Roadend sparse point cloud traffic small target detection method based on multilayer pruning and density perception enhancement
By employing multi-layer pruning and density-aware enhancement methods in roadside point cloud detection, a sparse voxel tensor is constructed and dynamically pruned. Combining point density and reflection intensity information, the problem of low accuracy and efficiency in small target detection under sparse point cloud environment is solved, achieving efficient and accurate small target detection.
Patent Information
- Application Number
- CN202510832307.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-11-18
AI Technical Summary
Existing roadside point cloud detection methods have low accuracy and efficiency in detecting small targets in sparse point cloud environments, and waste significant computational resources, failing to effectively focus on real small target areas.
We employ a multi-layer pruning and density-sensing enhancement approach. By constructing a sparse voxel tensor, we dynamically prune non-target regions and combine point density and reflection intensity information with multi-scale voxel pyramids and detection branches to optimize the detection process.
It improves the accuracy and efficiency of small target detection, reduces redundant computation, enhances the feature representation of small targets, realizes efficient detection of targets at different scales, and ensures the real-time performance and robustness of the method.
Smart Images
Figure CN120976515A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of point cloud target detection, and particularly relates to a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement. BACKGROUND
[0002] In the roadside traffic scene, three-dimensional point cloud target detection technology is of great significance to improve traffic safety and traffic management efficiency. Laser radar can provide high-precision three-dimensional spatial information. However, since the roadside laser radar is usually installed at a high position, the point cloud data generated by it shows obvious sparseness in the long distance and low area, which makes the geometric structure information of small targets insufficient. The existing detection methods mainly divide the point cloud data into fixed resolution grids through voxelization processing, and then use sparse convolution network for feature extraction and target detection. Although these methods can process sparse point cloud data to some extent, there are still deficiencies in small target detection. Specifically, the conventional voxelization processing further weakens the detail information of small targets, and the multi-layer down-sampling operation may cause the small target features to be lost in the encoding stage, so that the detection head cannot accurately regress the position of the small target. In addition, the existing methods uniformly process the global point cloud, which not only leads to waste of computing resources, but also disperses the limited attention weight, which cannot focus on the real small target area, reducing the detection sensitivity. These problems seriously affect the detection accuracy and efficiency of the existing technology for traffic small targets in the road end point cloud environment. SUMMARY
[0003] The present application aims at: for this purpose, a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement is proposed, which solves the technical problems of poor detection accuracy and low efficiency of traffic small targets in the existing road end point cloud environment.
[0004] Specifically, the present application provides a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement, comprising the following steps: S1, acquiring the original point cloud data of the roadside laser radar and preprocessing to obtain the ROI point cloud; S2, constructing a sparse voxel tensor according to the ROI point cloud; S3, using a sparse convolution backbone to extract L4→L3→L2→L1 four-layer features from the sparse voxel tensor; S4, configuring a detection head on the middle layer L3, predicting a center point heat map, and extracting a candidate center point set P coarse ; S5, constructing a mask according to each candidate point of the candidate center point set, and merging all the masks to generate a three-dimensional mask; using the three-dimensional mask to filter the voxels of L2 and L1 layers to obtain the reserved voxels; S6, after splicing the reserved voxel statistical point density D and the average reflection intensity I with the original features, inputting an MLP or a voxel-Attention module for weighting to generate enhanced features; S7, in the reserved voxel region of each layer, a detection head is configured based on the enhanced features, and target box information is respectively predicted; non-maximum suppression (NMS) is performed on the outputs of each layer from a high-altitude perspective to obtain a final small target bounding box set.
[0005] A storage device stores instructions and data for implementing a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement.
[0006] A road end sparse point cloud traffic small target detection device based on multi-layer pruning and density perception enhancement includes a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement.
