Multi-frame joint detection method based on graph neural network link prediction task

Abstract

The invention provides a multi-frame joint detection method based on a graph neural network link prediction task, and the method comprises the steps: carrying out the correlation test of observation data passing a first threshold through the priori knowledge of target physical motion, constructing a trace point correlation graph, and carrying out the detection of a multi-feature link prediction network based on the graph neural network design. The multi-feature link prediction network integrates the echo structure of single-frame observation, Doppler, space-time coupling of multi-frame observation and other features, and completes a multi-frame detection task. According to the method, a multi-frame detection problem is converted into link prediction in a graph, track search and track judgment are integrally completed, the target track can be directly observed and output from multiple frames to simplify processing logic, and track judgment can be more accurately carried out while track search performance loss is reduced. Compared with a traditional single-frame and multi-frame detection method, the method has the advantages that false trace points can be effectively suppressed, and meanwhile, the detection performance of the target is improved.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a multi-frame joint detection method based on a graph neural network link prediction task, which can be used for long-distance weak target detection. Background Art

[0002] Target detection aims to determine whether the given echo signal strength originates from the target. Effective target detection is the basis for subsequent operations such as target tracking and target identification. In the traditional radar signal processing process, detection and tracking are two independent steps. The sequential processing method of first detecting and then tracking is particularly disadvantageous when dealing with low signal-to-noise ratio targets. In order to successfully detect low signal-to-noise ratio targets, it is usually necessary to lower the threshold. However, lowering the threshold will bring a large number of false alarms that are difficult for the system to handle. On the other hand, it will increase the possibility of misassociation between the target and the false alarm, making it difficult to correctly start the target track.

[0003] In order to deal with the defects of traditional signal processing flow, multi-frame detection MFD is considered to be a practical method for detecting weak targets. It improves the detection performance of low signal-to-noise ratio targets by fusing continuous multi-frame echo data. However, the traditional multi-frame detection method has obvious drawbacks in high-resolution detection scenarios with a large number of detection units. On the one hand, a large number of original observations will bring a large computational burden to the system, and on the other hand, the observation uncertainty generated by too many false points will make it difficult to correctly associate the target. An effective countermeasure is the multi-frame detection method of searching first and then starting. This method uses the physical motion constraints of the target to search for potential tracks for multi-frame observations through a low threshold; finally, a fusion decision is made along the potential track based on information such as echo energy. However, the track search and track decision of this method are two sequential steps. This series connection method makes its final multi-frame detection performance limited by the target's potential track search performance. The model-driven track search method it adopts is difficult to use unified constraint parameters to make targets with different motion characteristics better associated, which leads to more missed associations of targets during track search. In addition, this type of method does not fully utilize the echo information and the spatial-temporal coupling characteristics of inter-frame traces, which is not conducive to multi-frame joint detection. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a multi-frame joint detection method based on graph neural network link prediction task. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0005] A multi-frame joint detection method based on graph neural network link prediction task, comprising:

[0006] S100, acquiring multiple frames of continuous raw echo data, and setting a first threshold for the raw echo data to perform detection, to obtain multiple frames of continuous observation data; the observation data includes measurement information of multiple point traces;

[0007] S200, for each point trace in the observation data of the current frame, using the maximum speed constraint, perform an association test on the point trace with the point trace in the observation data of the subsequent frame of the current frame; connect the point traces that pass the association test with edges, and use the point traces as nodes to obtain a point trace association graph;

[0008] S300, using NFEN to perform feature mining on each node in the point trace association graph to obtain a multi-dimensional extraction feature, and then obtain a multi-dimensional fusion feature of each node, and the multi-dimensional fusion features of all nodes are combined into a point trace extraction feature matrix;

[0009] S400, using STEF-GAN to mine multi-frame space-time coupling features in the point trace association graph, and in the mining process, using the point trace extraction feature matrix to perform message transmission to obtain the final feature of each node;

[0010] S500, for any two nodes with a common edge, use OAJN to fuse the final features of the two nodes to calculate the classification score of the common edge between the two nodes, and remove the common edges that do not meet the classification score of the common edge, and retain the common edges between the targets.

