Point cloud detection method and system based on local graph attention and multi-scale fusion

Abstract

The application discloses a point cloud detection method and system based on local graph attention and multi-scale fusion, and belongs to the technical field of automatic driving environment perception. The method first performs dynamic self-adaptive filtering on original point clouds, and removes noise while retaining sparse small targets; then, a KNN local topological graph is constructed in a body column, point features are aggregated through a graph attention network, and small target geometric expression is enhanced; subsequently, multi-scale BEV features are extracted, and bidirectional cross-scale fusion is realized through an adaptive weighted BiFPN; finally, three-dimensional detection results are obtained through a detection head decoding. The application effectively solves the problems of weak small target feature expression and high missing detection rate in sparse point clouds, significantly improves the detection accuracy and environmental robustness of small targets such as pedestrians and cyclists, meets the real-time requirements of vehicle-mounted systems, and can be widely applied to laser radar perception systems of intelligent networked vehicles.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving environmental perception technology, specifically relating to a three-dimensional target detection method and system for LiDAR point cloud, which is particularly suitable for high-precision and high-robust detection of small-sized targets with physical dimensions less than 1.2 meters and sparse point cloud quantity, such as pedestrians, cyclists, and traffic cones in autonomous driving scenarios. Background Technology

[0002] With the rapid development of autonomous driving technology, 3D object detection, as a core component of autonomous driving environmental perception systems, directly determines the decision-making safety and driving reliability of autonomous vehicles through its detection accuracy and robustness. LiDAR, capable of providing precise 3D spatial geometric information and unaffected by changes in lighting, has become a core sensor in the perception module of autonomous driving systems.

[0003] In real-world autonomous driving scenarios, small targets such as pedestrians, cyclists, and traffic cones are key targets for environmental perception, but also present significant challenges in detection. These targets typically exhibit extremely sparse point cloud data in LiDAR data, with fewer than 50 effective points. Existing 3D target detection methods face significant technical bottlenecks when processing such sparse small targets.

[0004] First, the ability to represent small targets is insufficient. Traditional point cloud feature extraction methods based on voxels / volume pillars usually aggregate features through max pooling or multilayer perceptrons, ignoring the spatial proximity and geometric topology between points. This results in the key geometric information of small targets being buried during the feature aggregation process, failing to form an effective feature response.

[0005] Second, the quality of point clouds degrades severely in complex environments. Under adverse weather conditions such as rain, fog, and low light, lidar point clouds become severely sparsified and accompanied by a large number of noise points. Traditional global threshold filtering methods are prone to misclassifying effective points of sparse small targets as outliers and filtering them out, further exacerbating the difficulty of detecting small targets.

[0006] Third, multi-scale feature fusion is insufficient. After multiple downsampling operations, although high-level features possess strong semantic information, their spatial resolution is significantly reduced, making it difficult to accurately locate small targets. Furthermore, traditional Feature Pyramid Networks (FPNs) often employ a fixed-weight unidirectional fusion method, which cannot adaptively balance the contribution of high-level semantics and low-level details, resulting in severe loss of detailed information about small targets.

[0007] Although existing research attempts to introduce attention mechanisms or improve feature fusion structures to enhance the detection performance of small targets, there is still a lack of systematic solutions for point cloud sparsity, noise robustness, and multi-scale feature fusion. These solutions cannot effectively address the issues of missed and false detections of small targets in sparse point clouds while maintaining real-time performance on vehicles. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides a point cloud small target detection method and system based on local graph attention and multi-scale fusion. This method solves the problems of weak feature representation ability of sparse small targets, difficulty in balancing point cloud noise filtering and target preservation in complex environments, and low detection accuracy and high false negative rate of small targets due to insufficient multi-scale feature fusion in the existing technologies. While meeting the real-time requirements of vehicle in-vehicle systems, this invention significantly improves the detection accuracy and environmental robustness of small targets in autonomous driving scenarios.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A point cloud small target detection method based on local graph attention and multi-scale fusion is applied to the 3D target detection scenario of LiDAR point clouds in autonomous driving environmental perception, including the following steps:

[0011] Step S1: Point cloud data preprocessing and dynamic adaptive filtering. The original LiDAR point cloud is normalized in coordinates and the region of interest is clipped. Then, the outlier filtering threshold is adaptively adjusted by local grid density. While suppressing noise, sparse small target point cloud is retained to obtain the denoised effective point cloud.