[0007] The road side point cloud traffic small target detection method provided by the application has the beneficial effects that: the road side point cloud traffic small target detection method provided by the application reduces redundant calculation through dynamic spatial pruning, improves calculation efficiency, introduces point density and reflection intensity information, and combines an attention mechanism to enhance the feature expression of small targets and improve detection accuracy, constructs a multi-scale voxel pyramid and deploys detection branches at each layer to realize efficient detection of targets of different scales, adopts a center expansion and IoU-free matching strategy in cooperation with Focal Loss to optimize the stability and recall rate of small target training, sets a reasonable ROI for a road end scene to ensure the real-time performance and robustness of the method, and enables the method to efficiently adapt to a road side traffic environment. BRIEF DESCRIPTION OF DRAWINGS
[0008] Figure 1 is a flowchart of the road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement of the application; Figure 2 is a sparse voxel pyramid encoding and decoding process diagram in the embodiment of the application; Figure 3 is a mask construction flowchart in the embodiment of the application; Figure 4 is a density enhancement module structure diagram in the embodiment of the application; Figure 5 is a dynamic spatial pruning mechanism diagram in the embodiment of the application; Figure 6 is a hardware device working diagram of the application. DETAILED DESCRIPTION
[0009] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be further described below with reference to the drawings.
[0010] Before formally describing the present application, the scheme of the present application is first described generally for convenience of understanding.
[0011] Please refer to Figure 1 The present application provides a road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement, comprising the following steps: S1, acquiring road side laser radar original point cloud data and pre-processing to obtain ROI point cloud; It should be noted that the pre-processing in step S1 includes noise elimination, ground separation, coordinate normalization and ROI setting.
[0012] Specifically, the present application receives road side LiDAR original point cloud, performs noise elimination, ground separation, coordinate normalization, and sets spatial ROI according to road geometry, and retains the ROI point cloud for subsequent processing.
[0013] For example, the set ROI region is: x∈[0, 50m], y∈[-15m, 15m], z∈[0, 3m], only the point cloud in this region is retained for subsequent processing.
[0014] S2, constructing a sparse voxel tensor according to the ROI point cloud; It should be noted that step S2 is specifically: the ROI point cloud is divided into a voxel grid with a set size, the initial features of each voxel are generated by using maximum or average pooling, and a sparse voxel tensor is constructed.
[0015] The present application divides the processed ROI point cloud into a sparse voxel grid at a fixed resolution, and constructs a sparse tensor based on the aggregated features of the points in the voxel, as the input of the backbone network.
[0016] S3, using a sparse convolution backbone to extract L4→L3→L2→L1 four-layer features from the sparse voxel tensor; It should be noted that step S3 is specifically: S31, using a sparse convolution network to extract coarse semantic features from the L4 layer voxel graph; S32, using transposed sparse convolution for upsampling to map the retained voxels of the upper layer to a higher resolution space; S33, fusing with the current layer backbone network features to construct a pyramid multi-scale voxel feature map.
[0017] The voxel pyramid feature extraction and upsampling decoding are described in Figure 2As shown in the structure, the backbone network extracts voxel feature maps from L4 to L1 through multi-layer sparse convolution, and the decoding stage up-samples the reserved area and fuses the backbone features of the same layer to generate a pyramid multi-scale voxel feature map.
[0018] S4, configure a detection head on the middle layer L3, predict a center point heat map, and extract a candidate center point set P through threshold value or non-maximum suppression coarse ; It should be noted that step S4 is specifically: S41, set an anchor-free detection head in the middle layer L3 feature map, and respectively predict a center point probability heat map, a category probability and a size offset for each voxel; S42, perform non-maximum suppression NMS and confidence threshold screening on the center point probability heat map to obtain a candidate center point set P coarse . Please refer to Figure 3 , Figure 3 is a structure diagram of the candidate center point generation and mask construction process in the embodiment of the application, and specifically includes the following steps: Deploy an anchor-free detection head on the middle layer L3 feature map, and output a center point probability heat map, a size offset, a direction angle and other prediction results for each voxel position.
[0019] Perform non-maximum suppression (NMS) and confidence threshold screening on the center point heat map, extract a group of center point candidate positions with the highest confidence, and constitute a candidate point set P coarse .
[0020] S5, construct a mask according to each candidate point of the candidate center point set, and merge all the masks to generate a three-dimensional mask; use the three-dimensional mask to screen the voxels of the L2 and L1 layers to obtain reserved voxels; It should be noted that step S5 is specifically: S51, construct a mask M with each candidate point as the center of a sphere and a radius R, and merge all the spheres to generate a three-dimensional mask; Specifically, for each candidate point p∈P coarse , a spherical region B(p, R) with a radius of R is constructed to represent its influence range.