[0011] Beneficial effects:

[0012] The present invention provides a multi-frame joint detection method GLP-MFD based on the graph neural network link prediction task. It uses the prior knowledge of the physical movement of the target to perform an association test on the observation data that passes the first threshold to construct a point track association graph. Then, a multi-feature link prediction network MFLPN designed based on the graph neural network is designed. The MFLPN integrates the echo structure, Doppler, and space-time coupling of multi-frame observations of single-frame observations to complete the multi-frame detection task. The present invention converts the multi-frame detection problem into link prediction in the graph, and completes the track search and track judgment in an integrated manner, so as to simplify the processing logic by being able to directly output the target track from the multi-frame observation, which is conducive to more accurate track judgment while reducing the loss of track search performance. Compared with traditional single-frame and multi-frame detection methods, the present invention can improve the detection performance of the target while effectively suppressing false point tracks.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of a multi-frame joint detection method based on a graph neural network link prediction task provided by the present invention;

[0015] Figure 2 The present invention provides Figure 1 Schematic diagram of the curve of the relationship between detection probability and sliding window number when the signal-to-noise ratio is 9dB;

[0016] Figure 3 It is a curve diagram of the relationship between the detection probability and the number of sliding windows when the signal-to-noise ratio is 12 dB provided by the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0018] like Figure 1 As shown, the present invention provides a multi-frame joint detection method GLP-MFD based on a graph neural network link prediction task, including:

[0019] S100, acquiring multiple frames of continuous raw echo data, and setting a first threshold for the raw echo data for detection, to obtain multiple frames of continuous observation data; the observation data includes measurement information of multiple points; the first threshold can be set in advance.

[0020] In a complex Gaussian white noise scene containing a target, the coherent radar system can improve the signal-to-noise ratio of the target echo through pulse compression, non-coherent accumulation and other steps, and obtain Doppler measurement information at the same time. After radar target detection, the measurement information of the kth point in the nth frame can be expressed as follows:

[0021]

[0022] In the formula, t k,n Indicates the time when the point is detected; r k,n Indicates the measured radial distance; θ k,n Indicates the measured azimuth; v k,n Indicates that there may be ambiguity in the measured radial velocity; d k,n Indicates the Doppler channel where the point is located; s k,n Indicates the signal-to-noise ratio of the trace; i k,n Indicates the time sequence coding of the trace; a k,n The range Doppler diagram of the points; superscript Represents a transpose operation;

[0023] In the nth frame, there are a total of D n The points are represented as:

[0024]

[0025] S200. For each trace in the observation data of the current frame, use the maximum speed constraint to perform an association test between this trace and the traces in the observation data of the subsequent frames of the current frame; connect the traces that pass the association test with edges, and use the traces as nodes to obtain a trace association graph;

[0026] There is a certain connection in the spatio-temporal characteristics between the target multi-frame data. The prior knowledge of the target's physical movement can be used to perform an association test on the multi-frame observations that pass the low threshold. Using the traces as nodes and connecting the successfully associated nodes with edges to form an undirected graph, which is called the trace association graph in the present invention. The correct target track exists in the trace association graph. The present invention needs to complete the integrated track search and track decision-making process from the trace association graph. The present invention constructs the trace association graph using the maximum speed constraint.

[0027] For two traces (t1, r1, θ1, v1) and (t2, r2, θ2, v2) of different frames, where t1 < t2, the maximum speed constraint is expressed by the formula:

[0028]

[0029] In the formula, the two traces are respectively expressed as (t1, r1, θ1, v1) and (t2, r2, θ2, v2), v max represents the maximum movement speed of the target of interest, t1 < t2, r1, θ1, v1 respectively represent the measured radial distance, measured azimuth angle, and possibly ambiguous measured radial speed of trace 1 in the current frame; r2, θ2, v2 respectively represent the measured radial distance, measured azimuth angle, and possibly ambiguous measured radial speed of trace 2 in the subsequent frame.

[0030] The task of multi-frame target detection is to judge whether there is a target according to the multi-frame observations In the multi-frame observation data associate two traces that satisfy the maximum speed constraint, where L represents the number of frames for joint processing.