[0012] Step S2: Volume column partitioning and local graph attention feature enhancement. The denoised effective point cloud is discretized into a uniform volume column mesh on the bird's-eye view plane. A local topological graph is constructed in each non-empty volume column with point cloud as nodes and K-nearest neighbor relationships as edges. The features of neighboring nodes are aggregated through a graph attention network. The output features of all nodes in the volume column are globally pooled to generate an enhanced volume column-level feature vector.

[0013] Step S3: BEV feature extraction and multi-scale adaptive fusion. All volumetric feature vectors are mapped to two-dimensional bird's-eye view BEV feature maps according to their spatial positions. Multi-scale feature maps are extracted through the backbone network. Then, the adaptive weighted bidirectional feature pyramid network BiFPN is used to construct a bidirectional fusion path from top to bottom and bottom to top. Adaptive weighted fusion of cross-scale features is achieved through learnable weights, and the enhanced multi-scale BEV features are output.

[0014] Step S4: Decode the 3D detection results. Input the fused multi-scale BEV features into the detection head, decode to obtain the target category probability and 3D bounding box parameters, perform non-maximum suppression post-processing, and output the final 3D target detection results.

[0015] As a further technical solution of the present invention, step S1 specifically includes:

[0016] S1.1: Acquire raw LiDAR point cloud data Each point contains three-dimensional coordinates (x, y, z) and reflection intensity information. The point cloud is then normalized to a unified coordinate system.

[0017] S1.2: Define the 3D region of interest (ROI), crop and retain the point cloud within the ROI to obtain the ROI point cloud. The cutting formula is as follows:

[0018]

[0019] in, , , , These are the boundary values ​​of the ROI on the three coordinate axes;

[0020] S1.3: Divide the ROI point cloud into multiple equal-sized local grids in the XOY plane. For each point in each grid, calculate the average distance to its K nearest neighbors. Where K=20;

[0021] S1.4: Calculate the mean of the average distances between all points within the grid. and standard deviation Generate dynamic filter threshold ,in This is an adjustable sensitivity parameter, with a value range of [value range missing]. ;

[0022] S1.5: Average distance within the grid Points exceeding the dynamic filtering threshold are identified as outliers and filtered out, resulting in a denoised effective point cloud. The formula is:

[0023] .

[0024] As a further technical solution of the present invention, in step S1.3, the size of the local mesh is 10m×10m; in step S1.2, the range of ROI in the XOY plane is [-80m, 80m]×[-40m, 40m].

[0025] As a further technical solution of the present invention, step S2 includes:

[0026] S2.1: Discretize the denoised effective point cloud into a uniform cylindrical mesh in the XOY plane. The planar size of a single cylindrical mesh is 0.16m × 0.16m. Using the point cloud in each non-empty cylindrical mesh as a node, search for the K nearest neighbors of each node using the K nearest neighbor algorithm to construct an undirected local topological graph G=(V,E), where V is the set of nodes, E is the set of edges formed by the nearest neighbor relationship, and K=8.

[0027] S2.2: Perform a linear transformation on the node features of the local topology graph, and calculate the unnormalized attention score between a node and its neighboring nodes. The formula is as follows:

[0028]

[0029] in, , Let be the initial features of node i and node j, W be the shared weight matrix, and a be the attention weight vector. This is a vector concatenation operation; LeakyReLU is a non-linear activation function.

[0030] S2.3: Normalize the attention score using the softmax function to obtain the final attention weights, as shown in the following formula:

[0031]

[0032] in, Let i be the set of nearest neighbors of node i;

[0033] S2.4: The linear transformation features of neighboring nodes are weighted and summed using attention weights, and then nonlinearly activated to obtain the output features of the nodes. The formula is as follows:

[0034]

[0035] in, It is a non-linear activation function;

[0036] S2.5: Perform global max pooling on the final layer output features of all nodes within the body column to generate the enhanced body column-level feature vector corresponding to the body column.

[0037] As a further technical solution of the present invention, a multi-head attention mechanism is adopted in step S2 to concatenate or average the output features of multiple independent attention heads as the final output features of the node; the graph attention network has a stacking layer of 2 layers.

