Radar point cloud target detection method and device, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of radar wave detection, and provides a radar point cloud target detection method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining radar point cloud data of a target vehicle, processing the radar point cloud data of the target vehicle through a target detection model, and obtaining a target recognition result of the target vehicle; the target detection model is obtained through training of the MPNN with the NLNN inserted and the combined loss function, the point cloud spatial relation topological graph is constructed, local spatial features and global context information are fused, effective representation of target features is achieved, and the target classification accuracy in complex scene perception is improved.

Description

Technical Field

[0001] This invention relates to the field of radar wave detection technology, and in particular to a radar point cloud target detection method, device, electronic device, and storage medium. Background Technology

[0002] Vehicle-mounted millimeter-wave radar has been used in the field of automotive environmental perception for over two decades. With its all-weather capability, robustness in harsh environments, precise ranging and speed measurement performance, and wide-area detection advantages, it is irreplaceable in target tracking and collision avoidance warning systems. However, due to the inherent limitations of point cloud data—including high sparsity of detection results, low information density, and inability to acquire obstacle geometric features—it suffers from bottlenecks in complex scene perception, such as insufficient target classification capabilities and limited environmental feature analysis. This makes it difficult to independently meet the complete obstacle attribute perception requirements of autonomous driving systems. Summary of the Invention

[0003] In order to at least solve one of the technical problems existing in the prior art, the present invention provides a radar point cloud target detection method, device, electronic device and storage medium.

[0004] One aspect of the present invention provides a radar point cloud target detection method, comprising:

[0005] The radar point cloud data of the target vehicle is acquired, and the radar point cloud data of the target vehicle is processed using a target detection model to obtain the target recognition result of the target vehicle.

[0006] The training steps for the target detection model are as follows:

[0007] Obtain the dataset, which includes radar point cloud data;

[0008] Radar point cloud data is mapped as nodes into a graph structure, and the adjacency relationship of nodes is dynamically constructed using the K-nearest neighbor algorithm;

[0009] The edge features of the graph structure are determined based on the adjacency relationships of the nodes. The nodes, their adjacency relationships, and the edge features are then transformed into a matrix to obtain the graph structure matrix.

[0010] The graph structure matrix is ​​updated by performing graph convolution operations of a message-passing neural network on the nodes and their neighboring nodes to obtain the feature space;

[0011] The feature space is semantically segmented, and the object detection model is trained using a combined loss function, wherein the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

[0012] According to the radar point cloud target detection method described above, obtaining the dataset further includes:

[0013] The radar point cloud data is preprocessed, including integrating radar point cloud data within a preset time period and cropping the integrated range in front of and behind the vehicle. The cropping process involves determining the range in front of and behind the vehicle based on point-by-point annotations and trajectory markers for each point cloud.

[0014] According to the radar point cloud target detection method, radar point cloud data is mapped as nodes into a graph structure, and the adjacency relationship of nodes is dynamically constructed using the K-nearest neighbor algorithm, including:

[0015] While keeping the number of radar point cloud data and nodes consistent, the spatial location features of the point cloud data are discarded, and the velocity vector, radar cross section, timestamp, and connectivity are retained for mapping. The connectivity is used to characterize the number of associated edges of a node.

[0016] The K-nearest neighbor algorithm is used to determine the adjacency relationship of nodes based on the Euclidean distance between nodes and the computational complexity.

[0017] According to the radar point cloud target detection method, the edge features of the graph structure are determined based on the adjacency relationship of nodes. The nodes, their adjacency relationships, and the edge features are then transformed into a matrix to obtain the graph structure matrix, including:

[0018] Based on the adjacency relationship of nodes, calculate the relative position of adjacent nodes, and use the relative position as the edge feature;

[0019] The nodes, their adjacency relationships, and edge features are transformed into a matrix to obtain a graph structure matrix, which includes a node feature matrix, an edge feature matrix, and an adjacency matrix.

