Vessel Target Point-Track Association Method and System Based on Graph Representation Learning

Abstract

The present invention discloses a method and system for vessel target point-track association based on graph representation learning, which relates to the technical field of radar target tracking and data association. The method includes obtaining an R-D spectrum from the acquired echo signal, dividing the target foreground, and constructing an undirected weighted graph structure; using an encoder to convert the graph structure into a low-dimensional latent embedding, and then using a decoder to reconstruct the adjacency matrix; after training, using the encoder to extract features from the graph structure to obtain graph feature vectors; at the point-track association gate, calculate the similarity between the graph feature vectors of the candidate measurement points and the graph feature vectors of the current track history for three frames respectively, and select the candidate measurement point with the highest similarity as the point associated with the current track. The present invention realizes accurate association between points and tracks, and overcomes the problem of association errors caused by low signal-to-noise ratio and single target features.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target tracking and data association, and in particular to a method and system for ship target point-track association based on graph representation learning. Background Art

[0002] With the continuous improvement of the demand for maritime surveillance, compact ground-wave radars have been widely used in the field of maritime ship target detection due to their advantages such as small size, low cost, and flexible deployment. However, due to system limitations (such as low transmit power and low azimuth resolution) and the influence of interference such as sea clutter and ground clutter, radar echoes often have problems such as low signal-to-noise ratio and high measurement uncertainty. Traditional maritime ship point-track association methods mainly rely on target kinematic parameters, and it is often difficult to distinguish adjacent targets and clutter in complex environments, easily causing point-track association ambiguity problems, resulting in frequent occurrence of incorrect associations or track breaks.

[0003] With the continuous development of deep learning technology, in recent years, in some studies, attempts have been made to introduce deep learning technology into the field of data association. However, compact ground-wave radars are severely affected by complex working environments and variable target characteristics, and still face the problem of weak target features in the point-track association process. Existing target detection frameworks often only focus on extracting the kinematic parameters of targets, ignoring the rich spectral features in radar echoes. In fact, the range-Doppler (R-D) spectrum of radar can not only reflect the geometric position and velocity information of targets, but also contains rich energy distribution and spatial correlation data, which is of great value for distinguishing different targets and targets from clutter. However, due to the complexity of electromagnetic interaction, traditional signal processing methods are difficult to effectively extract and utilize these spectral features. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a method and system for ship target point-track association based on graph representation learning, which uses graph representation learning to extract graph structure features reflecting target scattering characteristics from the R-D spectrum, constructs recognizable target feature vectors, thereby realizing accurate association of points and tracks, and overcoming the association error problems caused by low signal-to-noise ratio and single target features.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] In the first aspect, the present invention provides a method for ship target point-track association based on graph representation learning, including:

[0007] Obtain an echo signal containing ship target data and maritime environment data, and preprocess the echo signal to obtain an R-D spectrum;

[0008] Based on the R-D spectrum, the target foreground is segmented, and the cells in the target foreground are used as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix;

[0009] The encoder is used to convert the graph structure into a low-dimensional latent embedding, and then the decoder reconstructs the adjacency matrix according to the latent embedding; the encoder and the decoder are trained. After the training is completed, the encoder is used to extract features from the graph structure to obtain a graph feature vector;

[0010] The graph feature vectors of the candidate measurement point tracks and the graph feature vectors of the last three frames of the current track history are respectively obtained. At the point track-track association gate, the similarities between the graph feature vectors of the candidate measurement point tracks and the graph feature vectors of the last three frames of the current track history are calculated respectively, and the candidate measurement point track with the highest similarity is selected as the point track associated with the current track.

[0011] As a further technical solution, the preprocessing includes: first, the echo signal is converted into a digital signal through A / D conversion, and then through pulse compression, beamforming, and Doppler processing, it is converted into an R-D spectrum; each resolution cell of the R-D spectrum is represented by a triple composed of distance, Doppler velocity, and echo amplitude.

