A radar point cloud recognition method, device, equipment and computer storage medium

Abstract

The application discloses a radar point cloud recognition method, device and equipment and a computer storage medium, and comprises the following steps: generating training data sets based on different point cloud data, constructing cloud picture data sets according to the point cloud data in the training data sets, extracting the advanced features F of the cloud picture time sequence in the cloud picture data sets by using a feature extraction module, extracting the point cloud space structure features S of the advanced features F, capturing the time dependence of the point cloud data of the advanced features F, obtaining point cloud time dependence features T, inputting the point cloud space structure features S and the point cloud time dependence features T into a full connection neural network model for training, obtaining a trained neural network model, detecting to-be-recognized point cloud data by using the trained neural network model, and outputting a recognition result. The application improves the efficiency and accuracy of feature extraction, retains the three-dimensional space characteristics of point clouds, reduces the calculation complexity, and improves the feature extraction efficiency requirement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of millimeter wave radar, in particular to a radar point cloud recognition method, device, equipment and computer storage medium. BACKGROUND

[0002] 3D data can provide rich coordinate, geometry, shape and other information, and has many applications in the fields of autonomous driving, medical treatment, AR / VR and the like. 3D data can be represented in different formats, including depth images, point clouds, meshes and volumetric meshes and the like. Among them, the point cloud is a definition of a point set in a three-dimensional space, and has become one of the most important data formats in three-dimensional representation, and is increasingly attracting the attention of researchers.

[0003] Deep learning is now the most powerful data processing tool in computer vision, and has become the preferred technology for classification, recognition and detection tasks, and is mainly applied to data with structured grids, while point clouds are unstructured. The non-structured and unordered nature of point clouds makes it very challenging to directly process them using deep learning.

[0004] Current deep learning methods for processing point clouds can be divided into three categories: multi-view based methods, voxel based methods and point based methods. The multi-view based method converts a three-dimensional object into two-dimensional images at different angles, uses two-dimensional convolution to extract features, and has good performance in classification tasks, but loses the three-dimensional spatial position information in the process of converting into two-dimensional images, and cannot well utilize the three-dimensional characteristics of point cloud data; the voxel based method converts point cloud into a three-dimensional voxel structure of XxYxZ, and respectively convolves with XxYxZ three-dimensional kernel and X, Y, Z, but this method retains the three-dimensional spatial position information of the data, but the calculation amount is huge and the efficiency is slow; the point based method uses the original point cloud data as input, first multiplies with a T-Net learned conversion matrix, then performs feature extraction through multiple MLPs, and then aligns the features using a T-Net, and finally obtains global features through maximum pooling operation, but this method increases the computational burden to solve the unordered nature of point cloud data, increases the computational complexity and reduces the feature extraction efficiency.

[0005] As can be seen from the above, the current point cloud recognition method cannot meet the needs of retaining three-dimensional spatial characteristics while reducing computational complexity and improving feature extraction efficiency. SUMMARY

[0006] The purpose of the present application is to provide a radar point cloud recognition method, device, equipment and computer storage medium, to solve the problem that the prior art cannot meet the needs of retaining three-dimensional spatial characteristics while reducing computational complexity and improving feature extraction efficiency.

[0007] To solve the above technical problems, the present application provides a radar point cloud recognition method, comprising:

[0008] Generating a training data set based on different point cloud data;

[0009] Constructing a cloud map data set according to the point cloud data in the training data set;

[0010] Extracting the advanced features F of the cloud map time sequence in the cloud map data set by using a feature extraction module;

[0011] Extracting the point cloud spatial structure features S of the advanced features F;

[0012] Capturing the time dependence of the point cloud data of the advanced features F to obtain point cloud time dependence features T;

[0013] Inputting the point cloud spatial structure features S and the point cloud time dependence features T into a fully connected neural network model for training to obtain a trained neural network model;

[0014] Detecting the to-be-recognized point cloud data by using the trained neural network model and outputting a recognition result.

[0015] Preferably, constructing a cloud map data set according to the point cloud data in the training data set comprises:

[0016] Defining a point cloud with N points as a set P = {p1, p1, …, pN}, wherein p = (X, S), n = 1, 2, …, N, X represents the three-dimensional spatial position information of the node, and S is the state vector of the node; n n n n n n

[0017] Given a point cloud P, a cloud map is constructed according to the points in the point cloud and the points within a fixed radius distance r centered on each point, P is taken as a vertex, and adjacent points within the fixed radius r are connected to obtain a cloud map G = (P, E);

[0018] wherein E = {(p i , p j ) || | x i -x j ||2 < r}, P represents a node set, E represents an edge set, p i is a target node, pj is a neighbor node, x i is the three-dimensional spatial position information of the target node, and x j is the three-dimensional spatial position information of the neighbor node.