[0020] According to the radar point cloud target detection method, the graph structure matrix is ​​updated by performing graph convolution operations of a message-passing neural network on the nodes and their neighboring nodes to obtain a feature space, including:

[0021] The graph convolution operation in a message-passing neural network is as follows:

[0022]

[0023] in, for Layer node characteristics, Layer node features Based on the features of the front-layer nodes Updated together with the neighborhood aggregation result, neighborhood aggregation is performed through a function Features of the sending node Receiver node characteristics and connected edge features Perform joint encoding, For aggregation operations, Neighboring nodes For node update functions, For nodes The set of neighboring nodes;

[0024] The feature space is obtained through multi-layer graph convolution and feature pooling operations.

[0025] According to the radar point cloud target detection method, the method further includes:

[0026] The message-passing neural network, which includes a nonlocal neural network structure, is used for processing. The nonlocal neural network structure is as follows:

[0027]

[0028] In this nonlocal neural network structure, the features of the current node are... Features of adjacent nodes Update; It is a transformation of the feature vectors of adjacent nodes. It is a linear embedding function, i.e. , It is a learnable weight function; It is the normalization factor, the function The expression is:

[0029]

[0030] in, and All are learnable embedding functions. Indicates a connection operation. It is a weight vector used to project a vector into a scalar. This is the activation function used for nonlinear transformations.

[0031] According to the radar point cloud target detection method, the feature space is semantically segmented, and a target detection model is obtained by training using a combined loss function, including:

[0032] A semantic segmentation head, including a shared MLP and softmax activation function, is used to perform segmentation processing to obtain the confidence and class probability distribution of each point cloud; a GAT detection head, including multiple hidden layers and probes, is used for multi-level information extraction and target classification; and a non-maximum suppression method is used to process overlapping bounding boxes.

[0033] The target detection model is obtained by training using a combined loss function, wherein the combined loss function... for:

[0034]

[0035] in, The FocalLoss multi-class loss function is used; The Huber loss function with a Δ value of 1 is used for object detection loss function; The loss function is L2 regularization. , , As weight.

[0036] Another aspect of the present invention provides a radar point cloud target detection device, comprising:

[0037] The first module is used to acquire radar point cloud data of the target vehicle, process the radar point cloud data of the target vehicle using a target detection model, and obtain the target recognition result of the target vehicle.

[0038] The target detection model is obtained through the following modules:

[0039] The second module is used to acquire a dataset, which includes radar point cloud data.

[0040] The third module is used to map radar point cloud data as nodes into a graph structure, and dynamically constructs the adjacency relationship of nodes using the K-nearest neighbor algorithm.

[0041] The fourth module is used to determine the edge features of the graph structure based on the adjacency relationship of the nodes, and to perform matrix transformation on the nodes, the adjacency relationship of the nodes, and the edge features to obtain the graph structure matrix;

[0042] The fifth module is used to update the features of the nodes and their adjacent nodes on the graph structure matrix through graph convolution operations of a message-passing neural network to obtain the feature space.

[0043] The sixth module is used to perform semantic segmentation on the feature space and to train the target detection model using a combined loss function, wherein the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

[0044] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0045] The memory is used to store programs;

[0046] The processor executes the program to implement the method as described above.

[0047] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0048] The beneficial effects of this invention are as follows: by constructing a point cloud spatial relationship topology map and integrating local spatial features with global contextual information, the effective representation of target features is achieved, thereby improving the target classification accuracy in complex scene perception. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the radar point cloud target detection process according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the front point cloud cropping according to an embodiment of the present invention.

[0051] Figure 3 This is a diagram of the MPNN network structure with inserted NLNN according to an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of IoU according to an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of a radar point cloud target detection device according to an embodiment of the present invention. Detailed Implementation

[0054] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0055] refer to Figure 1 , Figure 1This is a schematic diagram of the radar point cloud target detection process according to an embodiment of the present invention, which includes, but is not limited to, steps S100~S600:

[0056] S100: Acquire radar point cloud data of the target vehicle, process the radar point cloud data of the target vehicle using a target detection model, and obtain the target recognition result of the target vehicle.

