A radar target detection method based on an inter-frame graph network model combined with attention mechanism

Abstract

The present invention relates to a method for detecting radar targets based on graph data and a graph attention network, and belongs to the field of radar signal processing. In order to solve the problem that the time-domain and spatial-domain relationship information between signal samples in existing methods is not fully utilized, resulting in limited performance, the method for detecting radar targets based on graph data and a graph attention network includes seven steps: obtaining scanning radar echo data, over-threshold detection, obtaining a detection sample set, setting an azimuth correlation threshold and a distance correlation threshold to obtain radar signal graph data, model training, and result display. This method fully utilizes the temporal and spatial correlation information between the target and the false alarm point, improves the target detection performance in complex environments, and enhances the accuracy of detection.

Description

Technical Field

[0001] The present invention relates to a radar target detection method using an inter-frame graph network model combined with an attention mechanism, and belongs to the field of radar signal processing. Background Art

[0002] Target detection has widespread applications in both military and civilian fields. Radar is a key means of target detection and surveillance. However, due to the clutter generated by complex environments and the diverse types of targets, reliable and robust target detection remains a key technology in need of research. In recent years, graph theory has rapidly developed in the field of signal processing. Representing signals using graphs preserves signal characteristics and also has the ability to represent data associations, providing new approaches to signal processing. In the field of radar signal processing, graph-theoretic methods have been applied to processing signals such as SAR images.

[0003] Previous researchers have attempted to use various convolutional neural networks to process radar signal time series, distinguishing targets from clutter signals through binary classification to achieve target detection. These methods rely solely on signal sequence characteristics for classification. During training and testing, each signal sample is completely independent of each other, and the temporal and spatial relationship information between signal samples is not fully utilized, limiting performance. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem mentioned in the above background technology that the time domain and spatial domain relationship information between each signal sample in the existing method is not fully utilized and the performance is limited, and to provide a detection method with high radar template detection accuracy. The method is based on graph data and graph attention network scanning radar targets.

[0005] In order to solve the above problems, the detection method of scanning radar targets based on graph data and graph attention network of the present invention is implemented through the following technical solutions:

[0006] A method for detecting radar targets based on graph data and graph attention network scanning is characterized by the following steps:

[0007] Step 100: Acquire a scanning radar echo signal;

[0008] The radar echo signal adopts continuous multi-frame clutter and target radar echo signal, signal S∈C AZ×R×L ;

[0009] Where: AZ is the number of azimuth units, R is the number of range units, L is the number of scanning circles, and C is the complex domain symbol;

[0010] Step 200: Threshold crossing detection: performing threshold crossing detection on each scanning cycle of the radar echo signal. A detection threshold is independently set for the signal of each scanning cycle to obtain multiple binary detection result graphs. Each scanning cycle signal corresponds to a binary detection result graph. The value of the distance-angle unit that crosses the threshold is 1, and the corresponding value of other units is 0.

[0011] S' ijk ∈{0,1}

[0012] Where S' is the binary detection result image, S' is a three-dimensional data, i, j and k are indexes corresponding to the orientation unit number, distance unit number and scanning frame number of each data;

[0013] Step 300: Obtain a test sample set: Perform 4 or 8 connected domain detection on each binary test result graph to obtain t connected domains.

[0014] t i Indicates the number of connected domains in the binary detection result graph corresponding to the i-th scanning cycle signal;

[0015] Number each connected domain, and count the scanning cycle number corresponding to the binary detection result graph of each connected domain as the frame number of the connected domain. i ; The number of angle-distance units contained in each connected domain is taken as the area c of the connected domain i ;

[0016] Calculate the average orientation, average distance, frame number and number of angle-distance units contained in each connected domain to obtain the detection sample set;

[0017] Calculate the average orientation az of each connected region i : Extract the orientation unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average orientation of the connected domain: Where az ij is the azimuth number of the jth angle-distance unit in the i-th node;

[0018] Calculate the average observation distance r of each connected region i : Extract the distance unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average observed distance of the connected domain:

[0019] Where r ij is the azimuth number of the jth angle-distance unit in the i-th node;

[0020] Get the detection sample set V = {az, r, l, c};

[0021] Each connected domain corresponds to a target / false alarm sample, corresponding to a node, and the node feature is the average orientation az of the corresponding connected domain. i , average distance r i , frame number l i and area c i ;