[0012] As a further technical solution, the method for segmenting the target foreground is: in the R-D spectrum, a window is selected with the cell where the target peak is located as the center, and the Otsu method is used to determine the optimal amplitude segmentation threshold between the foreground and the background within the window; a target foreground segmentation mask is generated based on the optimal amplitude segmentation threshold, and the cells with an amplitude greater than or equal to the optimal amplitude segmentation threshold are segmented into the target foreground.

[0013] As a further technical solution, the method for constructing an undirected weighted graph structure is: first, the node corresponding to the target peak cell in the target foreground is set as the central node and numbered "1", and other nodes are numbered in the order of their relative positions to the central node; then the central node is connected to other nodes, and finally an undirected weighted graph structure is constructed.

[0014] As a further technical solution, in the training stage, the graph structure is input into the encoder, and the encoder uses a graph convolutional network to convert the graph structure into a low-dimensional latent embedding. Its encoding process can be expressed as:

[0015] ; where is the low-dimensional latent embedding, is the graph convolutional network, is the node feature matrix, is the adjacency matrix;

[0016] The formula for the decoder to reconstruct the adjacency matrix according to the latent embedding is:

[0017] ; wherein, is the reconstructed adjacency matrix, is the activation function.

[0018] As a further technical solution, the encoder adopts a two-layer GCN architecture; after training is completed, first, the first-layer GCN converts the input graph structure into a hidden embedding through a message passing operation, and then uses the Leaky ReLU activation function to introduce non-linearity; the second-layer GCN performs the message passing mechanism again on the basis of the hidden embedding generated by the first layer to generate node embeddings; finally, through a global average pooling operation, the feature vectors of all nodes are aggregated to obtain a graph feature vector.

[0019] As a further technical solution, calculate the similarity between the graph feature vector of the candidate measurement point track and the graph feature vectors of the three frames of the current track history. The specific method is: calculate the cosine similarity between the graph feature vector of the candidate measurement point track and the graph feature vectors of the three frames of the current track history, and then obtain the comprehensive similarity through weighted fusion. The calculation formula is:

[0020] ; wherein, , , are weight coefficients, is the graph feature vector of the candidate measurement point track, , , are the graph feature vectors of the three frames of the current track history.

[0021] In the second aspect, the present invention provides a vessel target point track - track association system based on graph representation learning, including the following modules:

[0022] A data acquisition and preprocessing module, configured to: acquire an echo signal containing vessel target data and marine environment data, and preprocess the echo signal to obtain an R-D spectrum;

[0023] A graph structure construction module, configured to: based on the R-D spectrum, divide the target foreground, and use the cells in the target foreground as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix;

[0024] A feature extraction module, configured to: use an encoder to convert the graph structure into a low-dimensional latent embedding, and then use a decoder to reconstruct the adjacency matrix according to the latent embedding; train the encoder and the decoder, and after training is completed, use the encoder to perform feature extraction on the graph structure to obtain a graph feature vector;

[0025] The tracklet-track association module is configured to: respectively obtain the graph feature vector of the candidate measurement tracklet and the graph feature vectors of the three previous frames of the current track history, and calculate the similarity between the graph feature vector of the candidate measurement tracklet and the graph feature vectors of the three previous frames of the current track history at the tracklet-track association gate, and select the candidate measurement tracklet with the highest similarity as the tracklet associated with the current track.

[0026] One or more technical solutions of the present invention have the following beneficial effects:

[0027] 1. By obtaining the echo signal, the present invention deeply excavates the spectrum information in the echo signal, extracts the energy distribution and spatial structure of the target in the R-D spectrum, and uses it as an important basis for association judgment, providing richer and more discriminative information support for distinguishing different tracklets, and significantly improving the accuracy and reliability of tracklet-track association.

[0028] 2. By extracting the graph feature vector, the present invention converts the spectrum feature into effective information available for association. Specifically, by constructing an undirected weighted graph from the local energy distribution of the target in the R-D spectrum, using the echo amplitude, Doppler index, and range index within the energy peak and its surrounding neighborhood as node information, and characterizing the scattering characteristics of the target by setting the weights of the edges between nodes. With the help of GAE, this variable-sized graph structure is further converted into a fixed-dimensional graph feature vector, enabling the effective extraction of the electromagnetic scattering information of the target and its quantification in vector form.