[0057] The training steps for the object detection model are S200~S600:

[0058] S200, acquire the dataset, which includes radar point cloud data.

[0059] In some embodiments, the radar point cloud data is preprocessed, including integrating radar point cloud data within a preset time period and cropping the integrated range in front of and behind the vehicle. The cropping includes determining the range in front of and behind the vehicle based on point-by-point annotations and trajectory markers for each point cloud.

[0060] For example, refer to Figure 2 The diagram shown illustrates the cropping of the front point cloud data. This embodiment of the invention uses the RadarScenes dataset for neural network training. Before model training, the raw data needs to be preprocessed, including integrating the radar point cloud data accumulated within a 500ms time interval to serve as the model's input; and cropping the resulting point cloud to a 100m×100m area in front of or behind the vehicle's rear axle. Figure 2 Since the RadarScenes dataset does not directly provide traditional ground truth bounding boxes, it is necessary to infer all points belonging to the same object through point-by-point annotation and trajectory ID, and then use this information to construct bounding boxes.

[0061] The S300 maps radar point cloud data as nodes to a graph structure and uses the K-nearest neighbor algorithm to dynamically construct the adjacency relationships of the nodes.

[0062] In some embodiments, while keeping the radar point cloud data consistent with the number of nodes, the spatial location features of the point cloud data are discarded, and the velocity vector, radar cross section, timestamp, and connectivity are retained for mapping. The connectivity is used to characterize the number of associated edges of a node. The K-nearest neighbor algorithm is used to determine the adjacency relationship of nodes based on the Euclidean distance between nodes and the computational complexity.

[0063] In some embodiments, the original point cloud points are mapped to graph structure points, and the original point cloud set is defined as follows: , , , ,in For the number of point clouds, It is a finite set. A set of nodes is generated through projection. In the mapping transformation, the number of point cloud points and the number of node set points remain consistent to ensure full coverage of the original point cloud. Node feature selection needs to balance information integrity and computational efficiency. To avoid interference from absolute coordinates on the model, spatial coordinates are discarded for node features, retaining only the velocity vector. , The design incorporates radar cross-section (RCS) and associated timestamp t, and introduces connectivity c (number of edges associated with a node) as an additional feature. This design achieves structural information transmission through relative relation encoding, providing a foundation for subsequent edge generation.

[0064] In some embodiments, edge generation includes:

[0065] The edge set is defined as , , The generation strategy needs to consider the ambiguity of adjacency relationships caused by the sparsity of radar point clouds. Due to the characteristics of radar point clouds, the nodes in the graph structure are far apart. If the radius adjacency graph method is used to determine the edges of the graph, the radius setting will be difficult. Using a small radius will result in some nodes of the same object not being able to form connections; using a large radius will result in large objects covering small objects, specifically, the nodes of two objects are connected together and cannot be distinguished. The embodiment of this invention uses the K-nearest neighbor algorithm to dynamically construct adjacency relationships, using Euclidean distance as the metric, and setting K to 20 to balance topological integrity and computational complexity.

[0066] S400: Determine the edge features of the graph structure based on the adjacency relationships of the nodes, and perform matrix transformation on the nodes, node adjacency relationships, and edge features to obtain the graph structure matrix.

[0067] In some embodiments, the relative positions of adjacent nodes are calculated based on the adjacency relationship of the nodes, and the relative positions are used as edge features; the nodes, the adjacency relationships of the nodes, and the edge features are transformed into a matrix to obtain a graph structure matrix, wherein the graph structure matrix includes a node feature matrix, an edge feature matrix, and an adjacency matrix.

[0068] It is understood that, in order to preserve the spatial invariance property, the edge features in this embodiment of the invention use the relative positions of adjacent nodes. We abandon absolute coordinates to achieve translation invariance.

[0069] In some embodiments, matrix transformation includes transforming graph structures Convert to matrix form to fit neural network input, node feature matrix Integrating attributes such as speed and RCS, edge feature matrix Encoding relative position information, adjacency matrix Representing the node connection state, the final output tuple As a standardized input for graph neural networks.