[0021] Take the union of all spherical regions to obtain a three-dimensional space mask M(x, y, z), wherein: .
[0022] Wherein, M(x, y, z) represents the voxel mask value of the coordinates (x, y, z); p represents the candidate center point coordinate vector; P coarseThe candidate target center point set is predicted by the anchor-free detection head on the refined L3 layer and obtained through non-maximum suppression and confidence threshold screening; R is a radius hyperparameter of a mask sphere, used to control the spatial range of the reserved region.
[0023] S52, using the three-dimensional mask map M, the L2 and L1 layer voxels are cropped, and only the M=1 region is reserved; The mask M is projected to the high-resolution feature map L2, L1, and the position alignment processing is performed, and the reserved voxel region is selected.
[0024] S53, the voxels in the pruning region continue to be calculated by upsampling and detection branch, and the rest is skipped.
[0025] S6, the point density D and the average reflection intensity I of the reserved voxels are counted, and after splicing with the original features, the MLP or voxel-Attention module is input for weighting to generate enhanced features; the enhanced features are used to replace the original voxel features as the input of the subsequent detection branch.
[0026] Specifically, please refer to Figure 4 , Figure 4 is the structure diagram of the density perception enhancement module of the application.
[0027] The module introduces two types of spatial structure information: point density and reflection intensity for the reserved voxel region. For each reserved voxel v i , let the point set contained in it be P i , then the point density of D i and the average reflection intensity I i are defined as follows:
[0028] The original voxel feature f raw ( v i ) and log(D i +1), I i are spliced and input into a three-layer perception machine MLP to generate enhanced features:
[0029] The enhanced features are then sent to the voxel-attention module for position perception weighting. Specifically, the voxel-attention module is designed as a lightweight attention module, and the input is multiple voxel features (f enh), the output is a weighted voxel representation (f att The module realizes focused modeling of the target region by calculating a semantic importance score of each voxel in the local receptive field.
[0030] S7, in each layer of the retained voxel region, a detection head is configured based on the enhanced feature configuration, and target box information is respectively predicted; non-maximum suppression (NMS) is performed on the outputs of the layers in a high-altitude overhead perspective to obtain a final small target bounding box set.
[0031] It should be noted that the step S7 includes: S71, a detection branch is deployed in each layer of the retained voxel region to predict three-dimensional bounding box parameters (x, y, z, l, w, h, yaw) and target class confidence; S72, after merging all detection results, non-maximum suppression (NMS) is performed in a high-altitude overhead BEV perspective to remove overlapping duplicate boxes, and a final detection box set is output.
[0032] Please refer to Figure 5 , Figure 5 is a structure diagram of the dynamic spatial pruning mechanism in the embodiment of the application.
[0033] The mechanism performs position filtering on the refined L2 and refined L1 feature maps based on the spatial mask M(x, y, z) generated in Figure 3 , only performs upsampling, feature fusion, and detection head processing on the retained voxel region with a mask value of 1, and skips non-target regions, thereby significantly reducing computational redundancy and improving modeling accuracy and efficiency for small targets. The figure shows how the retained voxel is expanded from the candidate points and forms a sparse pruning path on the feature map.
[0034] Please refer to Figure 6 , Figure 6 is a hardware device working schematic diagram of the embodiment of the application, and the hardware device specifically includes: a road end sparse point cloud traffic small target detection device 401 based on multi-layer pruning and density perception enhancement, a processor 402, and a storage device 403.
[0035] A road end sparse point cloud traffic small target detection device 401 based on multi-layer pruning and density perception enhancement: the road end sparse point cloud traffic small target detection device 401 based on multi-layer pruning and density perception enhancement implements the road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement.
[0036] The processor 402 loads and executes instructions and data in the storage device 403 for implementing the method for detecting small traffic targets based on multi-layer pruning and density-aware enhanced road sparse point cloud.
[0037] The storage device 403 stores instructions and data; the storage device 403 is used to implement the method for detecting small traffic targets based on multi-layer pruning and density-aware enhanced road sparse point cloud.