[0019] ​​​​​​Preferably, the feature extraction module is composed of multiple convolutional layers, each of which is cascaded with a PReLU activation function.

[0020] Preferably, the convolutional layer introduces an expansion coefficient ε, assuming that the original convolution kernel size is k, and after adding d holes, the receptive field size of the convolution kernel is k+(k-1)*d.

[0021] Preferably, the extraction of the point cloud spatial structure feature S of the high-level feature F includes:

[0022] The mutual influence degree between the target node and the adjacent node in the high-level feature F is learned through the graph attention network, and different weights are assigned to the target node and the adjacent node to obtain the point cloud spatial structure feature S.

[0023] Preferably, the graph attention network is given a target node p i and an adjacent node p j , the feature vectors of the target node p i and the adjacent node p j are h i and h j respectively, and the correlation coefficient e ij between the two nodes is calculated =LeakyReLU(a T ·(h i ||h j )), wherein LeakyReLU is a nonlinear activation function, a is a learnable parameter vector, and || is a vector splicing operation.

[0024] The normalized correlation coefficient between the target node p i and the adjacent node p j is obtained by using the softmax function for normalization operation. wherein K is the number of adjacent nodes of the target node p i , and exp(e ik ) is an exponential function.

[0025] The normalized correlation coefficient and the corresponding feature are linearly combined through the Sigmoid activation function σ(·) to obtain the point cloud spatial structure feature of the target node p i output by the graph attention network.

[0026] Preferably, the capture of the time dependence of the high-level feature F point cloud data obtains a point cloud time dependence feature T, which includes:

[0027] The time dependence of the time step distance in the time sequence is captured by using the gated recurrent unit, and the point cloud time dependence feature is calculated by the fully connected layer of the activation function sigmoid.

[0028] The application further provides a radar point cloud recognition device, comprising:

[0029] A training set generation module generates training data sets based on different point cloud data;

[0030] A cloud map construction module is configured to construct a cloud map data set according to the point cloud data in the training data set;

[0031] A feature extraction module is configured to extract high-level features F of a cloud map time sequence in the cloud map data set;

[0032] A graph attention network module is configured to extract point cloud spatial structure features S of the high-level features F;

[0033] A gated recurrent module is configured to capture time dependence of the high-level features F of the point cloud data, to obtain point cloud time dependence features T;

[0034] A training module is configured to input the point cloud spatial structure features S and the point cloud time dependence features T into a fully connected neural network model for training, to obtain a trained neural network model;

[0035] A detection module is configured to detect to-be-recognized point cloud data by using the trained neural network model, and output a recognition result.

[0036] The application further provides a radar point cloud recognition device, comprising:

[0037] A millimeter wave radar is configured to generate point cloud data;

[0038] A memory is configured to store a computer program;

[0039] A processor is configured to implement the steps of the radar point cloud recognition method according to any one of the above when executing the computer program.

[0040] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the radar point cloud recognition method according to any one of the above.

[0041] The radar point cloud recognition method provided by the application constructs a cloud map data set according to point cloud data, uses a graph structure to represent the point cloud data, provides an effective means for point cloud data recognition, extracts point cloud spatial structure features and point cloud time dependence features of the high-level features, uses a network structure in parallel to extract point cloud space-time features, improves the efficiency and accuracy of feature extraction, retains the three-dimensional spatial characteristics of the point cloud while reducing the computational complexity, and improves the efficiency requirement of feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to make the technical scheme of the present application or prior art clearer, the accompanying drawings needed in the description of the embodiments or prior art will be briefly introduced. Obviously, the accompanying drawings described are only some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work on the basis of the embodiments of the present application shall fall within the scope of the present application.