[0070] S500 updates the feature space by performing graph convolution operations on nodes and their neighboring nodes on the graph structure matrix through a message-passing neural network.

[0071] In some embodiments, the graph convolution (MPNN) operation of a message-passing neural network is as follows:

[0072]

[0073] in, for Layer node characteristics, Layer node features Based on the features of the front-layer nodes Updated together with the neighborhood aggregation result, neighborhood aggregation is performed through a function Features of the sending node Receiver node characteristics and connected edge features Perform joint encoding, For aggregation operations, Neighboring nodes For node update functions, For nodes The set of neighboring nodes; the feature space is obtained through multi-layer graph convolution and feature pooling operations.

[0074] In some embodiments, node feature embedding is implemented using a four-layer MLP, while edge feature embedding uses a three-layer MLP. The parameters of both are shared among similar elements to maintain feature consistency. During the message passing stage, the embodiments of the present invention adopt the MPNN framework. Through the above embodiments, the network can adaptively adjust the feature dimension, providing flexibility for cross-layer feature fusion.

[0075] In yet another embodiment, such as Figure 3 The MPNN network structure diagram with inserted NLNN shown is processed using a message-passing neural network that includes a nonlocal neural network structure (NLNN). The nonlocal neural network structure is as follows:

[0076]

[0077] In this nonlocal neural network structure, the features of the current node are... Features of adjacent nodes Update; It is a transformation of the feature vectors of adjacent nodes. It is a linear embedding function, i.e. , It is a learnable weight function; It is the normalization factor, the function The expression is:

[0078]

[0079] in, and All are learnable embedding functions. Indicates a connection operation. It is a weight vector used to project a vector into a scalar. This is the activation function used for nonlinear transformations.

[0080] Understandably, the NLNN structure can capture long-distance node dependencies to enhance node representation.

[0081] S600 performs semantic segmentation on the feature space and trains the object detection model using a combined loss function, which includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

[0082] In some embodiments, a semantic segmentation head including a shared MLP and softmax activation function is used to perform segmentation processing to obtain the confidence and class probability distribution of each point cloud; a GAT detection head including multiple hidden layers and probes is used for multi-level information extraction and target classification; and a non-maximum suppression method is used to process overlapping bounding boxes.

[0083] The object detection model is obtained by training with a combined loss function, which consists of multiple task-specific loss functions. for:

[0084]

[0085] in, The FocalLoss multi-class loss function is used; The Huber loss function with a Δ value of 1 is used for object detection loss function; The loss function is L2 regularization. , , For weights, the weights are set to... , and To balance the losses between different items.

[0086] In some embodiments, the proposed radar point cloud target recognition model adopts a multi-task collaborative learning paradigm, simultaneously performing semantic segmentation and target detection tasks. This framework achieves cross-task knowledge transfer by sharing a low-level feature extraction module, while maintaining task decoupling through independent task-specific headers. Its architecture can be flexibly extended to other perception tasks.

[0087] The semantic segmentation head consists of a shared MLP, with the softmax activation function at the end generating a pointwise class probability distribution. The final rank of each point is determined by the highest confidence score among all ranks. The object detection head is built using the GAT attention mechanism, which employs multiple hidden layers and multiple probes. This construction achieves multi-level information extraction and object classification. The hidden layers are specifically graph attention layers. For each attention head, the input features and adjacency matrix are passed in to obtain the attention structure of each head. The results are concatenated, then passed through a dropout and an additional attention layer, and finally a non-linear activation function is applied to output the final result. Here, the input features and adjacency matrix are the node features and adjacency matrix output by the MPNN message passing module. For the resulting bounding boxes, the non-maximum suppression (NMS) method is used to handle overlapping bounding boxes, retaining only the bounding box with the highest confidence score.

[0088] In some embodiments, such as Figure 4 As shown, in image-based target detection, the conventional method for determining whether a predicted bounding box matches a real target is to calculate the pixel-based Intersection over Union (IoU). Analogous to the IoU proposed in radar point cloud target detection, this method replaces the pixel-based calculation with the radar point cloud.

