Laser radar point cloud target detection method and device based on graph attention network, and medium

Abstract

The invention discloses a laser radar point cloud target detection method and device based on a graph attention network, and a medium. The method comprises the steps of obtaining a laser radar original data set P; a laser radar original data set P is defined as a point cloud set, a point Pi serves as a vertex of a graph, Pi belongs to P, K nearest neighbor points are found for each point, edges are established between each point and the corresponding K neighbor points, and therefore a graph structure is formed; performing down-sampling on the graph structure to obtain a vertex feature H = {h1, h2,..., hN}; on the basis of feature mapping and a shared attention mechanism, feature aggregation is carried out through a maximum pooling function, so that vertex features are updated; and classifying the updated vertex features through a pre-trained multi-layer perceptron (MLP), and predicting a bounding box of the target at each vertex.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser radar target detection, and in particular relates to a laser radar point cloud target detection method, device, and medium based on a graph attention network. Background Art

[0002] One of the core aspects of autonomous driving technology is environmental perception, and LiDAR, with its high precision and real-time performance, plays a vital role in this field. In autonomous driving, the point cloud data generated by LiDAR can provide rich three-dimensional spatial information for detecting and identifying surrounding objects, providing the necessary input for decision-making systems. To improve the accuracy of target recognition, researchers have used several advanced models for 3D object detection. These mainly include the following categories:

[0003] The first category is the voxel-based three-dimensional laser point cloud target detection algorithm. The voxel-based detection method discretizes the point cloud data into small cubes (i.e., voxels). Each voxel represents a small area in space and contains the point cloud information within the area. This method can effectively structure the point cloud data so that they can be better processed by existing convolutional neural networks. The voxelization process retains the spatial distribution characteristics of the point cloud, making it easier for the network to capture three-dimensional geometric features. Methods for voxelizing point clouds are widely adopted, such as SECOND and PointPillars. Although voxelization is effective, it will lead to a certain loss of spatial information, especially when the voxel size is not chosen properly. In addition, the requirements for computing resources are still high when applied on a large scale.

[0004] The second category is point-based object detection algorithms. Point-based 3D laser point cloud object detection algorithms operate directly on raw point clouds without converting them to voxel or image representations. The core advantage of this type of algorithm lies in its ability to preserve the complete geometric information of point cloud data, which is particularly important for capturing fine details in 3D space. The development of this type of algorithm began in 2017 with the introduction of PointNet, the first deep learning model to directly process point cloud data. It uses a symmetric spatial transformation network to ensure order invariance of input points. Finally, fully connected layers are used to extract global features for point cloud classification and segmentation. Subsequently, PointNet++ introduced local dependencies, extracting local features layer by layer and then gradually integrating them into global features, effectively improving the model's ability to handle complex 3D scenes. In recent years, point-based algorithms have been further developed, such as PointRCNN and 3DSSD. These algorithms have significantly improved detection accuracy and efficiency by improving feature extraction and candidate region generation. Furthermore, new convolutional operations such as KPConv have been proposed, which further enhance the model's feature extraction capabilities by considering the local neighborhood of each point and dynamically aggregating local information. Although point-based algorithms have demonstrated efficient information extraction capabilities when processing point clouds, most existing algorithms have high computational complexity and cannot meet the needs of real-time processing.

[0005] The third category is point-voxel-based methods, which are novel detection frameworks that combine point cloud and voxel technologies. They aim to overcome the limitations of traditional point cloud processing and pure voxel-based methods. These hybrid algorithms achieve high-precision recognition of three-dimensional objects by combining detailed point cloud geometric information with the efficient computation of voxels during the feature extraction phase. With technological advancements, research in this field has gradually differentiated into single-stage and two-stage detection strategies. Single-stage detection frameworks such as PVCNN and SPVNAS optimize feature extraction through point-to-voxel and voxel-to-point conversions. Two-stage detection frameworks, such as PV-RCNN and LiDAR R-CNN, first use voxel-based methods to roughly locate candidate objects and then refine key regions using point cloud operators. Compared to single voxel-based methods, these methods significantly improve detection accuracy while reducing inference time, demonstrating the advancement of point cloud data processing.