[0038] As a further technical solution of the present invention, step S3 specifically includes:

[0039] S3.1: Map all volumetric feature vectors to the bird's-eye view plane according to their spatial coordinates to generate a two-dimensional BEV feature map. Where H and W are the height and width of the feature map, and C is the number of feature channels.

[0040] S3.2: Input the BEV feature map into the convolutional backbone network to extract multi-scale feature maps downsampled by 4 times, 8 times, and 16 times. The convolutional backbone network adopts a lightweight ResNet-18 network.

[0041] S3.3: Input the multi-scale feature maps into the adaptive weighted bidirectional feature pyramid network BiFPN to construct a top-down semantic propagation path and a bottom-up detail propagation path. After upsampling or downsampling to align features with mismatched resolutions, adaptively weighted fusion of the multi-path input features is performed using learnable non-negative weights. The output formula of the fusion node is:

[0042]

[0043] in, These are the resolution-aligned input features. For the learnable non-negative weights corresponding to the input features, The minimum constant to ensure numerical stability is set to a value of . ;

[0044] S3.4: After multi-level bidirectional fusion, the enhanced multi-scale BEV features are output.

[0045] As a further technical solution of the present invention, the detection head in step S4 adopts a center-based detection paradigm based on key points, specifically including the following sub-steps:

[0046] S4.1: Input the fused multi-scale BEV features into the detection head, and predict the target center point heatmap through convolutional branches. In the heat map, each peak position corresponds to the center point of the BEV plane of a potential target, and the peak height represents the confidence level of the target's existence.

[0047] S4.2: For each center point identified in the heatmap, regress the target's 3D bounding box attributes through multiple parallel convolutional branches; including the local offset of the center point. Used for sub-pixel level precise positioning; target size Target orientation angle Using sine-cosine encoding , To avoid the problem of periodic discontinuity in angles;

[0048] S4.3: Perform non-maximum suppression on the initial 3D bounding box output by the detection head and set the cross-union ratio threshold. If the IoU is set to 0.25, redundant detection boxes with IoU exceeding the threshold in the same category are removed, and only the detection boxes with the highest confidence are retained, outputting the final 3D object detection result.

[0049] A point cloud detection system based on local graph attention and multi-scale fusion includes:

[0050] The point cloud preprocessing module is used to normalize the coordinates and crop the region of interest of the original LiDAR point cloud, and then adaptively adjust the outlier filtering threshold through local grid density to output the denoised effective point cloud.

[0051] The local graph attention feature enhancement module is used to divide the denoised effective point cloud into volumetric grids. Within a single volumetric grid, a local topological graph is constructed with point clouds as nodes and K-nearest neighbor relationships as edges. The features of neighboring nodes are aggregated through a graph attention network to generate enhanced volumetric feature vectors.

[0052] The multi-scale feature fusion module is used to map the volumetric feature vector to a two-dimensional bird's-eye view BEV feature map. After extracting multi-scale features, it completes cross-scale feature bidirectional fusion through the adaptive weighted bidirectional feature pyramid network BiFPN and outputs enhanced multi-scale BEV features.

[0053] The 3D detection and decoding module is used to input the fused multi-scale BEV features into the detection head, decode them to obtain the target's class probability and 3D bounding box parameters, and output the final 3D target detection result after non-maximum suppression post-processing.

[0054] As a further technical solution of the present invention, the point cloud preprocessing module has a built-in dynamic outlier filtering unit, which is used to divide the ROI point cloud into multiple local grids, calculate the dynamic filtering threshold based on the local density of the point cloud in a single grid, and achieve an adaptive balance between noise filtering and sparse small target retention.

[0055] As a further technical solution of the present invention, the local graph attention feature enhancement module has a built-in multi-head graph attention network unit, which is used to aggregate the neighborhood geometric features of the point cloud within the column through the multi-head attention mechanism to enhance the feature expression ability of sparse small targets.

[0056] The multi-scale feature fusion module incorporates an adaptive weighted BiFPN unit, which is used to achieve bidirectional adaptive fusion of high-level semantic features and low-level detail features through learnable weights.

[0057] This technology proposes a point cloud detection method and system based on local graph attention and multi-scale fusion, which has the following advantages and beneficial effects:

[0058] This invention effectively models the spatial topological relationships within a small target point cloud through a local graph attention mechanism within a column, assigning differentiated attention weights to different neighboring points, significantly enhancing the geometric feature representation ability of sparse small targets. Compared with traditional pooling aggregation methods, the feature response intensity of small targets such as pedestrians and cyclists is increased by more than 40%, effectively reducing the false negative rate.