[0089]

[0090] Here, Predicted Points represent predicted point clouds, and True Points represent true point clouds. If the predicted IoU is greater than or equal to a certain threshold, the object instance is defined as a match. This threshold is typically set to 0.5. However, for point cloud-based IoU definitions, an IoU greater than or equal to 0.5 is a very stringent condition. For example, the figure simulates a real-world point cloud of a pedestrian surrounded by noisy data. The noise and the elongated shape of the object cause the IoU to drop significantly even if the prediction differs slightly from the true label. Therefore, to avoid limiting the model's performance, this embodiment sets the IoU to be greater than or equal to 0.3.

[0091] When a bounding box successfully matches a ground truth label, it is classified as a True Positive (TP) if its predicted class matches the ground truth class. Other redundant bounding boxes on the same ground truth target (i.e., the same target detected multiple times) are classified as False Positives (FP). Ground truth target instances that are not matched are counted as False Negatives (FN), and all other cases are considered True Negatives (TN). The matching order of bounding boxes is determined by the object confidence score (c) attached to each object detection output. For each class, bounding boxes with higher confidence scores are matched first, while those with lower confidence scores are processed later.

[0092] The most commonly used evaluation metrics in object detection are Average Precision (AP) (for each class) and Mean Average Precision (mAP) (a combined metric across all classes). Precision and recall based on the confidence score c are calculated as follows:

[0093]

[0094]

[0095] AP is calculated using precision and recall, as shown in the following formula:

[0096]

[0097] For mAP, first calculate the AP for each category independently, then take the average:

[0098]

[0099] in Indicates the number of target categories ( (This represents the total number of categories; subtracting 1 indicates background exclusion). Individual AP can demonstrate the model's learning ability for different targets, facilitating targeted model improvements. mAP provides an overall performance representation, conveniently showing the overall performance of different models. This invention uses both methods for better comparison and optimization directions.

[0100] In radar-related target detection research, a commonly used metric is the F1 score, which is the harmonic mean of precision (Pr) and recall (Re). For target class k, its maximum F1 score is defined as:

[0101]

[0102] Where c is the confidence threshold. Typically, the model provides the average F1 score for each category as the result.

[0103] To verify the advancements achieved by the network in this embodiment, the experimental section compares the model with the PointPillars, YOLOv3, and RadarGNN networks. Table 1 shows the mAP (measurable accuracy) for object detection across different networks, and also provides AP scores for specific categories such as pedestrians, pedestrian groups, two-wheeled vehicles, cars, and large vehicles, based on the evaluation method used in the RadarScenes dataset.

[0104] Table 1 Comparison of results from different models

[0105] method AP pedestrians AP pedestrian group AP two-wheeler AP Auto AP Trucks mAP PointPillars 11.2 22.5 38.8 54.2 53.7 36.1 YOLOv3 34.4 55.7 57.4 70.2 61.9 55.9 RadarGNN 33.1 54.8 62.0 70.4 62.0 56.5 Ours 26.8 58.1 59.8 72.7 72.1 57.9

[0106] As shown in Table 1, the network proposed in this embodiment of the invention achieved a mAP of 57.9% on the RadarScenes validation set. Among specific categories, cars and large vehicles had the highest AP values, while pedestrians had the lowest. Compared with other object detection methods, the model proposed in this embodiment of the invention achieved leading results on the RadarScenes dataset, outperforming other object detection methods. It improved by 1.4% compared to similar networks like RadarGNN, surpassing the hybrid PointPillars model and YOLOv3-based methods.

[0107] This invention addresses the limitations of traditional deep learning networks in processing non-geometric information and capturing local features. It proposes a method using graph neural networks to process radar point cloud signals. By introducing an MPNN convolutional architecture to transmit point cloud information, and addressing the performance shortcomings of MPNN in capturing long-distance node dependencies, an NLNN module is embedded within the MPNN network to enhance its long-distance dependency modeling capabilities. Experimental results demonstrate that this model outperforms traditional models in various category recognition tasks.