[0022] Step 400: Setting the position correlation threshold thr az And the distance association threshold thr range ;

[0023] Position correlation threshold thr az And the distance association threshold thr range It is an adjustable variable that is adjusted according to information such as the application scenario;

[0024] Step 500: Obtain radar signal image data:

[0025] According to the orientation correlation threshold thr az , distance association threshold thr range , the feature T of each node i , constructing an adjacency matrix to obtain radar signal graph data, wherein the graph data includes the adjacency matrix and node features of each node;

[0026] Specifically, the weight of the edge between any two nodes is determined using the following formula:

[0027]

[0028] Where V i,2 is the average orientation of the i-th node, V j,2 is the average orientation of the jth node; V i,3 is the average distance of the i-th node, V j,3 is the average distance of the jth node; V i,4 is the frame number of the i-th node, V j,4 is the frame number of the jth node; thr az and thr range is the preset variable;

[0029] The characteristics of each node are determined using the following formula:

[0030]

[0031] Where Δ az and Δ range These are preset variables that control the feature size, set according to the distance and azimuth unit size of the radar signal;

[0032] Then, based on the edge weights between all nodes, each node is numbered sequentially and an adjacency matrix is ​​constructed. For example, the adjacency matrix A ij is the edge weight between the i-th node and the j-th node. Finally, based on the adjacency matrix, the node characteristics of each node, the average orientation and average distance of each node, the signal graph data can be obtained;

[0033] Step 600: Model training: inputting the signal graph data into the graph attention network to train the model;

[0034] Specifically, the graph attention network consists of a graph attention layer, a fully connected layer, and an output layer. The operation of the graph attention layer is:

[0035]

[0036] Where h i ' is the feature of node i in the output of the graph attention layer, h j is the feature of node j in the input, and T in the first i ,σ is the activation function,α ij is the attention coefficient of node j to node i, W is the trainable parameter of this layer, and the attention coefficient α ij The calculation method is as follows:

[0037]

[0038] Where N is N i The number of nodes, g(i,j,k), is calculated as follows:

[0039]

[0040] Step 700: Result display: Input the graph attention network output vector into the classifier, and output the classification result of each node; the classification result of each node is a target or a false alarm.

[0041] This method makes full use of the temporal and spatial correlation information between the target and the false alarm point, improves the target detection performance in complex environments, increases the detection accuracy, and can effectively suppress false alarms, especially in areas far away from the real target, where the false alarm suppression effect is more obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] For ease of explanation, the present invention is described in detail with reference to the following specific implementations and accompanying drawings.

[0043] Figure 1 : Algorithm flow chart of this application;

[0044] Figure 2 : Variable flow chart of this application;

[0045] Figure 3 : is the radar signal;

[0046] Figure 4 : is the threshold point diagram;

[0047] Figure 5 : Target location map for detection. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Example 1

[0050] like Figure 1-2 The method for detecting radar targets based on graph data and graph attention network includes the following steps:

[0051] Step 100: Acquire a scanning radar echo signal;

[0052] The radar echo signal in this embodiment is the echo data after pulse compression; the echo signal is obtained by a scanning radar, and the azimuth-range-scanning frame three-dimensional data is obtained after pulse compression;

[0053] Using continuous multi-frame clutter and target radar echo signals, signal S∈C AZ×R×L , AZ is the number of azimuth units, R is the number of range units, L is the number of scanning circles, and C is the complex domain symbol; for example, in the selected radar signal, AZ = 500, R = 100, and L = 20.

[0054] Step 200: Threshold crossing detection: performing threshold crossing detection on each scanning cycle of the radar echo signal. A detection threshold is independently set for the signal of each scanning cycle to obtain multiple binary detection result graphs. Each scanning cycle signal corresponds to a binary detection result graph. The value of the distance-angle unit that crosses the threshold is 1, and the corresponding value of other units is 0.

[0055] S' ijk ∈{0,1}

[0056] Where S' is the binary detection result image,

[0057] S' is a three-dimensional data, i, j and k are indexes, corresponding to the orientation unit number, distance unit number and scanning frame number of each data;

[0058] Step 300: Obtain the detection sample set: perform 8-connected domain detection on each binary detection result graph to obtain t connected domains.