[0006] The fourth category is graph-based point cloud object detection. The core of this approach is to model the relationships between points in point cloud data. Each point in the point cloud is treated as a node in a graph, and edges between nodes represent neighborhood relationships or similarities between points. This graph structure effectively captures the spatial distribution characteristics and local geometric properties of point clouds. Point-GNN is a typical example of graph-based object detection. In Point-GNN, a point cloud is represented as a graph, with each point as a node. Nodes are connected by edges, which represent their positional relationships. Specifically, neighboring points within a fixed radius are connected as edges, allowing feature information to be transferred within this local area. By constructing a graph representation of the point cloud and leveraging GNN for feature extraction and object detection, PointGNN demonstrates a novel point cloud processing framework. Its advanced performance on the KITTI dataset demonstrates the great potential of this approach in future 3D object detection research.

[0007] Among these four types of point cloud object detection algorithms, graph-based methods have a unique advantage in processing point cloud data: they can naturally capture the disorder and local structural characteristics of point clouds while maintaining a high degree of flexibility and expressiveness. Although point-based methods can preserve complete three-dimensional spatial information, their high computational cost and sensitivity to noise limit their practical application. While voxel-based methods improve computational efficiency, they may affect detection accuracy due to quantization errors and challenges in sparse point cloud areas. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the present invention provides a lidar point cloud target detection method, device, and medium based on a graph attention network.

[0009] In a first aspect, an embodiment of the present invention provides a method for detecting target in a laser radar point cloud based on a graph attention network, the method comprising:

[0010] Get the original lidar dataset P;

[0011] The original lidar dataset P is defined as a point cloud set, and the point P i As a vertex of the graph, P i ∈P, find the nearest K neighbor points for each point, and establish edges between each point and its corresponding K neighbor points to form a graph structure;

[0012] Downsample the graph structure to obtain vertex features H = {h1,h2,...,h N};

[0013] Based on feature mapping and shared attention mechanism, feature aggregation is performed through the maximum pooling function to update vertex features;

[0014] The updated vertex features are classified by a pre-trained multi-layer perceptron (MLP) to predict the bounding box of the object at each vertex.

[0015] In a second aspect, an embodiment of the present invention provides an electronic device comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned lidar point cloud target detection method based on graph attention network.

[0016] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned lidar point cloud target detection method based on a graph attention network.

[0017] In a fourth aspect, an embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-mentioned lidar point cloud target detection method based on a graph attention network.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The lidar point cloud target detection method based on graph attention network provided by the present invention can effectively extract target features from lidar point cloud data, and perform target detection and classification, providing reliable environmental perception information for autonomous driving vehicles.

[0020] Traditional graph convolutional neural networks assume a fixed neighborhood structure, while the graph attention network (Point-GAT) can handle disordered and sized neighborhoods, making it more suitable for processing point cloud data. By introducing a shared attention mechanism, Point-GAT can dynamically focus on the most relevant parts of the neighborhood for feature learning. It can effectively process spatial relationships in point cloud data and extract richer feature representations, thereby better extracting the structural features of objects. In addition, Point-GAT integrates target classification and localization tasks in an end-to-end learning framework, improving the efficiency and accuracy of the algorithm. Because Point-GAT can better extract the structural features of objects, it is more robust to noise and occlusion, and can maintain high detection accuracy even in complex environments.

[0021] By downsampling the graph structure to obtain vertex features, the graph pyramid structure can handle objects of different scales and extract feature information at different levels. A multi-layer perceptron (MLP) can classify and regress vertex features to predict the object category and bounding box parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A schematic diagram of the process of a LiDAR point cloud target detection method based on a graph attention network provided by an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] It should be noted that, unless there is any conflict, the features in the following embodiments and implementations may be combined with each other.

[0027] like Figure 1 As shown, an embodiment of the present invention provides a lidar point cloud target detection method based on a graph attention network (Point-GAT), the method comprising the following steps:

[0028] Step S1: Obtain the original lidar data set P.

