CN-121995340-A - Lei Dadian cloud target detection method and device, electronic equipment and storage medium

CN121995340ACN 121995340 ACN121995340 ACN 121995340ACN-121995340-A

Abstract

The invention belongs to the technical field of radar wave detection, and provides a radar point cloud target detection method, a device, electronic equipment and a storage medium, wherein the method comprises the steps of acquiring radar point cloud data of a target vehicle, and processing the radar point cloud data of the target vehicle by adopting a target detection model to obtain a target recognition result of the target vehicle; the target detection model is obtained through the MPNN network inserted into NLNN and the combined loss function training, and the target classification accuracy in the complex scene perception is improved by constructing the point cloud space relation topological graph and fusing the local space characteristics and the global context information to realize the effective characterization of the target characteristics.

Inventors

YUAN YUHENG
HU SHAOKANG
PENG YIRAN
WANG YIMING
LIU YINGNAN
WU ZIMEI
LIU JIANYU

Assignees

武汉工程大学

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (10)

1. The radar point cloud target detection method is characterized by comprising the following steps of: Acquiring radar point cloud data of a target vehicle, and processing the radar point cloud data of the target vehicle by adopting a target detection model to obtain a target recognition result of the target vehicle; The training steps of the target detection model are as follows: acquiring a data set, wherein the data set comprises radar point cloud data; mapping radar point cloud data serving as nodes into a graph structure, and dynamically constructing an adjacency relationship of the nodes by adopting a K-nearest neighbor algorithm; determining edge characteristics of the graph structure according to the adjacent relation of the nodes, and performing matrix conversion on the nodes, the adjacent relation of the nodes and the edge characteristics to obtain a graph structure matrix; carrying out feature update on nodes and adjacent nodes of the graph structure matrix through graph convolution operation of a message passing neural network to obtain a feature space; and carrying out semantic segmentation on the feature space, and training by adopting a combined loss function to obtain the target detection model, wherein the combined loss function comprises a multi-classification loss function, an object detection loss function and an L2 regularized loss function.
2. The radar point cloud target detection method of claim 1, wherein the acquiring the data set further comprises: Preprocessing Lei Dadian cloud data, namely integrating Lei Dadian cloud data in a preset time period, and cutting the front and rear ranges of the integrated vehicle, wherein cutting comprises point-by-point annotation and track identification according to each point cloud, and further determining the front and rear ranges of the vehicle.
3. The method for detecting the radar point cloud target according to claim 1, wherein the mapping the radar point cloud data as the nodes into the graph structure and dynamically constructing the adjacency relationship of the nodes by adopting a K-nearest neighbor algorithm comprises: Under the condition that the number of the radar point cloud data is kept consistent with that of the nodes, discarding the spatial position characteristics of the point cloud data, and reserving a speed vector, a radar scattering cross section, a time stamp and connectivity for mapping, wherein the connectivity is used for representing the associated edge number of the nodes; and determining the adjacency relation of the nodes by adopting a K-nearest neighbor algorithm according to the Euclidean distance between the nodes and the computational complexity.
4. The radar point cloud target detection method according to claim 3, wherein determining edge features of the graph structure according to the adjacency relation of the nodes, performing matrix conversion on the adjacency relation of the nodes and the edge features to obtain a graph structure matrix, and comprising: According to the adjacent relation of the nodes, calculating the relative positions of the adjacent nodes, and taking the relative positions as the edge characteristics; and performing matrix conversion on the nodes, the adjacent relations of the nodes and the edge characteristics to obtain a graph structure matrix, wherein the graph structure matrix comprises a node characteristic matrix, an edge characteristic matrix and an adjacent matrix.
5. The radar point cloud target detection method according to claim 1, wherein the performing feature update on the nodes and their neighboring nodes through the graph convolution operation of the message passing neural network on the graph structure matrix to obtain a feature space includes: the graph convolution operation of the messaging neural network is: ; Wherein, the Is that The node characteristics of the layer(s), Layer node features By front-level node features Updating with the result of neighborhood aggregation by a function For sender node characteristics Receiver node features And connected edge features The joint coding is performed such that, In order to operate the polymerization process, the polymerization reaction, Are the nodes of the neighboring nodes of each other, The function is updated for the node and, Is a node Is a neighbor node set; the feature space is obtained through the operations of multi-layer graph convolution and feature pooling.
6. The radar point cloud target detection method of claim 5, further comprising: Processing by adopting a message transmission neural network comprising a non-local neural network structure, wherein the non-local neural network structure is as follows: ; wherein the non-local neural network structure characterizes the current node And characteristics of adjacent nodes Updating; is a feature vector conversion for the neighboring node, As a linear embedded function, i.e. , Is a learnable weight function; is a normalization factor, a function The expression of (2) is: ; Wherein, the And Are all a function of the embedment that can be learned, The connection operation is represented by a number of steps, Is a weight vector, for projecting the vector as a scalar, For activating the function, for non-linear conversion.
7. The method for detecting a radar point cloud target according to claim 6, wherein the performing semantic segmentation on the feature space and training with a combined loss function to obtain the target detection model includes: performing segmentation processing by using a semantic segmentation head comprising a shared MLP and a softmax activation function to obtain confidence coefficient and category probability distribution of each point cloud, performing multi-level information extraction and target classification by using a GAT detection head comprising a plurality of hidden layers and detection heads, and processing an overlapped boundary frame by using a non-maximum suppression method; training by combining loss functions to obtain the target detection model, wherein the combined loss functions The method comprises the following steps: ; Wherein, the Adopting FocalLoss multiple classification loss functions; detecting a loss function for an object, and adopting a Huber loss function with a delta value of 1; Regularizing the loss function for L2; 、、 Is the weight.
8. A radar point cloud target detection apparatus, comprising: The first module is used for acquiring radar point cloud data of the target vehicle, and processing the radar point cloud data of the target vehicle by adopting a target detection model to obtain a target recognition result of the target vehicle; The target detection model is obtained through the following modules: A second module for acquiring a dataset, wherein the dataset comprises radar point cloud data; The third module is used for mapping the radar point cloud data as nodes into a graph structure, and dynamically constructing the adjacency relationship of the nodes by adopting a K-nearest neighbor algorithm; A fourth module, configured to determine edge features of the graph structure according to the adjacency relationship of the nodes, and perform matrix conversion on the nodes, the adjacency relationship of the nodes, and the edge features to obtain a graph structure matrix; a fifth module, configured to perform feature update on the nodes and adjacent nodes of the graph structure matrix through graph rolling operation of the message passing neural network, so as to obtain a feature space; and a sixth module, configured to perform semantic segmentation on the feature space, and perform training by using a combined loss function to obtain the target detection model, where the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularized loss function.
9. An electronic device comprising a processor and a memory; the memory is used for storing programs; The processor executing the program implements the radar point cloud target detection method of any of claims 1-7.
10. A computer-readable storage medium, wherein the storage medium stores a program that is executed by a processor to implement the radar point cloud target detection method according to any one of claims 1 to 7.