[0059] t i Indicates the number of connected domains in the binary detection result graph corresponding to the i-th scanning cycle signal;

[0060] Number each connected domain, and count the scanning cycle number corresponding to the binary detection result graph of each connected domain as the frame number of the connected domain. i ; The number of angle-distance units contained in each connected domain is taken as the area c of the connected domain i ;

[0061] Calculate the average orientation, average distance, frame number and number of angle-distance units contained in each connected domain to obtain the detection sample set;

[0062] Calculate the average orientation az of each connected region i : Extract the orientation unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average orientation of the connected domain.

[0063]

[0064] Calculate the average observation distance r of each connected region i : Extract the distance unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average observation distance of the connected domain;

[0065]

[0066] Get the detection sample set V = {az, r, l, c};

[0067] Each connected domain corresponds to a target / false alarm sample, corresponding to a node, and the node feature is the average orientation az of the corresponding connected domain. i , average distance r i , frame number l i and area c i ;

[0068] Step 400: Setting the position correlation threshold thr az And the distance association threshold thr range ;

[0069] Position correlation threshold thr az And the distance association threshold thr range It is an adjustable variable, which is adjusted according to the application scenario and other information; for example, setting thr az =10,thr range =10;

[0070] Step 500: Obtain radar signal image data:

[0071] According to the orientation correlation threshold thr az , distance association threshold thr range , the feature T of each node i , constructing an adjacency matrix to obtain radar signal graph data; the graph data includes the adjacency matrix and node features of each node;

[0072] Specifically, the weight of the edge between any two nodes is determined using the following formula:

[0073]

[0074] Where V i,2 is the average orientation of the i-th node, V j,2 is the average orientation of the jth node; V i,3 is the average distance of the i-th node, V j,3 is the average distance of the jth node; V i,4 is the frame number of the i-th node, V j,4 is the frame number of the jth node; thr az and thr range is the preset variable;

[0075] Specifically, the following formula is used to determine the characteristics of each node:

[0076]

[0077] Where Δ az and Δ range are preset variables that control the feature size; for example, setting Δ az =5,Δ range =5;

[0078] Then, based on the edge weights between all nodes, each node is numbered sequentially and an adjacency matrix is ​​constructed. For example, the adjacency matrix A ij is the edge weight between the i-th node and the j-th node. Finally, according to the adjacency matrix and the node characteristics of each node, the signal graph data can be obtained;

[0079] Step 600: Model training:

[0080] Inputting the signal graph data into the graph attention network to train the model;

[0081] Specifically, the graph attention network consists of a graph attention layer, a fully connected layer, and an output layer. The operation of the graph attention layer is:

[0082]

[0083] Where h i ' is the feature of node i in the output of the graph attention layer, h j is the feature of node j in the input, and T in the first i ,σ is the activation function,α ij is the attention coefficient of node j to node i, and W is the trainable parameter of this layer. ij The calculation method is as follows:

[0084]

[0085] Where N is N i The number of nodes, g(i,j,k), is calculated as follows:

[0086]

[0087] Step 700: Result display: Input the graph attention network output vector into the classifier, and output the classification result of each node; the classification result of each node is a target or a false alarm.

[0088] Example 2

[0089] The detection method of scanning radar targets based on graph data and graph attention network in this embodiment is basically the same as that in embodiment 1, except that: in step 300, a 4-connected domain detection is performed on each binary detection result graph to obtain t connected domains.

[0090] t i Indicates the number of connected domains in the binary detection result graph corresponding to the i-th scanning cycle signal;

[0091] Number each connected domain, and count the scanning cycle number corresponding to the binary detection result graph of each connected domain as the frame number of the connected domain. i ; The number of angle-distance units contained in each connected domain is taken as the area c of the connected domain i ;

[0092] Calculate the average orientation, average distance, frame number and number of angle-distance units contained in each connected domain to obtain the detection sample set;

[0093] Calculate the average orientation az of each connected region i : Extract the orientation unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average orientation of the connected domain.

[0094]

[0095] Calculate the average observation distance r of each connected region i : Extract the distance unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average observation distance of the connected domain;

[0096]

[0097] Get the detection sample set V = {az, r, l, c};

[0098] Each connected domain corresponds to a target / false alarm sample, corresponding to a node, and the node feature is the average orientation az of the corresponding connected domain. i , average distance r i , frame number l i and area c i .