[0031] S300. Use NFEN to perform feature mining on each node in the trace association graph to obtain multi-dimensional extraction features, and then obtain the multi-dimensional fusion features of each node, and form a trace extraction feature matrix with the multi-dimensional fusion features of all nodes;

[0032] The NFEN designed in the present invention for d k,n , s k,n , i k,n , a k,nThe whole network is used to mine features, so that it can better learn the mapping from multi-dimensional attribute information to the track-derived target confidence. It includes three fully connected neural networks and one convolutional neural network. The fully connected layer is usually used for feature extraction and feature dimension transformation in NFEN. The convolutional structure not only reduces the number of parameters through parameter sharing, but also improves the learning ability of NFEN for local similarity information, thereby more effectively mining point track features. It represents the feature vector describing the point Doppler channel after feature mining. Represents the feature vector describing the signal-to-noise ratio of the point trace after feature mining, Represents the feature vector that describes the time series information of the trace after feature mining. Represents the feature vector that describes the structure information of the trace signal after feature mining. k,n Represents the multi-dimensional fusion features that integrate different attribute information.

[0033] In a specific implementation of the present invention, S300 includes:

[0034] S310, using NFEN to perform feature mining on each node in the point trace association graph to obtain multi-dimensional extraction features, wherein the multi-dimensional extraction features include: a feature vector describing the point trace Doppler channel Characteristic vector describing the signal-to-noise ratio of a point trace Characteristic vector describing the timing information of the trace and the feature vector describing the structure information of the trace signal

[0035] In a specific implementation of the present invention, S310 includes:

[0036] S311, using three fully connected neural networks of the NFEN module to calculate the d of each node in the point trace association graph k,n 、s k,n 、i k,n Perform feature mining to obtain the feature vector describing the point Doppler channel Characteristic vector describing the signal-to-noise ratio of a point trace and the feature vector describing the timing information of the trace in,

[0037]

[0038] In the formula, Represents the operations between layers in the NFEN module, Represents a function composite operation, N d ,N s ,N i represents the number of fully connected network layers, Represents the learnable parameters of each layer of the network, φ D Represents the nonlinear activation function in the NFEN module, which can enhance the NFEN's ability to learn nonlinear mapping.

[0039] S312, using the convolutional neural network of NFEN to calculate the a of each node in the point trace association graph k,n Perform feature mining to obtain feature vectors that describe the structure information of the point trace signal It is expressed as:

[0040]

[0041] Where N p Indicates the number of convolutional layers, * indicates the convolution operation, Represents the convolution kernel used in each layer of convolution operation.

[0042] S320, for each node, the multi-dimensional extraction features of the node are fused to obtain a multi-dimensional fusion feature, and the multi-dimensional fusion features of all nodes are combined into a point trace extraction feature matrix.

[0043] In a specific implementation of the present invention, S320 includes:

[0044] S321, will Splicing is performed to obtain the multi-dimensional fusion feature f of each node k,n , expressed as:

[0045]

[0046] S322, multi-dimensional fusion features f of all nodes k,n Rewrite and form the point trace extraction feature matrix f.

[0047] In a specific implementation manner of the present invention, S322 includes:

[0048] S3221, rewrite the multidimensional fusion feature of each point trace according to the measurement time sequence of the point traces in multiple frames to obtain the rewritten multidimensional fusion feature, which is expressed as:

[0049]

[0050] Where L represents the total number of frames of observation data, l represents the lth frame, and there are D l A dot trace, Represents the rewritten multi-dimensional fusion features of the lth frame;

[0051] S3222, rewrite the multi-dimensional fusion features to form a point trace extraction feature matrix f.