[0043] Figure 1 A flow chart of a first specific embodiment of a radar point cloud recognition method provided by the present application;

[0044] Figure 2 A structure diagram of a radar point cloud recognition method;

[0045] Figure 3 A point cloud data format diagram;

[0046] Figure 4 A feature extraction module schematic diagram composed of a hollow convolution;

[0047] Figure 5 A graph attention network principle schematic diagram;

[0048] Figure 6 A graph attention network implementation mechanism schematic diagram;

[0049] Figure 7 A radar point cloud recognition method confusion matrix obtained by space-time feature fusion;

[0050] Figure 8 A structure block diagram of a radar point cloud recognition device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0051] The core of the present application is to provide a radar point cloud recognition method, device, equipment and computer storage medium, which retains the point cloud three-dimensional space characteristics, reduces the calculation complexity and improves the feature extraction efficiency.

[0052] In order to make the technical scheme of the present application or prior art clearer, the accompanying drawings needed in the description of the embodiments or prior art will be briefly introduced. Obviously, the accompanying drawings described are only some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work on the basis of the embodiments of the present application shall fall within the scope of the present application.

[0053] Please refer to Figure 1 , Figure 1 A flow chart of a first specific embodiment of a radar point cloud recognition method provided by the present application; the specific operation steps are as follows:

[0054] Step S101: generating a training data set based on different point cloud data.

[0055] Step S102: constructing a cloud graph data set according to the point cloud data in the training data set;

[0056] A point cloud with N points is defined as a set P = {p1, p1, …, p n}, wherein p n = (X n , S n ), n = 1, 2, …, N, X n represents the three-dimensional spatial position information of the node, and S n is the state vector of the node;

[0057] Given a point cloud P, a cloud graph is constructed according to the points in the point cloud and the points within a fixed radius distance R centered at each point. P is taken as a vertex, and adjacent points within a fixed radius r are connected to obtain a cloud graph G = (P, E);

[0058] wherein E = {(p i , p j )|||x i -x j ||2 < r}, P represents a node set, E represents an edge set, p i is a target node, p j is a neighbor node, x i is the three-dimensional spatial position information of the target node, and x i is the three-dimensional spatial position information of the neighbor node.

[0059] Step S103: extracting high-level features F of a cloud graph time sequence in the cloud graph data set by using a feature extraction module;

[0060] The feature extraction module composed of a hollow convolution includes multiple hollow convolution layers, and a PReLU activation function is cascaded after each hollow convolution layer.

[0061] Further, the hollow convolution layer is a kind of convolution, which adds a hollow in the filter of the standard convolution, introduces a hyperparameter called “expansion coefficient ε”, and increases the receptive field by adding a hollow. Assuming that the original convolution kernel size is k, then the receptive field size of the convolution kernel after adding d hollows is k + (k-1)*d, so as to obtain larger range of feature information.

[0062] Step S104: extracting a point cloud spatial structure feature S of the high-level feature F;

[0063] The mutual influence degree between the target node and the adjacent node in the high-level feature F is learned by a graph attention network, and different weights are assigned to the target node and the adjacent node to obtain a point cloud spatial structure feature S.

[0064] The graph attention network gives a target node p i and an adjacent node p j , the feature vectors of the target node p i and the adjacent node p j are h i and h j respectively, and the correlation coefficient e ij between the two nodes is calculated = LeakyReLU(a T ·(h i ||h j )), where LeakyReLU is a nonlinear activation function, a is a learnable parameter vector, and || is a vector splicing operation;

[0065] The normalized correlation coefficient between the target node p i and the adjacent node p j is obtained by using the softmax function for normalization operation where K is the number of adjacent nodes of the target node p i , and exp(e ik ) is an exponential function;

[0066] The normalized correlation coefficient is linearly combined with the corresponding features by the Sigmoid activation function σ(·) to obtain the target node p i The point cloud spatial structure feature is output by the graph attention network.

[0067] Step S105: Capture the time dependence of the high-level feature Fpoint cloud data to obtain a point cloud time dependence feature T.

[0068] Step S106: Input the point cloud spatial structure feature S and the point cloud time dependence feature T into a fully connected neural network model for training to obtain a trained neural network model.

[0069] The time dependence of the time step distance in the time sequence is captured by using a gated recurrent unit, and the point cloud time dependence feature is calculated by a fully connected layer of an activation function sigmoid.

[0070] Step S107: Use the trained neural network model to detect the to-be-identified point cloud data and output an identification result.