[0099] As attached Figure 3-5 As shown, Figure 3 The original radar signal consists of 100 range units and 500 azimuth units. Figure 4 This is the result after threshold detection. The white area is the target and a large number of false alarms, among which the white area in the white circle is the real target. Figure 5 This is the detection result of the proposed method. The white areas are the target and false alarms, and the white area in the white circle is the real target. This method can effectively suppress false alarms, especially in areas far away from the target, where the false alarm suppression effect is more obvious.

Claims

1. A method for detecting radar targets based on graph data and graph attention network, characterized by: The following steps are involved: Step 100: Acquire a scanning radar echo signal; Step 200: Threshold crossing detection: performing threshold crossing detection on each scanning cycle of the radar echo signal, independently setting a detection threshold for the signal of each scanning cycle, and obtaining a plurality of binary detection result images; Step 300: Obtain the detection sample set: perform connected domain detection on each binary detection result graph, number each connected domain, Calculate the average orientation, average distance, frame number and number of angle-distance units contained in each connected domain to obtain the detection sample set; Step 400: Setting the position correlation threshold thr az And the distance association threshold thr range ; Step 500: Obtain radar signal image data: According to the orientation correlation threshold thr az , distance association threshold thr range , the feature T of each node i , construct the adjacency matrix and obtain the radar signal graph data; Step 600: Model training: Inputting the signal graph data into the graph attention network to train the model; Step 700: Result display: The output vector of the graph attention network is input into the classifier, and the classification result of each node is output; the classification result of each node is a target or a false alarm.

2. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 1, characterized in that: In step 100, continuous multi-frame clutter and target radar echo signals are used, and the signal S∈C AZ×R×L , AZ is the number of azimuth units, R is the number of distance units, and L is the number of scanning circles.

3. The method for detecting radar targets based on graph data and a graph attention network according to claim 1 or 2, characterized in that: In step 200: each scanning cycle signal corresponds to a binary detection result image, the distance-angle unit value that exceeds the threshold is 1, and the corresponding value of other units is 0; S’ ijk ∈{0,1}。 4. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 3 is characterized by: In step 300, a 4- or 8-connected domain detection is performed on each binary detection result graph to obtain t connected domains. t i Indicates the number of connected domains in the binary detection result graph corresponding to the i-th scanning cycle signal; Number each connected domain, and count the scanning cycle number corresponding to the binary detection result graph of each connected domain as the frame number of the connected domain. i ; The number of angle-distance units contained in each connected domain is taken as the area c of the connected domain i .

5. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 4, characterized in that: In step 300: Calculate the average orientation az of each connected region i : Extract the orientation unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average orientation of the connected domain; 6. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 4, characterized in that: In step 300: Calculate the average observation distance r of each connected region i : Extract the distance unit numbers of all angle-distance units in each connected domain and calculate the weighted average as the average observation distance of the connected domain; The detection sample set V = {az, r, l, c} is obtained.

7. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 4, characterized in that: In step 300: Each connected domain corresponds to a target / false alarm sample, corresponding to a node, and the node feature is the average orientation az of the corresponding connected domain. i , average distance r i , frame number l i and area c i .

8. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 4, characterized in that: In step 500: The graph data includes an adjacency matrix and node features of each node; Specifically, the weight of the edge between any two nodes is determined using the following formula: Where V i,2 is the average orientation of the i-th node, V j,2 is the average orientation of the jth node; V i,3 is the average distance of the i-th node, V j,3 is the average distance of the jth node; V i,4 is the frame number of the i-th node, V j,4 is the frame number of the jth node; thr az and thr range is the preset variable; Then, based on the edge weights between all nodes, each node is numbered sequentially and an adjacency matrix is ​​constructed. For example, the adjacency matrix A ij is the edge weight between the i-th node and the j-th node; finally, according to the adjacency matrix, the node features of each node, the node features of each node, the average orientation and average distance of each node, the signal graph data can be obtained.

9. The method for detecting radar targets based on graph data and graph attention network scanning according to claim 4, characterized in that: In step 600: The graph attention network consists of a graph attention layer, a fully connected layer, and an output layer. In the graph attention layer, the attention coefficient is obtained by the correlation between the node's orientation az, distance r, and frame l.

Images