[0052] S400, using STEF-GAN to mine multi-frame space-time coupling features in the point trace association graph, and in the mining process, using the point trace extraction feature matrix to perform message transmission to obtain the final feature of each node;

[0053] In a specific implementation of the present invention, S400 includes:

[0054] S410, using the point trace extraction feature matrix f as the input node feature matrix x in STEF-GAN (0) , so that each layer of STEF-GAN transmits messages according to the message passing formula, and in the transmission process, the multi-frame space-time coupling features of the point trace association graph are mined to obtain the final features of each point trace; wherein the message passing formula is expressed as:

[0055]

[0056] In the formula, k represents the number of layers; represents the feature of node i in the kth layer, the input node feature matrix x of STEF-GAN (0) That is, the point trace extraction feature matrix f output by the NFEN network; represents the set of neighbor nodes of node i; e j,i Represents the characteristics of the edge between node i and node j; A learnable matrix representing the features of nodes in the kth layer; Represents the learnable matrix of the k-th layer edge features, φ G Represents the nonlinear activation function in the GNN network, such as relu, sigmoid, softmax; represents a fully connected layer; ⊕ represents aggregation operations, such as addition, maximum value, and average value, which can aggregate the neighbor node information of node i into itself; || represents feature merging; represents the attention coefficient of the kth layer; Represents the learnable parameters in the attention mechanism; represents the learnable matrix of the edge in the attention mechanism, and The difference is, Focus on learning the influence of edge features in the attention mechanism, It is used for feature enhancement of edges, focusing on the impact in aggregation operations. It is this message passing mechanism that enables STEF-GAN to fuse multi-frame structural information, mine the complex motion law characteristics between nodes, and thus increase the separability of targets and false alarms.

[0057] There is a coupling relationship between the measurement time, measurement distance, measurement angle, and measurement speed of the target multi-frame traces with certain physical motion laws, but this coupling relationship is usually more complicated. k,n,r k,n ,θ k,n ,v k,n Directly inputting the node features into the network will increase the difficulty of network training and the computational burden. Based on this consideration, the t-values ​​between multi-frame point traces can be pre-constructed based on the target domain knowledge. k,n ,r k,n ,θ k,n ,v k,n After obtaining the feature input of nodes and edges, STEF-GAN mines multi-frame information through multi-layer message passing to obtain the final feature, which can be expressed as:

[0058]

[0059] Where G· represents the inter-layer operation in STEF-GAN; Indicates function composite operation; N G represents the number of STEF-GAN network layers; e represents the edge feature matrix of the input STEF-GAN network.

[0060] After deeply exploring the multi-dimensional attribute characteristics of point traces and the spatial-temporal coupling characteristics between multi-frame point traces, the present invention uses OAJN to complete intelligent track judgment in the point trace association graph. Compared with the traditional judgment method using energy accumulation threshold, this method can learn the mapping from multi-dimensional measurement information to whether the point trace association originates from the target, and better evaluate the track quality.

[0061] S500, for any two nodes with a common edge, use OAJN to fuse the final features of the two nodes to calculate the classification score of the common edge between the two nodes, and remove the common edges that do not meet the classification score of the common edge, and retain the common edges between the targets.

[0062] In a specific implementation of the present invention, S500 includes:

[0063] S510, for any two nodes having a common edge, using OAJN to fuse the final features of the two nodes, and calculating the classification score of the common edge between the two nodes;

[0064] Considering the edge feature e j,i It reflects the space-time coupling relationship between the traces. In order to make full use of this feature, the present invention will j,i Input it into OAJN again. In addition, the Doppler channel difference of the target is an important representation of the target motion characteristics. For example, for a target moving at a constant speed, the Doppler channel difference between different frames is usually small. The present invention calculates the Doppler channel difference d between node j and node i. j,i =|d j-d i | is also input into OAJN to better learn the classification prediction of edge attributes. In order to facilitate the subsequent setting of thresholds to distinguish target tracks from false tracks, the present invention selects the softamx function as the nonlinear activation during decoding. The edges in the point trace association graph are divided into three categories: edges between false alarms, edges between false alarms and targets, and edges between targets. Their true label values ​​are represented as For any two nodes with edges output by STEF-GAN and The score of the common edge between any two nodes is expressed as:

[0065]

[0066] In the formula, and Represents any two nodes, Indicates the scores of the edges between two nodes predicted as false alarm-false alarm edges, false alarm-target edges, and target-target edges, respectively. is the learnable parameter in OAJN; φ M represents the nonlinear activation function in OAJN, d j,i Represents the Doppler channel difference between node i and node j.