[0029] 3. During the tracklet-track association process, in order to quantitatively compare the graph feature vectors of the candidate tracklets and the historical tracklets on the target track, the present invention uses the cosine similarity as the similarity metric. The cosine similarity calculation can effectively eliminate the influence of the vector amplitude and pay more attention to the direction information of the feature distribution, thereby accurately quantifying the matching degree between the graph feature vectors. By calculating the cosine similarity between the graph feature vector of the candidate measurement tracklet and the graph feature vectors of the corresponding frames in the historical track, and combining the multi-frame weighted fusion strategy, the candidate measurement tracklet with the highest similarity is finally selected as the correct association result. Description of the Drawings

[0030] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0031] Figure 1 It is an example diagram of the R-D spectrum in the present invention;

[0032] Figure 2 It is an example diagram of the target foreground in the present invention;

[0033] Figure 3 It is the structural diagram of the GAE network in the present invention;

[0034] Figure 4 Schematic diagram for extracting the feature vector of the figure of the present invention;

[0035] Figure 5 Schematic diagram for point-track association of the present invention;

[0036] Figure 6 Comparison diagram of track tracking results under the multi-target proximity scenario of the present invention;

[0037] Figure 7 Comparison diagram of track tracking results under the sea clutter interference scenario of the present invention;

[0038] Figure 8 Flowchart of the point-track association method of the present invention. Detailed implementation manners

[0039] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0040] Embodiment 1

[0041] In this embodiment, a point-track association method for ship targets based on graph representation learning is provided, as Figure 8 shown in the flowchart of the point-track association method, and the specific steps include:

[0042] S1: Obtain the echo signal containing ship target data and marine environment data, and preprocess the echo signal to obtain the R-D spectrum;

[0043] S2: Based on the R-D spectrum, divide the target foreground, and use the cells in the target foreground as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix;

[0044] S3: Use the encoder to convert the graph structure into a low-dimensional latent embedding, and then use the decoder to reconstruct the adjacency matrix according to the latent embedding; train the encoder and the decoder, and after the training is completed, use the encoder to extract features from the graph structure to obtain the graph feature vector;

[0045] S4: Respectively obtain the graph feature vectors of the candidate measurement points and the graph feature vectors of the three frames of the current track history. At the point-track association gate, calculate the similarity between the graph feature vectors of the candidate measurement points and the graph feature vectors of the three frames of the current track history, and select the candidate measurement point with the highest similarity as the point associated with the current track.

[0046] In step S1, an echo signal containing vessel target data and marine environment data is obtained using a compact ground wave radar. Specifically, after the electromagnetic wave of the transmitter interacts with the marine vessel target and the surrounding marine environment, the generated echo signal is received by the receiver.

[0047] The specific method for preprocessing the echo signal is as follows: First, the echo signal is converted into a digital signal through A / D conversion. Then, through a series of digital signal processing techniques, including pulse compression, beamforming, and Doppler processing, the echo signal is converted into an R-D spectrum. As Figure 1 shown, this R-D spectrum is the basic data for subsequent processing. Each resolution cell of the R-D spectrum is represented by a triple consisting of distance, Doppler velocity, and echo amplitude.

[0048] In step S2, the specific method for dividing the target foreground is as follows: Based on the above-mentioned R-D spectrum obtained in step S1, as Figure 1 shown, taking the cell where the target peak is located as the center, a 5×5 window is selected, and this window can cover the spatial expansion range of the target echo in the R-D spectrum. The Otsu method is used to determine the optimal amplitude segmentation threshold S between the foreground and the background (clutter / noise) within the window. The Otsu method determines the threshold by maximizing the between-class variance between the foreground and the background, so as to maximize the difference between the two categories of the foreground and the background. Based on the optimal amplitude segmentation threshold S a target foreground segmentation mask is generated. Cells with an amplitude greater than or equal to the optimal amplitude segmentation threshold S are divided into the extended area of the target, that is, the target foreground, as Figure 2 shown.