[0059] This invention employs a dynamic adaptive outlier filtering strategy, which dynamically adjusts the filtering threshold based on the local point cloud density. It performs strict noise reduction in dense point cloud regions and loose filtering in sparse point cloud regions. While effectively suppressing noise, it avoids the accidental deletion of valid points of sparse small targets. Compared with traditional global threshold filtering, the point cloud retention rate of small targets under complex weather conditions is improved by more than 35%, significantly enhancing the environmental robustness of the algorithm.

[0060] This invention constructs a bidirectional cross-scale fusion path through an adaptive weighted BiFPN network. By automatically balancing the contribution of features at different scales through learnable weights, it achieves full interaction between high-level semantic information and low-level detailed information, alleviating the problem of small target detail loss caused by multiple downsampling, and improving the localization accuracy of small targets by more than 25%.

[0061] This invention adopts a modular design, is compatible with mainstream BEV detection frameworks, and its overall algorithm inference speed meets the real-time requirements of vehicles. It does not require modification of existing LiDAR hardware and can be directly deployed in the domain controller of autonomous vehicles, thus possessing strong engineering application value. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the point cloud detection method based on local graph attention and multi-scale fusion of the present invention.

[0063] Figure 2 This is a schematic diagram of the architecture of the point cloud detection system based on local graph attention and multi-scale fusion of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the embodiments. It should be noted that the following content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should all be considered to fall within the protection scope of the present invention.

[0065] Please see Figure 1 , Figure 2 As shown, this invention is a method for detecting small targets in point clouds for autonomous driving that integrates local graph attention enhancement, anti-interference preprocessing, and adaptive multi-scale BEV fusion, including the following steps:

[0066] S1: Point cloud data preprocessing and adaptive filtering: This step normalizes and denoises the original point cloud, laying the foundation for subsequent feature extraction. The aim of this step is to normalize the original LiDAR point cloud and perform an adaptive noise filtering process to improve data quality while preserving sparse, small targets.

[0067] First, acquire the raw LiDAR point cloud frames. , where each point The point cloud contains three-dimensional coordinates and reflection intensity. Coordinate normalization is performed on the point cloud, transforming it to a right-handed coordinate system with the vehicle's center of mass as the origin. Then, a fixed cuboid region of interest is defined. Define its scope in The axis is rice, The axis is rice, The axis is set according to the sensor height. Point cloud within the ROI is obtained through spatial masking operations:

[0068]

[0069] Crucially, to suppress noise and preserve sparse targets, dynamic adaptive outlier removal is employed. Divide the XOY plane into sections with side lengths of... A square grid of meters. For the grid... any point inside Calculate the Euclidean distance to its K nearest neighbor and average it to obtain the local average distance of the point:

[0070]

[0071] in Point The set of K nearest neighbor indices, K=20. Next, the grid is calculated. All points inside mean and standard deviation :

[0072]

[0073]

[0074] Based on this, a dynamic threshold is generated. ,in For adjustable sensitivity parameters, the preferred value range is [value range missing]. Finally, filter out all those that satisfy the condition. The points are used to obtain the denoised point cloud:

[0075] This strategy enables the algorithm to... Strict filtering of small, dense regions, in Large sparse regions are treated leniently, thus intelligently balancing noise reduction and target preservation.

[0076] S2: Volumetric column segmentation and local graph attention feature enhancement, which enhances features by modeling the internal structural relationships of the point cloud. This step converts the point cloud into a structured bird's-eye view representation and uses graph neural networks to enhance the geometric features of small targets at the microscale.

[0077] First, point cloud Discretize the mesh into a uniform cylindrical grid on the XOY plane, with each cylinder having a size of [missing information]. Meters. For each non-empty column, its contained set of points is... The coordinates and intensity of each point are used as the initial node features. Inside this column, for each node The K nearest neighbor algorithm is used to search for the K nearest nodes in the spatial order, where K=8, to construct an undirected local topological graph. , among which the side If and only if yes One of K's closest neighbors.