[0067] S520, dividing the common edge into edges between false alarms and false alarms, edges between false alarms and targets, and edges between targets and targets according to the classification score of the common edge and the second threshold;

[0068] S530, removing the edges between false alarms and the edges between false alarms and targets as inconsistent common edges, and retaining the common edges between targets.

[0069] In summary, the multi-frame joint detection method GLP-MFD based on the graph neural network link prediction task provided by the present invention uses the multi-feature link prediction MFLPN network to perform integrated track search and judgment after constructing a point track association graph for observation data of multiple consecutive frames passing a low threshold. The multi-feature link prediction network MFLPN includes three modules: a node feature embedding network NFEN, an attention mechanism graph neural network STEF-GAN with space-time coupling edge features, and a point track association judgment network OAJN. NFEN extracts deep features of point track echo structure, Doppler and other information; STEF-GAN fully mines the association features containing complex physical motion laws between target point tracks; OAJN completes target detection through link prediction in the point track association graph based on the mined features. The present invention converts the multi-frame detection problem into link prediction in the graph, completes track search and judgment in an integrated manner, simplifies the processing logic by being able to directly output the target track from multi-frame observations, and is conducive to more accurate track judgment while reducing the loss of track search performance. Compared with traditional single-frame and multi-frame detection methods, the present invention can improve the detection performance of the target while effectively suppressing false points.

[0070] The advantages of the present invention are reflected below through experimental contents and results.

[0071] In order to demonstrate the improvement in detection performance brought about by the integrated multi-frame detection using the proposed GLP-MFD method, the present invention considers conducting simulation experiments in 2D scenes, but the proposed method can be easily extended to 3D scenes. The simulated radar parameters are set as follows: the radar operates in the L band, the beam width is 2°, the distance resolution is 100m, and the radar system transmits 32 pulses at each wave position. As mentioned above, the first threshold determines the optimal detection performance that can be achieved by the proposed method. Therefore, in order to ensure that the radar system has better detection performance for low signal-to-noise ratio targets without introducing too many false points that make the system calculation burden too heavy, the present invention sets the first threshold according to the corresponding false alarm probability P fa,1 =10 -3 set up.

[0072] In order to highlight the superiority of the proposed GLP-MFD method, the present invention uses the traditional MFD method NDP of searching first and then starting, and a model-driven MFD method MTS-NCI combining track initiation and incoherent accumulation for comparison. The NDP method uses the target position measurement to construct kinematic constraints to obtain potential tracks, and fuses the track energy through the incoherent accumulation method to compare with the second threshold for track judgment. The MTS-NCI method makes full use of Doppler information, constructs motion constraints based on the coupling relationship between position measurement and Doppler measurement to obtain potential tracks. By introducing Doppler information, the separability of target tracks and false tracks is increased, and then the track energy is incoherently accumulated and compared with the second threshold for track judgment. In order to make a fair comparison, the above methods all use the same window length for processing, and the false alarm probability after processing is guaranteed to be P for all methods. fa,2 =10 -6 .

[0073] The relationship curves between the target detection probability and the number of sliding windows for the above methods with signal-to-noise ratios of 9dB and 12dB are shown in the figure below. Figure 2 , Figure 3 As shown. Among them, P D,CA-CFAR Indicates that the unit average constant false alarm method is used in a single frame with a false alarm probability of 10 -6 The target detection probability under the condition P D,UPB Represents the upper limit of the detection performance of the proposed method and the comparison method. It can be seen that with the increase of the number of sliding windows, the detection probability of the multi-frame detection method gradually increases. In most cases, the multi-frame detection method can achieve better target detection performance than the single-frame detection method. Since the NDP method does not use Doppler information, the separability of targets and false alarms is poor. At the same time, its exclusion of false tracks mainly relies on energy accumulation, which makes its single sliding window performance weaker than the traditional unit average constant false alarm method at low signal-to-noise ratio. Under high signal-to-noise ratio conditions, its performance is higher than the unit average constant false alarm method. The MTS-NCI method introduces Doppler information, which has better separability between targets and false alarms, and its performance is better than the NDP method. However, weak targets will reduce the separability of targets and false alarms, and setting the false alarm probability P fa,1 =10 -3 This results in more false alarms passing the first threshold, leading to a performance bottleneck. The proposed GLP-MFD method fuses the multi-dimensional information of target points and the space-time coupling information between frames through a multi-feature link prediction network, which is closer to the optimal target detector that integrates multi-dimensional extraction features and has the highest detection performance. Moreover, at low signal-to-noise ratios, its superiority is more obvious compared to other methods, indicating that the proposed GLP-MFD method can effectively suppress false points while having good detection performance for weak targets.