[0049] In this embodiment, the target foreground segmentation mask is defined as follows:

[0050] ;

[0051] where is the echo amplitude of the th cell within the 5×5 window, and cells with are marked as target foreground cells.

[0052] In step S2, for each detected target, the cells in the target foreground are used as nodes in the constructed graph structure. Each node consists of a three-dimensional feature vector composed of Doppler velocity, distance, and echo amplitude, denoted as , where is the total number of nodes, is the Doppler velocity index, is the distance index, and is the echo amplitude.

[0053] The method for constructing an undirected weighted graph structure is as follows: First, set the node corresponding to the target peak cell in the target foreground as the central node and number it as "1", and number other nodes in the order of their relative positions to the central node to preserve the spatial layout information of the target scattering region; then connect the central node with other nodes to construct a graph structure. Among them, the weight of the edge is determined by the amplitude difference between the central node and the connected node, that is: Finally, an undirected weighted graph structure is constructed. where is the set of nodes, is the set of edges, and the node feature matrix integrates the features of the graph structure , and the adjacency matrix represents the connection relationship and weight of the edges.

[0054] In step S3, in this embodiment, a Graph Autoencoder (GAE) is used to extract features from the constructed graph structure. GAE consists of an encoder and a decoder and is an unsupervised learning framework. In the training stage, the encoder uses a Graph Convolution Network (GCN) to process the input graph structure. The graph convolution network learns the feature representation of nodes by aggregating the information of adjacent nodes. Its principle is to perform a convolution operation on the topological structure of the graph so that nodes can obtain the feature information of neighbor nodes.

[0055] Specifically: as Figure 3 shown, in the training stage, the graph structure is input into the encoder, and the encoder uses the graph convolution network to convert the graph structure (including the node feature matrix and the adjacency matrix ) into a low-dimensional latent embedding. Its encoding process can be expressed as:

[0056] ; where is the low-dimensional latent embedding, is the graph convolution network, is the node feature matrix, is the adjacency matrix.

[0057] The decoder reconstructs the adjacency matrix according to the latent embedding , and the reconstruction formula is:

[0058] ; where is the reconstructed adjacency matrix, is the activation function.

[0059] During the training process, by minimizing the original adjacency matrix The reconstruction with the reconstructed adjacency matrix The reconstruction loss Optimize the model. Use the Adam optimization algorithm to adjust the model parameters so that the encoder gradually learns effective embeddings to capture the structural features of the graph.

[0060] After the training is completed, use the encoder to extract features from the graph structure. As Figure 4 shown, the encoder adopts a two-layer GCN architecture. First, the first layer of GCN converts the input graph structure into a 64-dimensional hidden embedding through message passing operations, and then uses the Leaky ReLU activation function to introduce non-linearity to enhance the model's ability to capture complex patterns; the second layer of GCN further optimizes the hidden embedding. Based on the hidden embedding generated by the first layer, the second layer of GCN executes the message passing mechanism again, aggregates the features of the node itself and its neighbor nodes, further explores the complex correlation relationships between features, and generates more discriminative and representative node embeddings , where each row of data corresponds to the feature vector of a node in the graph structure. Finally, through the global average pooling operation, the feature vectors of all nodes are aggregated to obtain a fixed-length 64-dimensional global graph feature vector , and the calculation formula is: , and this graph feature vector effectively encodes the unique electromagnetic scattering characteristics of the target.

[0061] In step S4, Figure 5 is a schematic diagram of point-track association. Inside the point-track association gate, calculate the similarity between the graph feature vector of the candidate measurement point track and the graph feature vector of the current track history and , their cosine similarity can be calculated by the following formula:

[0062] ;

[0063] In this embodiment, obtain the graph feature vector of the candidate measurement point track and the graph feature vectors of the three frames of the current track history respectively, calculate the cosine similarity between the graph feature vector of the candidate measurement point track and the graph feature vectors of the three frames of the current track history respectively, and then obtain the comprehensive similarity through weighted fusion. The calculation formula is:

[0064] ; where , , are weight coefficients, satisfying > > and + + = 1 to highlight the role of the latest target state in the association while taking historical information into account to ensure stability; is the graph feature vector of the candidate measurement point track; , , are the graph feature vectors of the three frames of the current track history.