[0108] Figure 5 This is a schematic diagram of a radar point cloud target detection device according to an embodiment of the present invention. The device includes a first module 510, a second module 520, a third module 530, a fourth module 540, a fifth module 550, and a sixth module 550.

[0109] The system comprises six modules: First, acquiring radar point cloud data of the target vehicle; second, acquiring a dataset including radar point cloud data; third, mapping the radar point cloud data as nodes to a graph structure and dynamically constructing adjacency relationships using the K-nearest neighbor algorithm; fourth, determining edge features of the graph structure based on node adjacency relationships, performing matrix transformation on nodes, their adjacency relationships, and edge features to obtain a graph structure matrix; fifth, updating the features of nodes and their adjacent nodes using graph convolution operations of a message-passing neural network to obtain a feature space; and sixth, performing semantic segmentation on the feature space and training with a combined loss function to obtain the target detection model, which includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

[0110] For example, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned radar point cloud target detection methods, namely, acquiring radar point cloud data of the target vehicle, processing the radar point cloud data of the target vehicle using a target detection model, and obtaining the target recognition result of the target vehicle; the training steps of the target detection model are as follows: acquiring a dataset, wherein the dataset includes radar point cloud data; mapping the radar point cloud data as nodes to a graph structure, and dynamically constructing the adjacency relationship of the nodes using the K-nearest neighbor algorithm; determining the edge features of the graph structure based on the adjacency relationship of the nodes, performing matrix transformation on the nodes, the adjacency relationship of the nodes, and the edge features to obtain a graph structure matrix; updating the features of the nodes and their adjacent nodes through graph convolution operation of a message-passing neural network on the graph structure matrix to obtain a feature space; performing semantic segmentation on the feature space; and training using a combined loss function to obtain a target detection model, wherein the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function. The beneficial effects of this invention are as follows: by constructing a point cloud spatial relationship topology map and integrating local spatial features with global contextual information, the effective representation of target features is achieved, thereby improving the target classification accuracy in complex scene perception.

[0111] This invention also provides an electronic device, which includes a processor and a memory;

[0112] The memory stores the program;

[0113] The processor executes a program to perform the aforementioned radar point cloud target detection method; the electronic device has the function of carrying and running the software system for radar point cloud target detection provided in the embodiments of the present invention, such as a personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0114] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the radar point cloud target detection method described above.

[0115] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0116] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned radar point cloud target detection method.

[0117] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0118] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0119] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0120] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0121] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0122] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0123] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0124] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A radar point cloud target detection method, characterized in that, include: The radar point cloud data of the target vehicle is acquired, and the radar point cloud data of the target vehicle is processed using a target detection model to obtain the target recognition result of the target vehicle. The training steps for the target detection model are as follows: Obtain the dataset, which includes radar point cloud data; Radar point cloud data is mapped as nodes into a graph structure, and the adjacency relationship of nodes is dynamically constructed using the K-nearest neighbor algorithm; The edge features of the graph structure are determined based on the adjacency relationships of the nodes. The nodes, their adjacency relationships, and the edge features are then transformed into a matrix to obtain the graph structure matrix. The graph structure matrix is ​​updated by performing graph convolution operations of a message-passing neural network on the nodes and their neighboring nodes to obtain the feature space; The feature space is semantically segmented, and the object detection model is trained using a combined loss function, wherein the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

2. The radar point cloud target detection method according to claim 1, characterized in that, The acquisition of the dataset also includes: The radar point cloud data is preprocessed, including integrating radar point cloud data within a preset time period and cropping the integrated range in front of and behind the vehicle. The cropping process involves determining the range in front of and behind the vehicle based on point-by-point annotations and trajectory markers for each point cloud.