[0074] It is worth noting that the terms "first" and "second" in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0075] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0076] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A multi-frame joint detection method based on graph neural network link prediction task, characterized in that: include: S100, acquiring multiple frames of continuous raw echo data, and setting a first threshold for the raw echo data for detection, to obtain multiple frames of continuous observation data; The observation data includes measurement information of multiple points; S200, for each point trace in the observation data of the current frame, using the maximum speed constraint, perform an association test on the point trace with the point trace in the observation data of the subsequent frame of the current frame; connect the point traces that pass the association test with edges, and use the point traces as nodes to obtain a point trace association graph; S300, using NFEN to perform feature mining on each node in the point trace association graph to obtain a multi-dimensional extraction feature, and then obtain a multi-dimensional fusion feature of each node, and the multi-dimensional fusion features of all nodes form a point trace extraction feature matrix; S400, using STEF-GAN to mine multi-frame space-time coupling features in the point trace association graph, and in the mining process, using the point trace extraction feature matrix to perform message transmission to obtain the final feature of each node; S500, for any two nodes with a common edge, use OAJN to fuse the final features of the two nodes to calculate the classification score of the common edge between the two nodes, and remove the common edges that do not meet the classification score of the common edge, and retain the common edges between the targets.

2. The multi-frame joint detection method based on graph neural network link prediction task according to claim 1 is characterized in that: The measurement information of the kth point trace in the nth frame in S110 is expressed as: Where, t k,n Indicates the time when the point is detected; r k,n Indicates the measured radial distance; θ k,n Indicates the measured azimuth; v k,n Indicates that there may be ambiguity in the measured radial velocity; d k,n Indicates the Doppler channel where the point is located; s k,n Indicates the signal-to-noise ratio of the trace; i k,n Indicates the time sequence coding of the trace; a k,n Range Doppler diagram showing the point traces; Represents a transpose operation; In the nth frame, there are a total of D n The points are represented as:

3. The multi-frame joint detection method based on graph neural network link prediction task according to claim 2 is characterized in that: The maximum speed constraint in S200 is expressed as: In the formula, the two point tracks are respectively expressed as (t1, r1, θ1, v1) and (t2, r2, θ2, v2), where v max represents the maximum moving speed of the target of interest, t1 < t2, r1, θ1, v1 respectively represent the measured radial distance, measured azimuth angle, and possibly ambiguous measured radial speed of point track 1 in the current frame; r2, θ2, v2 respectively represent the measured radial distance, measured azimuth angle, and possibly ambiguous measured radial speed of point track 2 in the subsequent frame.

4. The multi-frame joint detection method based on graph neural network link prediction task according to claim 3 is characterized in that: S300 includes: S310, using NFEN mining to obtain multi-dimensional extraction features of each node in the point trace association graph, the multi-dimensional extraction features including: a feature vector describing the point trace Doppler channel Characteristic vector describing the signal-to-noise ratio of a point trace Characteristic vector describing the timing information of the trace and the feature vector describing the structure information of the trace signal S320, for each node, the multi-dimensional extraction features of the node are fused to obtain a multi-dimensional fusion feature, and the multi-dimensional fusion features of all nodes are combined into a point trace extraction feature matrix.