[0071] The embodiment provides a radar point cloud recognition method, uses a graph structure to represent point cloud data, provides an effective means for point cloud recognition, retains point cloud three-dimensional space characteristics, uses a high-level feature extraction module composed of a hollow convolution to extract high-level features of each time sequence input of point cloud data, reduces loss of feature information, uses a parallel network structure to extract point cloud space-time features, improves feature extraction efficiency and accuracy, realizes retention of point cloud three-dimensional space characteristics, reduces computational complexity, and improves feature extraction efficiency.

[0072] Based on the above embodiment, the embodiment specifically describes a radar point cloud recognition method by using specific devices and data, as shown in the following specific process: Figure 2

[0073] The data used in the embodiment is from a data set MMActivity. The data set is collected and made by a developer through a TI millimeter wave radar development board IWR1443. The specific collection scene is as follows: the millimeter wave radar is placed on a 1.3-meter tripod, and a tester performs five different behavior activities in front of the radar, including: Walking, Jumping, Jumping Jacks, Squats and Boxing, and data is collected within about 20 seconds of continuous time. The data of the radar is sent to a computer through a USB and a robot operating system (Robot Operating System, ROS), and the data is recorded and stored by using rogbag, and then the data is converted into a text format, as shown in the following formula: Figure 3 The data set collects a total of 93 minutes of data, and the point cloud data includes three-dimensional space coordinates, distance, speed azimuth angle and reflection intensity and the like.

[0074] Step S201: Construct a cloud map according to each point in the point cloud and the near neighbors with a fixed radius distance centered on the point, and each vertex feature in the cloud map includes not only 3D position information but also a fixed length state value vector;

[0075] Specifically, the construction of the graph according to the near neighbors with a fixed radius distance centered on each point in the point cloud refers to: defining a point cloud with N points as a set P={p1, p1, …, pN}, wherein p n = (X n , S n ), n = 1, 2, …, N, X n represents three-dimensional space position information of a node, x n = (x, y, z), s n is a state vector of the node, including three-dimensional space coordinate information and state information of the point, and s n = (s n ​=(x, y, z, range, velocity, doppler_bin, bearing, intensity). Given a point cloud P, construct a cloud graph G based on the point cloud P. Take P as the vertex and connect the adjacent points within a fixed radius r to obtain the cloud graph representation G = (P, E), E = {(p i , p j )|||x i -x j ||2<r}, P represents the node set, and E represents the edge set.

[0076] Step S202: Input the constructed cloud map data into a feature extraction module composed of dilated convolution to extract high-level features of each time series input of the point cloud data;

[0077] Specifically, the advanced feature extraction module composed of dilated convolutions contains four dilated convolution layers, each of which is cascaded with a PReLU activation function. The expansion coefficients of the four Conv dilated convolution layers are 1, 2, 4, and 8, respectively. The specific structure is as follows: Figure 4 shown.

[0078] Furthermore, the dilated convolution layer is characterized by the addition of holes to the filter of a standard convolution, introducing a hyperparameter called the dilation coefficient. Like convolution, a dilated convolution operation slides the convolution kernel over a feature map and performs a convolution operation to generate a new feature map. However, the addition of holes increases the receptive field, capturing feature information over a wider range. When the dilation coefficient is 1, the dilated convolution is equivalent to a standard convolution.

[0079] Step S203: Using the graph attention network module to learn the spatial structure features of the point cloud to obtain the spatial structure features of the point cloud;

[0080] Furthermore, using graph attention neural network to learn the spatial structural features of point cloud includes constructing the radar point cloud data into a graph, so that the point cloud data can be regarded as a complete graph structure. Figure 5 As shown in the figure, a node represents a feature, and an edge represents the relationship between the corresponding nodes. The graph attention network learns the degree of mutual influence between the target node and adjacent nodes, and assigns different weights to them in the subsequent training process to obtain updated features;

[0081] Furthermore, if Figure 6 As shown, the graph attention network is given a target node p i and an adjacent node p j , the target node p i and the adjacent node p j The eigenvectors of h are i and hj , calculate the correlation coefficient e between two nodes ij = LeakyReLU(a T · (h i || h j )), where LeakyReLU is a nonlinear activation function, a is a learnable parameter vector, and || is a vector concatenation operation; in order to make the sum of all neighbor node attention coefficients equal to 1, a normalization operation is performed using a softmax function to obtain the normalized correlation coefficient a i between nodes p j and p ij , which is used to compare the coefficient difference between nodes K is the number of neighboring nodes of node p i , and finally the normalized correlation coefficient is linearly combined with the corresponding feature through a Sigmoid activation function σ(·) to obtain the output feature of node p i after passing through a graph attention module:

[0082] Step S204: Use the gated recurrent module to capture the time dependence of the point cloud data to obtain point cloud time dependence features;

[0083] The input of the present gated recurrent unit is the output value via the atrous convolution module, the number of hidden units is set to 150, dropout is applied in the output layer with a dropout rate of 0.2, and no bias is set.