[0065] If the target state of a certain frame is a prediction rather than a measurement value (i.e., there is a lack of effective radar echo and the corresponding graph feature vector), the corresponding weight coefficient is set to 0. To enhance the similarity difference, the formula is used to transform the comprehensive similarity, where = is the numerical stability factor to prevent the logarithmic term from diverging when is close to 1. Finally, the candidate measurement point track with the highest similarity is selected as the track associated with the current track.

[0066] The following are the experimental results and analysis of this embodiment:

[0067] To quantitatively evaluate the performance of point track - track association, the target tracking duration is used as a key indicator in this experiment. This indicator reflects the ability of the radar system to maintain tracking continuity. The longer the tracking duration, the higher the accuracy of the point track - track association of the system. In addition, by qualitatively analyzing the association effect under the influence of multi - target proximity and sea clutter interference, the association accuracy and track integrity of different methods in complex environments are compared.

[0068] Quantitative analysis:

[0069] Statistical analysis is carried out on the average tracking duration of 10 representative targets. The method provided in this embodiment is compared with the traditional Nearest Neighbor Data Association (NNDA) and the advanced Joint Probability Data Association (JPDA) methods. The results are shown in Table 1:

[0070] Table 1 Comparison of average tracking duration

[0071] (Unit: minutes, "+" indicates the addition of the tracking durations of multiple track segments)

[0072]

[0073] The statistical results show that the average tracking duration of the method proposed in this embodiment reaches 135.6 minutes, which is significantly improved compared with NNDA (average 99.7 minutes) and JPDA (average 102.2 minutes), with an average extension of about 32% - 36%, indicating that this method has significant advantages in the continuity of target tracking.

[0074] Qualitative analysis:

[0075] 1. Multi-target proximity scenario

[0076] As Figure 6 shown, in the multi-target proximity case, the traditional NNDA method has a point-track association error caused by the interference of adjacent targets at the 70th minute (marked by the blue dot), resulting in the premature termination of the target track; while this embodiment effectively distinguishes the measurement point tracks of adjacent targets using the graph feature vector, realizes the correct point-track association, and increases the target tracking duration by 88 minutes.

[0077] 2. Sea clutter interference scenario

[0078] As Figure 7 shown, in the sea clutter interference case, due to the influence of sea clutter interference, the traditional method is difficult to achieve accurate point-track association in the starting stage, resulting in the track being unable to start in a timely and accurate manner; while this embodiment can establish the target track in a timely manner and maintain a high association accuracy within the sea clutter interference area, ultimately extending the target tracking duration by 14 minutes (the track segment marked by the blue dashed ellipse).

[0079] The above qualitative analysis cases show that this method has high robustness in complex environments, can achieve accurate point-track association, and ensure the continuous and stable tracking of the target track.

[0080] Embodiment 2

[0081] In this embodiment, a ship target point-track association system based on graph representation learning is provided, including the following modules:

[0082] Data acquisition and preprocessing module, configured to: acquire the echo signal containing ship target data and marine environment data, and preprocess the echo signal to obtain the R-D spectrum;

[0083] Graph structure construction module, configured to: based on the R-D spectrum, divide the target foreground, and use the cells in the target foreground as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix;

[0084] A feature extraction module, configured to: use an encoder to convert the graph structure into a low-dimensional latent embedding, and then use a decoder to reconstruct an adjacency matrix according to the latent embedding; train the encoder and the decoder, and after the training is completed, use the encoder to perform feature extraction on the graph structure to obtain a graph feature vector;

[0085] A measurement-track association module, configured to: respectively obtain the graph feature vector of a candidate measurement point and the graph feature vectors of the three previous frames of the current track history, and at a measurement-track association gate, respectively calculate the similarity between the graph feature vector of the candidate measurement point and the graph feature vectors of the three previous frames of the current track history, and select the candidate measurement point with the highest similarity as the measurement point associated with the current track.