[0029] Specifically, in this example, the KITTI dataset was selected as the raw lidar dataset P. The KITTI dataset is one of the largest and most scenario-rich autonomous driving computer vision evaluation datasets available internationally. It covers diverse driving scenarios, including urban, rural, and highway driving, as well as varying weather conditions, such as sunny, cloudy, and rainy days, as well as varying lighting conditions. Each image in the KITTI dataset is precisely annotated, including 3D bounding boxes and location information for objects such as vehicles and pedestrians. This annotation information includes not only the object category but also its pose and 3D shape, providing a rich reference for 3D object detection and tracking. For 3D object detection, the KITTI dataset provides a more detailed label classification, including car, van, truck, pedestrian, pedestrian (sitting), cyclist, tram, and misc (other).

[0030] Step S2: define the original laser radar data set P as a point cloud set, and point P i As a vertex of the graph, P i ∈P, find the nearest K neighbor points for each point through the K-Nearest Neighbors (KNN) algorithm, and establish edges between each point and its corresponding K neighbor points to form a graph structure.

[0031] Specifically, the original laser radar data set P containing N points is P = {p1, p2, ..., p i ,...,p N} is defined as a point cloud set, where p i ={x i ,s i} is a point, x i ∈R 3 For point cloud as and s i ∈R k is the state value, and the k-length vector represents the attributes of the point. i Indicates the laser reflection intensity.

[0032] Given a point cloud P, we use P as a vertex V to obtain the graph vertices V. We use the K-nearest neighbor algorithm to find the K nearest neighbors for each point, and then establish edges between these K neighbors. We then connect each point with its K corresponding points through an edge, ultimately obtaining the set of edges E corresponding to the graph of the entire point cloud data.

[0033] That is, for each point p i and its K nearest neighbors Create Edge Therefore, the edge set E can be expressed as:

[0034] and and j≠i}

[0035] Here Represents point p i and The edges between For all i from 1 to N, Indicates that p i Each nearest neighbor j of . Through the above steps, we get the graph G = (V, E).

[0036] Step S3: downsample the graph structure to obtain vertex features H = {h1,h2,...,h N}.

[0037] Specifically, in LiDAR point cloud object detection, since point cloud data typically contains a large number of dense points, performing feature learning and object detection directly on the raw point cloud is very time-consuming and computationally intensive. Therefore, downsampling of graph-structured data is necessary.

[0038] In this example, the Farthest Point Sampling (FPS) method is used to downsample the graph structure. The principle of the FPS method is to select a point from the point cloud as a seed point, then select the point farthest from the currently selected point from the remaining points as the next seed point, and repeat this process until a specified number of points are selected. The FPS method effectively preserves the overall shape and structural characteristics of the point cloud, and is therefore widely used in point cloud downsampling. By downsampling the point cloud using the FPS method, point cloud subsets of different scales can be obtained. Each subset contains a different number of points, representing a different level of detail in the point cloud data. Constructing each subset into a corresponding graph forms a graph pyramid. Each graph in the graph pyramid represents the structural information of the point cloud data at different scales and can be used for feature learning at different levels. For example, low-level graphs can be used to learn local features of the point cloud, while high-level graphs can be used to learn global features of the point cloud. By performing feature learning at different levels of the graph pyramid, multi-level features can be effectively extracted from the point cloud data, thereby improving the accuracy and robustness of object detection.

[0039] In step S3, based on feature mapping and shared attention mechanism, feature aggregation is performed through the maximum pooling function to update the vertex features.

[0040] Specifically, step S3 includes the following sub-steps:

[0041] Step S301, vertex feature H = {h1, h2, ..., h N}, where each feature Associated with the corresponding vertex i∈V in the graph, where F is the feature dimension of each vertex. The graph attention network Point-GAT is designed to learn a function It maps the input feature H to a set of new vertex features H′={h1′,h2′,...,h N '},in While maintaining the structural connections between these output features. At the same time, unlike the relatively fixed adjacency relationship in the image domain, the proposed Point-GAT should also be able to handle disordered and variable-sized neighborhoods while maintaining the weight sharing property.