[0084] Step S205: input the point cloud spatial structure features and the point cloud time dependence features into a fully connected neural network model for training to obtain a trained neural network model;

[0085] Figure 7 is the confusion matrix of the proposed model, further, the cost function used in the radar point cloud recognition network training process is a cross-entropy loss function. The behavior recognition classification task based on the data set MMActivity is a five-classification problem of point cloud data, and the model uses a cross-entropy loss function to measure the probability error between five types of behavior sample data in the point cloud data and to determine the category of each sample. The calculation formula of the cross-entropy loss function is: where y j represents the jth true value, and p j represents the jth predicted value output by the network. After training the network, an optimal model is obtained, and finally the accuracy of the space-time feature fusion point cloud recognition network for recognizing the data set MMActivity can reach 94.33%.

[0086] The radar point cloud recognition method provided in the embodiment adopts various three-dimensional space coordinates, distance, speed azimuth angle and reflection intensity specific data, uses a graph structure to represent point cloud data, retains three-dimensional space position information of the data, uses a high-level feature extraction module composed of a hollow convolution to extract high-level features of each time sequence input of the point cloud data, reduces loss of feature information, uses a parallel network structure to extract point cloud space-time features, improves feature extraction efficiency and accuracy, realizes retention of three-dimensional space characteristics of the point cloud, reduces calculation complexity and improves feature extraction efficiency.

[0087] Please refer to Figure 8 , Figure 8 The structure block diagram of the radar point cloud recognition device provided in the embodiment of the application is provided; the specific device can include:

[0088] The training set generation module 100 generates a training data set based on different point cloud data;

[0089] The cloud map construction module 200 is used to construct a cloud map data set according to point cloud data in the training data set;

[0090] The feature extraction module 300 is used to extract high-level features F of a cloud map time sequence in the cloud map data set;

[0091] The graph attention network module 400 is used to extract point cloud space structure features S of the high-level features F;

[0092] The gated recurrent module 500 is used to capture time dependence of the high-level features F of the point cloud data, to obtain point cloud time dependence features T;

[0093] The training module 600 is used to input the point cloud space structure features S and the point cloud time dependence features T into a full connection neural network model for training, to obtain a trained neural network model;

[0094] The detection module 700 is used to detect to-be-recognized point cloud data by using the trained neural network model, to output a recognition result.

[0095] The radar point cloud recognition device of the embodiment is used to implement the radar point cloud recognition method, and the specific embodiments of the radar point cloud recognition device can be found in the embodiment part of the radar point cloud recognition method, for example, the training set generation module 100, the cloud map construction module 200, the feature extraction module 300, the graph attention network module 400, the gated recurrent module 500, the training module 600, and the detection module 700 are respectively used to implement steps S101, S102, S103, S104, S105, S106 and S107 in the radar point cloud recognition method, so the specific embodiments can refer to the description of the corresponding embodiment part, and details are not repeated here.

[0096] The embodiment of the present application also provides a radar point cloud recognition device, comprising: a millimeter wave radar, used to generate point cloud data; a memory, used to store a computer program; and a processor, used to implement the steps of the radar point cloud recognition method when the computer program is executed.

[0097] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is used to implement the steps of the radar point cloud recognition method when the computer program is executed by a processor.

[0098] In the specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the related parts can be referred to the method part.

[0099] The skilled person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or combination of the two. In order to clearly show the interchangeability of hardware and software, the composition and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0100] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, hard disk can be used for tangibly embodying the software module.