5. The multi-frame joint detection method based on graph neural network link prediction task according to claim 4 is characterized in that: S310 includes: S311, using three fully connected neural networks of the NFEN module to calculate the d of each node in the point trace association graph k,n 、s k,n 、i k,n Transform to obtain the characteristic vector describing the point Doppler channel Characteristic vector describing the signal-to-noise ratio of a point trace and the feature vector describing the timing information of the trace in, In the formula, Represents the operations between layers in the NFEN module, Represents a function composite operation, N d ,N s ,N i represents the number of fully connected network layers, Represents the learnable parameters of each layer of the network, φ D Represents the nonlinear activation function in the NFEN module; S312, using the convolutional neural network of NFEN to calculate the a of each node in the point trace association graph k,n Transform to obtain the feature vector describing the structure information of the point trace signal It is expressed as: Where N p Indicates the number of convolutional layers, * indicates the convolution operation, Represents the convolution kernel used in each layer of convolution operation.

6. The multi-frame joint detection method based on graph neural network link prediction task according to claim 5 is characterized in that: S320 includes: S321, will Splicing is performed to obtain the multi-dimensional fusion feature f of each node k,n , expressed as: S322, multi-dimensional fusion features f of all nodes k,n Rewrite and form the point trace extraction feature matrix f.

7. The multi-frame joint detection method based on graph neural network link prediction task according to claim 6 is characterized in that: The S322 includes: S3221, rewrite the multidimensional fusion feature of each point trace according to the measurement time sequence of the point traces in multiple frames to obtain the rewritten multidimensional fusion feature, which is expressed as: In the formula, L represents the total number of frames of observation data, l represents the lth frame, and there are D l A dot trace, Represents the rewritten multi-dimensional fusion features of the lth frame; S3222, rewrite the multi-dimensional fusion features to form a point trace extraction feature matrix f.

8. The multi-frame joint detection method based on graph neural network link prediction task according to claim 7 is characterized in that: S400 includes: S410, using the point trace extraction feature matrix f as the input node feature matrix x in STEF-GAN (0) , so that each layer of STEF-GAN transmits messages according to the message passing formula, and in the transmission process, the multi-frame space-time coupling features of the point trace association graph are mined to obtain the final features of each point trace; wherein the message passing formula is expressed as: In the formula, k represents the number of layers; represents the feature of node i in the kth layer, the input node feature matrix x of STEF-GAN (0) That is, the point trace extraction feature matrix f output by the NFEN network; represents the set of neighbor nodes of node i; e j,i Represents the characteristics of the edge between node i and node j; A learnable matrix representing the features of nodes in the kth layer; Represents the learnable matrix of the k-th layer edge features, φ G Represents the nonlinear activation function in the GNN network; represents a fully connected layer; represents aggregation operation, and || represents feature merging; represents the attention coefficient of the kth layer; Represents the learnable parameters in the attention mechanism; A learnable matrix representing the edges in the attention mechanism; The final feature is expressed by the formula: In the formula, G . Represents the inter-layer operations in STEF-GAN; Indicates function composite operation; N G represents the number of STEF-GAN network layers; e represents the edge feature matrix of the input STEF-GAN network.

9. The multi-frame joint detection method based on graph neural network link prediction task according to claim 1 is characterized in that: S500 includes: S510, for any two nodes having a common edge, using OAJN to fuse the final features of the any two nodes, and using the fused final features to calculate the score of the common edge between the any two nodes; S520, dividing the common edge into an edge between false alarms and false alarms, an edge between false alarms and a target, and an edge between a target and a target according to the classification score of the common edge and the second threshold; S530: The edges between false alarms and false alarms, and the edges between false alarms and targets are regarded as non-compliant common edges, and the common edges between targets and targets are removed to obtain the retained common edges.

10. The multi-frame joint detection method based on graph neural network link prediction task according to claim 9 is characterized in that: The classification score of the common edge between any two nodes is expressed as: In the formula, and Represents any two nodes, Indicates the scores of the edges between two nodes predicted as false alarm-false alarm edges, false alarm-target edges, and target-target edges, respectively. is the learnable parameter in OAJN; φ M represents the nonlinear activation function in OAJN, d j,i Represents the distance between node i and node j.

Images