[0042] Step S302: Build a shared attention mechanism It is used to focus on the most relevant parts in the neighborhood for feature learning, so that the convolution kernel of Point-GAT can dynamically adapt to the structure of the object. Specifically, the attention weight of each adjacent vertex is calculated as follows:

[0043]

[0044] in, Represents the attention weight vector of vertex j to vertex i. Δp ij =p i -p j , Δh ij =M g (h j )-M g (h i ),in is the feature mapping function applied to each vertex, i.e., M g is a multilayer perceptron. The first term of α is Δp ij Represents the spatial relationship between adjacent vertices, which helps to expand disordered adjacent points to meaningful surfaces. ij It is used to measure the feature differences between vertex pairs, guiding us to allocate more attention to similar adjacent points.

[0045] The shared attention mechanism α can be implemented with any differentiable architecture. This paper uses a multi-layer perceptron, which can be expressed as follows:

[0046] α(Δp ij ,Δh ij )=M α ([Δp ij ||Δh ij ])

[0047] Among them, || is the connection operation, Mα Represents an applied multilayer perceptron.

[0048] In S303, in order to handle neighborhoods of variable sizes at different vertices and spatial scales, the attention weights are normalized across all neighborhoods of vertex i as follows:

[0049]

[0050] in, is the attention weight of vertex j to vertex i on the kth feature channel.

[0051] S304, iteratively update the vertex features. Each iterative step uses neighborhood information to update the vertex features, thereby extracting richer feature representations. The update formula is as follows:

[0052]

[0053] Where * represents the Hadamard product, which produces the element-wise product of two vectors. is a learnable bias and σ is the activation function.

[0054] S305, using the MaxPool function as the feature aggregation function, aggregates the features in the neighborhood to the central vertex. MaxPool can be expressed as:

[0055]

[0056] In step S4, the updated vertex features are classified by a pre-trained multi-layer perceptron (MLP) to predict the bounding box of the target at each vertex.

[0057] Specifically, the multi-layer perceptron (MLP) is trained using classification loss and localization loss. The classification loss is calculated using the cross-entropy loss function to measure the accuracy of classification predictions, as shown below:

[0058]

[0059] Where, L cls is the classification loss, N is the number of vertices, C is the number of categories, y ic is the indicator variable that vertex i belongs to category c, q ic is the predicted probability that vertex i belongs to category c.

[0060] Among them, the loss loss uses the smooth L1 loss function to calculate the loss of the positioning branch to measure the accuracy of the bounding box prediction. The expression is as follows:

[0061]

[0062] Where, Lloc is the positioning loss, D is the number of bounding box parameters, are the true bounding box parameters of vertex i, are the predicted bounding box parameters of vertex i. ρ(·) is a smooth L1 function that calculates the distance between the predicted value and the true value.

[0063] Specifically, the multi-layer perceptron MLP can be expressed as:

[0064] q i =softmax(MLP(h i ))

[0065] Where q i is the classification probability distribution of vertex i, h i is the feature of vertex i; softmax(·) is the softmax function that converts the feature vector into a probability distribution; MLP is a multi-layer perceptron used to learn the mapping to the probability distribution.

[0066] The expressions for identifying the bounding box parameters are as follows:

[0067] δ i =MLP(h i )

[0068] Where, δ i is the bounding box parameter of vertex i, and MLP is a multi-layer perceptron used to learn the mapping from features to bounding box parameters.

[0069] Furthermore, in this example, the detected bounding boxes are merged using the Non-Maximum Suppression (NMS) algorithm to remove overlapping boxes and retain the box with the highest confidence. The expression is as follows:

[0070]

[0071] Where B represents the set of detected bounding boxes, Th is the overlap threshold, and IoU is the intersection-over-union function used to calculate the degree of overlap between two bounding boxes. i is the retained bounding box.

[0072] It's important to note that the basic principle of the NMS algorithm is to first sort all detected bounding boxes by confidence. Then, starting with the box with the highest confidence, the intersection over union (IoU) between it and the other boxes is calculated. If the IoU between two boxes is greater than a preset threshold, the two boxes are considered overlapping, and the box with the higher confidence is retained, while the box with the lower confidence is removed. This process is repeated until all boxes have been processed.