[0101] The radar point cloud recognition method, device, equipment and computer storage medium provided by the present application are described in detail above. The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method of the present application and its core idea. It should be pointed out that for ordinary skilled persons in the art, without departing from the principles of the present application, the present application can be improved and modified in several ways. These improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A radar point cloud recognition method, characterized in that: include: Generate training datasets based on different point cloud data; Constructing a cloud map dataset based on the point cloud data in the training dataset; Extracting high-level features F of the cloud image time series in the cloud image dataset using a feature extraction module; Extracting a point cloud spatial structure feature S of the high-level feature F, including: learning the degree of mutual influence between the target node and the adjacent nodes in the high-level feature F through a graph attention network, and assigning different weights to the target node and the adjacent nodes to obtain the point cloud spatial structure feature S; Capturing the temporal dependency of the high-level feature F point cloud data to obtain a point cloud temporal dependency feature T, including: using a gated recurrent unit to capture the temporal dependency of a time series with a large time step distance, and calculating the point cloud temporal dependency feature through a fully connected layer with a sigmoid activation function; Inputting the point cloud spatial structure feature S and the point cloud time-dependent feature T into a fully connected neural network model for training to obtain a trained neural network model; The trained neural network model is used to detect the point cloud data to be identified and output the identification result.

2. The radar point cloud recognition method according to claim 1, wherein: The step of constructing a cloud image dataset based on the point cloud data in the training dataset includes: Define a point cloud with N points as a set P = {p1, p2, ..., p n }, where p n =(X n ,S n ), n=1,2,…,N,X n Represents the three-dimensional spatial position information of the node, S n is the state vector of the node; Given a point cloud P, construct a cloud graph based on the points in the point cloud and the points with a fixed radius r around each point. Use P as a vertex and connect the adjacent points within the fixed radius r to obtain a cloud graph G = (P, E). where \(E = {(p i , p j ) | \| x i - x j \|^2 < r \}\), \(P\) represents the set of nodes, \(E\) represents the set of edges, \(p i \) is the target node, \(p j \) is the neighbor node, \(x i \) is the three-dimensional spatial position information of the target node, \(x j \) is the three-dimensional spatial position information of the neighbor node.

3. The radar point cloud recognition method according to claim 1, wherein: The feature extraction module consists of multiple dilated convolutional layers, each of which is cascaded with a PReLU activation function.

4. The radar point cloud recognition method according to claim 3, wherein: The dilation coefficient ε is introduced in the dilation convolution. Assuming that the original convolution kernel size is k, the receptive field size of the convolution kernel after adding d holes is k+(k-1)*d.

5. The radar point cloud recognition method according to claim 1, wherein: The graph attention network is given a target node p i and an adjacent node p j , the target node p i and the adjacent node p j The eigenvectors of h are i and h j , calculate the correlation coefficient e between two nodes ij =LeakyReLU(a T ·(h i ||h j )), where LeakyReLU is a nonlinear activation function, a is a learnable parameter vector, and || is a vector concatenation operation; Use the softmax function to perform normalization operation to obtain the target node p i and the adjacent node p j Normalized correlation coefficient between Among them, K is the target node p i The number of adjacent nodes, exp(e ik ) is an exponential function; The normalized correlation coefficient is linearly combined with the corresponding feature through the Sigmoid activation function σ(·) to obtain the target node p i The spatial structure features of the point cloud are output through the graph attention network.

6. A radar point cloud recognition device, characterized in that: include: Generate training set module, generate training data set based on different point cloud data; A cloud map construction module, configured to construct a cloud map dataset based on the point cloud data in the training dataset; A feature extraction module is used to extract high-level features F of the cloud image time series in the cloud image dataset; A graph attention network module is used to extract the point cloud spatial structure feature S of the high-level feature F, including: learning the degree of mutual influence between the target node and the adjacent nodes in the high-level feature F through the graph attention network, and assigning different weights to the target node and the adjacent nodes to obtain the point cloud spatial structure feature S; A gated recurrent module is used to capture the temporal dependency of the high-level feature F point cloud data to obtain a point cloud temporal dependency feature T, including: using a gated recurrent unit to capture the temporal dependency of a time series with a large time step distance, and calculating the point cloud temporal dependency feature through a fully connected layer with a sigmoid activation function; A training module, configured to input the point cloud spatial structure feature S and the point cloud time-dependent feature T into a fully connected neural network model for training, thereby obtaining a trained neural network model; The detection module is used to use the trained neural network model to detect the point cloud data to be identified and output the identification result.

7. A radar point cloud recognition device, characterized in that: include: Millimeter-wave radar, used to generate point cloud data; Memory for storing computer programs; A processor, configured to implement the steps of a radar point cloud recognition method as claimed in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the radar point cloud recognition method according to any one of claims 1 to 5 are implemented.

Images