Description

Lei Dadian cloud target detection method and device, electronic equipment and storage medium Technical Field The present invention relates to the field of radar wave detection technologies, and in particular, to a method and apparatus for detecting a radar point cloud target, an electronic device, and a storage medium. Background The vehicle millimeter wave radar has been applied in the field of automobile environment sensing for over twenty years, and has irreplaceability in a target tracking and anti-collision early warning system by virtue of all-weather working capacity, robustness in severe environment, accurate distance and speed measurement performance and large-range detection advantages. However, the inherent defects of the point cloud data, including high sparsity of detection results, low information density, incapability of obtaining geometric features of the obstacle, and the like, have the bottlenecks of insufficient target classification capability, limited analysis of environmental features, and the like in the complex scene perception, and are difficult to independently meet the complete perception requirement of an automatic driving system on the attribute of the obstacle. Disclosure of Invention The invention aims to at least solve one of the technical problems in the prior art, and provides a radar point cloud target detection method, a radar point cloud target detection device, electronic equipment and a storage medium. One aspect of the present invention provides a radar point cloud target detection method, including: Acquiring radar point cloud data of a target vehicle, and processing the radar point cloud data of the target vehicle by adopting a target detection model to obtain a target recognition result of the target vehicle; The training steps of the target detection model are as follows: acquiring a data set, wherein the data set comprises radar point cloud data; mapping radar point cloud data serving as nodes into a graph structure, and dynamically constructing an adjacency relationship of the nodes by adopting a K-nearest neighbor algorithm; determining edge characteristics of the graph structure according to the adjacent relation of the nodes, and performing matrix conversion on the nodes, the adjacent relation of the nodes and the edge characteristics to obtain a graph structure matrix; carrying out feature update on nodes and adjacent nodes of the graph structure matrix through graph convolution operation of a message passing neural network to obtain a feature space; and carrying out semantic segmentation on the feature space, and training by adopting a combined loss function to obtain the target detection model, wherein the combined loss function comprises a multi-classification loss function, an object detection loss function and an L2 regularized loss function. The radar point cloud target detection method, wherein acquiring the data set further comprises: Preprocessing Lei Dadian cloud data, namely integrating Lei Dadian cloud data in a preset time period, and cutting the front and rear ranges of the integrated vehicle, wherein cutting comprises point-by-point annotation and track identification according to each point cloud, and further determining the front and rear ranges of the vehicle. According to the radar point cloud target detection method, radar point cloud data is mapped into a graph structure as nodes, and the adjacency relationship of the nodes is dynamically constructed by adopting a K-nearest neighbor algorithm, and the method comprises the following steps: Under the condition that the number of the radar point cloud data is kept consistent with that of the nodes, discarding the spatial position characteristics of the point cloud data, and reserving a speed vector, a radar scattering cross section, a time stamp and connectivity for mapping, wherein the connectivity is used for representing the associated edge number of the nodes; and determining the adjacency relation of the nodes by adopting a K-nearest neighbor algorithm according to the Euclidean distance between the nodes and the computational complexity. According to the radar point cloud target detection method, edge characteristics of a graph structure are determined according to the adjacent relation of nodes, and the adjacent relation of the nodes and the edge characteristics are subjected to matrix conversion to obtain a graph structure matrix, wherein the method comprises the following steps: According to the adjacent relation of the nodes, calculating the relative positions of the adjacent nodes, and taking the relative positions as the edge characteristics; and performing matrix conversion on the nodes, the adjacent relations of the nodes and the edge characteristics to obtain a graph structure matrix, wherein the graph structure matrix comprises a node characteristic matrix, an edge characteristic matrix and an adjacent matrix. According to the radar point cloud target detection method, the f