[0073] Table 1: Comparison of mAP values ​​of different models

[0074] Model Car Pdestrian Cyclist SECOND 72.27 68.70 60.62 LiDAR R-CNN 76.0 71.2 68.60 PV-RCNN 78.00 79.21 71.46 PV-RCNN++ 79.25 81.83 73.46 Ours 80.25 84.63 78.30

[0075] As can be seen from Table 1 above, Table 1 gives the mean average precision (mAP) values ​​of SECOND, LiDAR R-CNN, PV-RCNN, PV-RCNN++ and the model provided in this example when detecting three types of objects: vehicles (Car), pedestrians (Pedestrian) and cyclists (Cyclist). The model provided in this example achieved an mAP of 80.25% for vehicle (Car), 84.63% for pedestrian (Pedestrian), and 78.30% for cyclists (Cyclist). The model provided in this application achieved the highest accuracy in the detection of the three types of objects, especially in pedestrian and vehicle detection, showing significant superiority. The detection accuracy of cyclists has also been greatly improved, approaching the accuracy level of vehicle detection.

[0076] In summary, the lidar point cloud target detection method based on the graph attention network provided by the present invention can effectively extract target features from lidar point cloud data, and perform target detection and classification, providing reliable environmental perception information for autonomous driving vehicles. Specifically, the traditional graph convolutional neural network assumes that the neighborhood structure is fixed, while the graph attention network (Point-GAT) can handle disordered and sized neighborhoods, making it more suitable for processing point cloud data. By introducing a shared attention mechanism, Point-GAT can dynamically focus on the most relevant parts of the neighborhood for feature learning, effectively process the spatial relationships in point cloud data, and extract richer feature representations, thereby better extracting the structural features of the object. In addition, Point-GAT integrates target classification and positioning tasks in an end-to-end learning framework, which improves the efficiency and accuracy of the algorithm. Since Point-GAT can better extract the structural features of the object, it is more robust to noise and occlusion, and can maintain a high detection accuracy even in complex environments.

[0077] like Figure 2 As shown, an embodiment of the present application provides an electronic device, which includes a memory 101 for storing one or more programs and a processor 102. When the one or more programs are executed by the processor 102, any method of the first aspect described above is implemented.

[0078] The system also includes a communication interface 103. The memory 101, processor 102, and communication interface 103 are electrically connected to each other directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory 101 can be used to store software programs and modules, and the processor 102 executes the software programs and modules stored in the memory 101 to perform various functional applications and data processing. The communication interface 103 can be used to communicate signaling or data with other node devices.

[0079] Among them, the memory 101 can be, but is not limited to, a random access memory 101 (Random Access Memory, RAM), a read-only memory 101 (Read Only Memory, ROM), a programmable read-only memory 101 (Programmable Read-Only Memory, PROM), an erasable programmable read-only memory 101 (Erasable Programmable Read-Only Memory, EPROM), an electrically erasable read-only memory 101 (Electric Erasable Programmable Read-Only Memory, EEPROM), etc.

[0080] The processor 102 may be an integrated circuit chip with signal processing capabilities. The processor 102 may be a general-purpose processor 102, including a central processing unit 102 (CPU), a network processor 102

[0081] (Network Processor, NP), etc.; it can also be a digital signal processor 102 (Digital Signal Processing, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0082] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method and system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of the methods and systems, methods, and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment, or a portion of code, and the module, program segment, or a portion of code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0083] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0084] On the other hand, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements a method as described in any one of the first aspects above when executed by the processor 102. If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory 101 (ROM, Read-Only Memory), a random access memory 101 (RAM, Random Access Memory), a magnetic disk or an optical disk.

[0085] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the contents disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only.

[0086] It will be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A laser radar point cloud target detection method based on graph attention network, characterized in that: The method comprises: Get the original laser radar data set P; The original laser radar data set P is defined as a point cloud set. i As a vertex of the graph, P i ∈P, find the nearest K neighbor points for each point, and establish edges between each point and its corresponding K neighbor points to form a graph structure; Downsample the graph structure to obtain vertex features H = {h1,h2,...,h N }; Based on feature mapping and shared attention mechanism, feature aggregation is performed through the maximum pooling function to update vertex features; The updated vertex features are classified by a pre-trained multi-layer perceptron MLP to predict the bounding box of the object described by each vertex.