[0038] In summary, the beneficial effects of the present application are: the proposed road side point cloud traffic small target detection method reduces redundant calculation through dynamic spatial pruning, improves calculation efficiency. The point density and reflectivity information are introduced and combined with the attention mechanism to enhance the feature expression of small targets and improve the detection accuracy. A multi-scale voxel pyramid is constructed and detection branches are deployed at each layer to realize efficient detection of targets of different scales. The center expansion and IoU-free matching strategy are used in combination with FocalLoss to optimize the stability and recall rate of small target training. Reasonable ROI is set for the road end scene to ensure the real-time and robustness of the method, so that it can efficiently adapt to the road side traffic environment.
[0039] The above is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for detecting traffic small targets based on multi-layer pruning and density-aware enhanced road sparse point cloud, characterized in that: The method comprises the following steps: S1, acquiring roadside laser radar original point cloud data and preprocessing to obtain ROI point cloud; S2, constructing a sparse voxel tensor according to the ROI point cloud; S3, using a sparse convolution backbone to extract L4→L3→L2→L1 four-layer features from the sparse voxel tensor; S4, configure detection head on middle layer L3, predict center point heat map, and extract candidate center point set P through threshold or non-maximum suppression coarse ; S5, constructing a mask according to each candidate point of the candidate center point set, and merging all the masks to generate a three-dimensional mask; Using the three-dimensional mask to screen the voxels of the L2 and L1 layers to obtain reserved voxels; S6, after the point number density D and the average reflection intensity I of the reserved voxels are counted, the original features are spliced, and then input into an MLP or voxel-Attention module for weighting to generate enhanced features; S7, in the reserved voxel region of each layer, a detection head is configured based on the enhanced features, and target box information is respectively predicted; the outputs of each layer are subjected to non-maximum suppression NMS in a high-altitude overhead view angle to obtain a final small target bounding box set.
2. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. The preprocessing in step S1 includes noise removal, ground separation, coordinate normalization and ROI setting.
3. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. Step S2 specifically is: dividing the ROI point cloud into a voxel grid with a set size, using maximum or average pooling to generate initial features for each voxel, and constructing a sparse voxel tensor.
4. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. Step S3 specifically is: S31, using a sparse convolution network to extract coarse semantic features from the L4 layer voxel graph; S32, using transposed sparse convolution for upsampling to map the upper layer reserved voxels to a higher resolution space; S33, fusing with the current layer backbone network features to construct a pyramid multi-scale voxel feature map.
5. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. Step S4 specifically is: S41, setting an anchor-free detection head in the middle layer L3 feature map to respectively predict a center point probability heat map, a class probability and a size offset for each voxel; S42, performing non-maximum suppression (NMS) and confidence threshold screening on the center point probability heat map to obtain a candidate center point set P coarse。 6. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. Step S5 specifically is: S51, constructing a mask M with each candidate point as the center of a sphere and a radius R, and merging all the spheres to generate a three-dimensional mask; S52, using the above three-dimensional mask graph M to crop the L2 and L1 layer voxels, and only retaining the M=1 region; S53, the voxels in the pruning region continue to be calculated through upsampling and detection branches, and the rest are skipped.
7. The method of claim 1, wherein the method is based on multi-layer pruning and density-aware enhanced road sparse point cloud traffic small target detection. The step S7 comprises: S71, deploying a detection branch in each layer of the reserved voxel region to predict three-dimensional bounding box parameters (x, y, z, l, w, h, yaw) and target class confidence; S72, after merging all the detection results, performing non-maximum suppression NMS in a high-altitude overhead BEV view angle to remove overlapping duplicate boxes, and outputting a final detection box set.
8. A storage device, characterized by: The storage device stores instructions and data for implementing the road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement according to any one of claims 1-7.
9. A device for detecting traffic micro-objects based on multi-layer pruning and density-aware enhanced road-end sparse point cloud, characterized in that: It comprises: A processor and a storage device; the processor loads and executes the instructions and data in the storage device to implement the road end sparse point cloud traffic small target detection method based on multi-layer pruning and density perception enhancement according to any one of claims 1-7.