3. The radar point cloud target detection method according to claim 1, characterized in that, The process of mapping radar point cloud data as nodes to a graph structure and dynamically constructing the adjacency relationships of nodes using the K-nearest neighbor algorithm includes: While keeping the number of radar point cloud data and nodes consistent, the spatial location features of the point cloud data are discarded, and the velocity vector, radar cross section, timestamp, and connectivity are retained for mapping. The connectivity is used to characterize the number of associated edges of a node. The K-nearest neighbor algorithm is used to determine the adjacency relationship of nodes based on the Euclidean distance between nodes and the computational complexity.

4. The radar point cloud target detection method according to claim 3, characterized in that, The step of determining the edge features of the graph structure based on the adjacency relationships of nodes involves performing a matrix transformation on the nodes, their adjacency relationships, and the edge features to obtain the graph structure matrix, including: Based on the adjacency relationship of nodes, calculate the relative position of adjacent nodes, and use the relative position as the edge feature; The nodes, their adjacency relationships, and edge features are transformed into a matrix to obtain a graph structure matrix, which includes a node feature matrix, an edge feature matrix, and an adjacency matrix.

5. The radar point cloud target detection method according to claim 1, characterized in that, The step of updating the feature space by performing graph convolution operations on the graph structure matrix using a message-passing neural network to update the features of nodes and their neighboring nodes includes: The graph convolution operation in a message-passing neural network is as follows: ； in, for Layer node characteristics, Layer node features Based on the features of the front-layer nodes Updated together with the neighborhood aggregation result, neighborhood aggregation is performed through a function Features of the sending node Receiver node characteristics and connected edge features Perform joint encoding, For aggregation operations, Neighboring nodes For node update functions, For nodes The set of neighboring nodes; The feature space is obtained through multi-layer graph convolution and feature pooling operations.

6. The radar point cloud target detection method according to claim 5, characterized in that, The method further includes: The message-passing neural network, which includes a nonlocal neural network structure, is used for processing. The nonlocal neural network structure is as follows: ； In this nonlocal neural network structure, the features of the current node are... Features of adjacent nodes Update; It is a transformation of the feature vectors of adjacent nodes. It is a linear embedding function, i.e. , It is a learnable weight function; It is the normalization factor, the function The expression is: ； in, and All are learnable embedding functions. Indicates a connection operation. It is a weight vector used to project a vector into a scalar. This is the activation function used for nonlinear transformations.

7. The radar point cloud target detection method according to claim 6, characterized in that, The step of semantically segmenting the feature space and training the target detection model using a combined loss function includes: A semantic segmentation head, including a shared MLP and softmax activation function, is used to perform segmentation processing to obtain the confidence and class probability distribution of each point cloud; a GAT detection head, including multiple hidden layers and probes, is used for multi-level information extraction and target classification; and a non-maximum suppression method is used to process overlapping bounding boxes. The target detection model is obtained by training using a combined loss function, wherein the combined loss function... for: ； in, The FocalLoss multi-class loss function is used; The Huber loss function with a Δ value of 1 is used for object detection loss function; The loss function is L2 regularization. , , As weight.

8. A radar point cloud target detection device, characterized in that, include: The first module is used to acquire radar point cloud data of the target vehicle, process the radar point cloud data of the target vehicle using a target detection model, and obtain the target recognition result of the target vehicle. The target detection model is obtained through the following modules: The second module is used to acquire a dataset, which includes radar point cloud data. The third module is used to map radar point cloud data as nodes into a graph structure, and dynamically constructs the adjacency relationship of nodes using the K-nearest neighbor algorithm. The fourth module is used to determine the edge features of the graph structure based on the adjacency relationship of the nodes, and to perform matrix transformation on the nodes, the adjacency relationship of the nodes, and the edge features to obtain the graph structure matrix; The fifth module is used to update the features of the nodes and their adjacent nodes on the graph structure matrix through graph convolution operations of a message-passing neural network to obtain the feature space. The sixth module is used to perform semantic segmentation on the feature space and to train the target detection model using a combined loss function, wherein the combined loss function includes a multi-classification loss function, an object detection loss function, and an L2 regularization loss function.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the radar point cloud target detection method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program, which is executed by a processor to implement the radar point cloud target detection method as described in any one of claims 1-7.

Images