2. According to claim 1, a laser radar point cloud target detection method based on graph attention network is characterized in that: The laser radar original data set uses the KITTI data set.

3. According to claim 1, a laser radar point cloud target detection method based on graph attention network is characterized in that: The process of building a graph structure includes: The original laser radar data set P containing N points is P = {p1, p2, ..., p i ,...,p N } is defined as a point cloud set, where p i ={x i ,s i } is a point, x i ∈R 3 For the point cloud as and s i ∈R k is the status value; Use the point cloud P as vertex V to obtain the vertex V of the graph; For each point p i and its K nearest neighbors Create Edge Therefore, the edge set E can be expressed as: Here Represents point p i and The edge between For all i from 1 to N, Indicates that p i Each nearest neighbor j of Thus, the graph structure G = (V, E) is constructed.

4. The method for detecting target in laser radar point cloud based on graph attention network according to claim 1, characterized in that: The process of downsampling the graph structure and obtaining vertex features includes: The farthest point sampling method is used to downsample the graph structure to obtain vertex features.

5. The method for detecting target in laser radar point cloud based on graph attention network according to claim 1, characterized in that: Based on feature mapping and shared attention mechanism, the process of updating vertex features by performing feature aggregation through the maximum pooling function includes: Construct a graph attention network, the expression is: F is the characteristic dimension of each vertex; The graph attention network transforms the vertex features H = {h1,h2,...,h N }, mapped to a new set of vertex features H′={h1′,h2′,...,h N '},in Building a shared attention mechanism The convolution kernel of the graph attention network can dynamically adapt to the structure of the object, that is, in, represents the attention weight vector of vertex j to vertex i, Δp ij =p i -p j , Δh ij =M g (h j )-M g (h i ),in is the feature mapping function applied to each vertex; Δp ij Represents the spatial relationship between adjacent vertices; Δh ij Used to measure feature differences between vertex pairs; The attention weights are normalized across all neighbors of vertex i as follows: in, is the attention weight of vertex j to vertex i on the kth feature channel; Iteratively update the vertex features, the expression is as follows: Among them, * represents Hadamard, is a learnable bias, σ is the activation function; Use the maximum pooling function as the feature aggregation function to aggregate the features in the neighborhood to the central vertex. The expression is as follows:

6. The method for detecting target in laser radar point cloud based on graph attention network according to claim 1, characterized in that: The training process of the multi-layer perceptron MLP includes: The multi-layer perceptron MLP is trained with classification loss and localization loss; Among them, the expression of the classification loss is as follows: Where, L cls is the classification loss, N is the number of vertices, C is the number of categories, and y ic is the indicator variable that vertex i belongs to category c, q ic is the predicted probability that vertex i belongs to category c; The expression of loss determination is as follows: Where, L loc is the positioning loss, D is the number of bounding box parameters, are the true bounding box parameters of vertex i, are the predicted bounding box parameters of vertex i. ρ(·) is a smooth L1 function that calculates the distance between the predicted value and the true value.

7. The method for detecting target in laser radar point cloud based on graph attention network according to claim 1, characterized in that: The multi-layer perceptron MLP classifies the updated vertex features and predicts the bounding box of the target described by each vertex. The process includes: The expression of the classification probability distribution of vertex i is as follows: q i =softmax(MLP(h i )) In the formula, q i is the classification probability distribution of vertex i, h i is the feature of vertex i; softmax(·) is the softmax function, and MLP is a multi-layer perceptron; The expressions for identifying the bounding box parameters are as follows: δ i =MLP(h i ) In the formula, δ i is the bounding box parameter of vertex i, and MLP is a multi-layer perceptron used to learn the mapping from features to bounding box parameters; The detected bounding boxes are merged via the non-maximum suppression algorithm.

8. An electronic device, comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the lidar point cloud target detection method based on the graph attention network as described in any one of claims 1-7 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the laser radar point cloud target detection method based on a graph attention network as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the method for detecting target by a laser radar point cloud based on a graph attention network described in any one of claims 1 to 7 is implemented.

Images