[0078] Subsequently, a multi-layer graph attention network is applied to this local graph to aggregate neighborhood information. For the first... The layer and node feature update process is as follows:

[0079] First, perform linear projection:

[0080]

[0081] in This is a learnable weight matrix. Then, the nodes are computed. with his neighbors Attention coefficient between:

[0082]

[0083] in For learnable attention vectors, This represents vector concatenation. Next, the attention coefficients are normalized using the softmax function to obtain the attention weights:

[0084]

[0085] Finally, the projected features of neighboring nodes are weighted and summed using attention weights, and a new feature of the node is obtained through a non-linear activation function:

[0086]

[0087] in For activation functions. Through stacking... In layers, such as L=2, node features incorporate geometric information from multi-hop neighborhoods. For volumetric columns... Final layer features of all nodes Perform global max pooling to obtain the encoded feature vector of the column. .

[0088] S3: BEV Feature Extraction and Multi-Scale Fusion. This step converts point features into structured feature maps and performs efficient multi-scale information fusion. The volumetric features are organized into a bird's-eye view feature map, and contextual information from different scales is fused through an efficient pyramid network.

[0089] All column features According to its spatial location Placement generates a dense two-dimensional bird's-eye view feature map. The feature map is then fed into a lightweight convolutional backbone network, such as ResNet-18, to extract a multi-scale feature map set. , where the superscript indicates the downsampling factor relative to the input BEV map.

[0090] Crucially, an adaptively weighted bidirectional feature pyramid network is employed for feature fusion. This network constructs bidirectional paths, both top-down and bottom-up. For any input feature in the network that needs to be fused from multiple sources... The node, its output Calculated using learnable, normalized weights:

[0091]

[0092] in , These are the scalar weight parameters that the network learns for each input feature. ReLU ensures that the weights are non-negative. It is a very small constant used for numerical stability.

[0093] Specifically, in a BiFPN fusion node with bidirectional connectivity, two features are typically received: one is an upsampled result from higher-level features. One path is the downsampling result from low-level features. The fusion operation of this node can be specifically represented as follows:

[0094]

[0095] in, , The network automatically learns importance weights for these two features respectively.

[0096] This mechanism allows the network to dynamically adjust the contribution of each feature stream based on the input content. After multi-level bidirectional fusion, it outputs a set of enhanced feature maps with rich multi-scale contextual information. , used for final detection.

[0097] S4: Input the fused multi-scale BEV feature map into the detection head and decode it to obtain the final 3D object detection result. This invention is compatible with multiple mainstream detection head paradigms. Its core lies in using the enhanced features from the aforementioned steps for prediction, preferably adopting the center-based paradigm based on key points, as it is more suitable for sparse, irregularly arranged point cloud object detection. This step decodes the fused high-level features into specific 3D bounding boxes, and obtains the final detection result through post-processing.

[0098] This invention is compatible with multiple detection head paradigms. A detailed explanation will be provided using the high-precision Center-based paradigm as an example. For each scale of the fused feature map... The detection head performs two tasks in parallel:

[0099] Center point heatmap prediction: Predicting a heatmap using a single convolutional layer. ,in This represents the number of categories. Each location on the heatmap... value This indicates that a category exists at this location. The confidence level of the target center point. Focal Loss is used during training to address the imbalance between positive and negative samples.

[0100] For each potential center point predicted in the heatmap, its complete 3D bounding box properties are regressed. This is achieved through multiple convolutional branches, with the regression objective including: center point offset. Used for sub-pixel level refinement of discretized BEV mesh locations, 3D dimensions To avoid regression ambiguity caused by the periodicity of angles, the orientation angle is usually regressed. ,in This is the yaw angle.

[0101] In reasoning, start with the heatmap. Extract all peak points (i.e., local maxima points) as the predicted center point location. Then, based on this position index, the corresponding attribute is retrieved from the regression branch: offset. ,size Orientation Finally, the three-dimensional center coordinates of the predicted bounding box were calculated as follows: ,in This represents the resolution in meters per pixel corresponding to the scale of the feature map. It is usually obtained from a separate Z-coordinate regression branch. The orientation angle is obtained through... recover.

[0102] Finally, the initial set of detection boxes collected across all scales and categories. Non-maximum suppression is applied. First, the boxes are sorted in descending order of category confidence. For the current category, the box with the highest confidence is selected. Calculate its comparison with all other boxes of the same category. 3D intersection-union ratio Remove all boxes that meet the criteria. frame ,in The preset threshold is 0.25. Repeat this process until all bounding boxes have been processed, and output the final list of detected 3D objects. .

