusion; The output module is used to output the final features at the fourth scale. Convolution, batch normalization, and ReLU activation are performed to generate a fused sparse BEV feature map.

5. The LiDAR 3D target detection optimization method for small targets according to claim 1, characterized in that, The specific execution steps of the dynamic sample allocation strategy are as follows: S31. Use all non-empty voxels in the fused sparse BEV feature map output by the sparse bidirectional feature fusion module as candidate anchor boxes. S32. Based on the manually annotated ground truth (GT) boxes, calculate the two-dimensional Euclidean distance between the center point of the GT box and the center point of the candidate anchor boxes; sort the two-dimensional Euclidean distances in ascending order and select the top K candidate anchor boxes as the initial set of positive candidate boxes. S33. Using each positive candidate box in the initial positive candidate box set as the center, and through the size of the corresponding GT box, and uniformly rotating it around the z-axis by 12 rotation angles, generate a set of hypothesis boxes; the hypothesis boxes are: In the formula, It is the k-th hypothesis box; k is the index of the hypothesis box; It is the j-th candidate anchor box The center coordinates; It's a GT frame. The center height; It's a GT frame. Length, width, and height; It is the first Rotation angle in each sampling direction; S34. Calculate the 3D intersection-union ratio between each hypothetical box and its corresponding ground truth (GT) box to obtain a set of 3D intersection-union ratio values ​​between the hypothetical box and its corresponding GT box; based on the set of 3D intersection-union ratio values, obtain the maximum 3D intersection-union ratio value of the hypothetical box, and use it as the 3D intersection-union ratio value between the positive candidate box and the GT box. The expression for calculating the 3D intersection-union ratio of each hypothesis box and its corresponding ground truth box is as follows: In the formula, It refers to the first Hypothesis boxes With GT frame IoU between; The maximum 3D intersection-union ratio of the hypothesis box is: In the formula, This represents the maximum IoU value of the hypothesis box across all sampling directions; S35. Calculate the set of 3D intersection-union ratios (IURs) of positive candidate boxes and ground truth (GT) boxes, and use the median as the adaptive IUR threshold. The expression is: In the formula, This represents the median; This represents the initial set of positive candidate boxes; This indicates the operation of retrieving the median of the values ​​in the set; S36. If the 3D intersection-union ratio (IU) of the positive candidate box and the ground truth box is greater than or equal to the adaptive IU threshold, then the positive candidate box is labeled as a positive sample. The expression is: In the formula, This is the final set of positive samples; Otherwise, label the positive candidate box as a negative sample.

6. The LiDAR 3D target detection optimization method for small targets according to claim 5, characterized in that, The expression for calculating the two-dimensional Euclidean distance between the center point of the ground truth bounding box and the center point of the candidate anchor box is: In the formula, The distance between the i-th ground truth bounding box and the j-th candidate anchor box; For the i-th ground truth (GT) frame; Let j be the j-th candidate anchor box.

7. The LiDAR 3D target detection optimization method for small targets according to claim 5, characterized in that, After step S36, the positive candidate boxes labeled as positive samples are traversed. If a positive candidate box is labeled as a positive sample by multiple ground truth (GT) boxes, the positive candidate box is assigned to the GT box with the largest 3D intersection-union ratio, expressed as: In the formula, Represented as anchor frame The index of the assigned GT frame; This indicates that the GT box with the largest IoU value is selected.

Images