[0103] The above is an exemplary description of the invention. Obviously, the specific implementation of the invention is not limited to the above-described manner. Any non-substantial improvement made using the inventive concept and technical solution of the invention, or the direct application of the inventive concept and technical solution to other situations without modification, is within the protection scope of the invention.

Claims

1. A point cloud small target detection method based on local graph attention and multi-scale fusion, applied to the 3D target detection scenario of LiDAR point clouds in autonomous driving environmental perception, characterized in that... Includes the following steps: Step S1: Point cloud data preprocessing and dynamic adaptive filtering. The original LiDAR point cloud is normalized in coordinates and the region of interest is clipped. Then, the outlier filtering threshold is adaptively adjusted by local grid density. While suppressing noise, sparse small target point cloud is retained to obtain the denoised effective point cloud. Step S2: Volume column partitioning and local graph attention feature enhancement. The denoised effective point cloud is discretized into a uniform volume column mesh on the bird's-eye view plane. A local topological graph is constructed in each non-empty volume column with point cloud as nodes and K-nearest neighbor relationships as edges. The features of neighboring nodes are aggregated through a graph attention network. The output features of all nodes in the volume column are globally pooled to generate an enhanced volume column-level feature vector. Step S3: BEV feature extraction and multi-scale adaptive fusion. All volumetric feature vectors are mapped to two-dimensional bird's-eye view BEV feature maps according to their spatial positions. Multi-scale feature maps are extracted through the backbone network. Then, the adaptive weighted bidirectional feature pyramid network BiFPN is used to construct a bidirectional fusion path from top to bottom and bottom to top. Adaptive weighted fusion of cross-scale features is achieved through learnable weights, and the enhanced multi-scale BEV features are output. Step S4: Decode the 3D detection results. Input the fused multi-scale BEV features into the detection head, decode to obtain the target category probability and 3D bounding box parameters, perform non-maximum suppression post-processing, and output the final 3D target detection results.

2. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Acquire raw LiDAR point cloud data Each point contains three-dimensional coordinates (x, y, z) and reflection intensity information. The point cloud is then normalized to a unified coordinate system. S1.2: Define the 3D region of interest (ROI), crop and retain the point cloud within the ROI to obtain the ROI point cloud. The cutting formula is as follows: in, , , , These are the boundary values ​​of the ROI on the three coordinate axes; S1.3: Divide the ROI point cloud into multiple equal-sized local grids in the XOY plane. For each point in each grid, calculate the average distance to its K nearest neighbors. Where K=20; S1.4: Calculate the mean of the average distances between all points within the grid. and standard deviation Generate dynamic filter threshold ,in This is an adjustable sensitivity parameter, with a value range of [value range missing]. ; S1.5: Average distance within the grid Points exceeding the dynamic filtering threshold are identified as outliers and filtered out, resulting in a denoised effective point cloud. The formula is: 。 3. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 2, characterized in that, In step S1.3, the size of the local mesh is 10m×10m; in step S1.2, the range of ROI in the XOY plane is [-80m, 80m]×[-40m, 40m].

4. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 1, characterized in that, Step S2 includes: S2.1: Discretize the denoised effective point cloud into a uniform cylindrical mesh in the XOY plane. The planar size of a single cylindrical mesh is 0.16m × 0.16m. Using the point cloud in each non-empty cylindrical mesh as a node, search for the K nearest neighbors of each node using the K nearest neighbor algorithm to construct an undirected local topological graph G=(V,E), where V is the set of nodes, E is the set of edges formed by the nearest neighbor relationship, and K=8. S2.2: Perform a linear transformation on the node features of the local topology graph, and calculate the unnormalized attention score between a node and its neighboring nodes. The formula is as follows: in, , Let be the initial features of node i and node j, W be the shared weight matrix, and a be the attention weight vector. This is a vector concatenation operation; LeakyReLU is a non-linear activation function. S2.3: Normalize the attention score using the softmax function to obtain the final attention weights, as shown in the following formula: in, Let i be the set of nearest neighbors of node i; S2.4: The linear transformation features of neighboring nodes are weighted and summed using attention weights, and then nonlinearly activated to obtain the output features of the nodes. The formula is as follows: in, It is a non-linear activation function; S2.5: Perform global max pooling on the final layer output features of all nodes within the body column to generate the enhanced body column-level feature vector corresponding to the body column.

5. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 4, characterized in that, In step S2, a multi-head attention mechanism is used to concatenate or average the output features of multiple independent attention heads as the final output features of the node. The graph attention network has two stacked layers.

6. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 1, characterized in that, Step S3 specifically includes: S3.1: Map all volumetric feature vectors to the bird's-eye view plane according to their spatial coordinates to generate a two-dimensional BEV feature map. Where H and W are the height and width of the feature map, and C is the number of feature channels. S3.2: Input the BEV feature map into the convolutional backbone network to extract multi-scale feature maps downsampled by 4 times, 8 times, and 16 times. The convolutional backbone network adopts a lightweight ResNet-18 network. S3.3: Input the multi-scale feature maps into the adaptive weighted bidirectional feature pyramid network BiFPN to construct a top-down semantic propagation path and a bottom-up detail propagation path. After upsampling or downsampling to align features with mismatched resolutions, adaptively weighted fusion of the multi-path input features is performed using learnable non-negative weights. The output formula of the fusion node is: in, These are the resolution-aligned input features. For the learnable non-negative weights corresponding to the input features, The minimum constant to ensure numerical stability is set to a value of . ; S3.4: After multi-level bidirectional fusion, the enhanced multi-scale BEV features are output.

7. The point cloud small target detection method based on local graph attention and multi-scale fusion according to claim 1, characterized in that, The detection head in step S4 adopts a center-based detection paradigm based on key points, which specifically includes the following sub-steps: S4.1: Input the fused multi-scale BEV features into the detection head, and predict the target center point heatmap through convolutional branches. In the heat map, each peak position corresponds to the center point of the BEV plane of a potential target, and the peak height represents the confidence level of the target's existence. S4.2: For each center point identified in the heatmap, regress the target's 3D bounding box attributes through multiple parallel convolutional branches; Including local offset of the center point Used for sub-pixel level precise positioning; target size Target orientation angle Using sine-cosine encoding , To avoid the problem of periodic discontinuity in angles; S4.3: Perform non-maximum suppression on the initial 3D bounding box output by the detection head and set the cross-union ratio threshold. If the IoU is set to 0.25, redundant detection boxes with IoU exceeding the threshold in the same category are removed, and only the detection boxes with the highest confidence are retained, outputting the final 3D object detection result.

8. A point cloud detection system based on local graph attention and multi-scale fusion, used to perform the method of claim 1, characterized in that, include: The point cloud preprocessing module is used to normalize the coordinates and crop the region of interest of the original LiDAR point cloud, and then adaptively adjust the outlier filtering threshold through local grid density to output the denoised effective point cloud. The local graph attention feature enhancement module is used to divide the denoised effective point cloud into volumetric grids. Within a single volumetric grid, a local topological graph is constructed with point clouds as nodes and K-nearest neighbor relationships as edges. The features of neighboring nodes are aggregated through a graph attention network to generate enhanced volumetric feature vectors. The multi-scale feature fusion module is used to map the volumetric feature vector to a two-dimensional bird's-eye view BEV feature map. After extracting multi-scale features, it completes cross-scale feature bidirectional fusion through the adaptive weighted bidirectional feature pyramid network BiFPN and outputs enhanced multi-scale BEV features. The 3D detection and decoding module is used to input the fused multi-scale BEV features into the detection head, decode them to obtain the target's class probability and 3D bounding box parameters, and output the final 3D target detection result after non-maximum suppression post-processing.

9. The point cloud detection system based on local graph attention and multi-scale fusion according to claim 8, characterized in that, The point cloud preprocessing module has a built-in dynamic outlier filtering unit, which is used to divide the ROI point cloud into multiple local grids and calculate the dynamic filtering threshold based on the local density of the point cloud in a single grid, so as to achieve an adaptive balance between noise filtering and sparse small target retention.

10. The point cloud detection system based on local graph attention and multi-scale fusion according to claim 9, characterized in that, The local graph attention feature enhancement module has a built-in multi-head graph attention network unit, which is used to aggregate the neighborhood geometric features of the point cloud within the column through the multi-head attention mechanism to enhance the feature representation ability of sparse small targets. The multi-scale feature fusion module incorporates an adaptive weighted BiFPN unit, which is used to achieve bidirectional adaptive fusion of high-level semantic features and low-level detail features through learnable weights.

Images