[0086] For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A vessel target point-track association method based on graph representation learning, characterized in that Including: Obtain an echo signal containing vessel target data and marine environment data, and preprocess the echo signal to obtain an R-D spectrum; Based on the R-D spectrum, divide the target foreground, and use the cells in the target foreground as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix; Use an encoder to convert the graph structure into a low-dimensional latent embedding, and then use a decoder to reconstruct the adjacency matrix according to the latent embedding; train the encoder and the decoder, and after training is completed, use the encoder to extract features from the graph structure to obtain a graph feature vector; Respectively obtain the graph feature vectors of candidate measurement point traces and the graph feature vectors of the current track history for three frames. At the point trace-track association gate, calculate the similarity between the graph feature vectors of the candidate measurement point traces and the graph feature vectors of the current track history for three frames respectively, and select the candidate measurement point trace with the highest similarity as the point trace associated with the current track.

2. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein The preprocessing includes: first convert the echo signal into a digital signal through A / D conversion, and then through pulse compression, beamforming, and Doppler processing, convert it into an R-D spectrum; each resolution cell of the R-D spectrum is represented by a triple composed of distance, Doppler velocity, and echo amplitude.

3. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein The method for dividing the target foreground is: in the R-D spectrum, take the cell where the target peak is located as the center, select a window, and use the Otsu method to determine the optimal amplitude segmentation threshold between the foreground and the background within the window; generate a target foreground segmentation mask based on the optimal amplitude segmentation threshold, and the cells with an amplitude greater than or equal to the optimal amplitude segmentation threshold are divided into the target foreground.

4. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein The method for constructing an undirected weighted graph structure is: first set the node corresponding to the target peak cell in the target foreground as the central node and number it as "1", and number the other nodes in the order of their relative positions to the central node; then connect the central node to the other nodes, and finally construct an undirected weighted graph structure.

5. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein In the training stage, the graph structure is input into the encoder, and the encoder uses a graph convolutional network to convert the graph structure into a low-dimensional latent embedding, and its encoding process can be expressed as: ; wherein, is a low-dimensional latent embedding, is a graph convolutional network, is a node feature matrix, is an adjacency matrix; The formula for the decoder to reconstruct the adjacency matrix according to the latent embedding is: ; among them, is the reconstructed adjacency matrix, is the activation function.

6. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein, The encoder adopts a two-layer GCN architecture; after training is completed, first the first layer of GCN converts the input graph structure into a hidden embedding through a message passing operation, and then uses the Leaky ReLU activation function to introduce non-linearity; The second layer of GCN executes the message passing mechanism again on the basis of the hidden embedding generated by the first layer to generate node embeddings; finally, through a global average pooling operation, the feature vectors of all nodes are aggregated to obtain a graph feature vector.

7. The method for associating vessel target point tracks based on graph representation learning according to claim 1, wherein, Calculate the similarity between the graph feature vectors of the candidate measurement point traces and the graph feature vectors of the current track history for three frames respectively. The specific method is: calculate the cosine similarity between the graph feature vectors of the candidate measurement point traces and the graph feature vectors of the current track history for three frames respectively, and then obtain the comprehensive similarity through weighted fusion. The calculation formula is: ; wherein, , , are weight coefficients, is the graph feature vector of the candidate measurement point track, , , are the graph feature vectors of the three frames of the current track history.

8. A vessel target point-track association system based on graph representation learning, characterized in that Including the following modules: The data acquisition and preprocessing module is configured to: acquire an echo signal containing vessel target data and marine environment data, and preprocess the echo signal to obtain an R-D spectrum; The graph structure construction module is configured to: based on the R-D spectrum, divide the target foreground, and use the cells in the target foreground as nodes to construct an undirected weighted graph structure including a node feature matrix and an adjacency matrix; The feature extraction module is configured to: use an encoder to convert the graph structure into a low-dimensional latent embedding, and then use a decoder to reconstruct the adjacency matrix according to the latent embedding; train the encoder and the decoder, and after the training is completed, use the encoder to extract features from the graph structure to obtain a graph feature vector; The measurement-track association module is configured to: respectively obtain the graph feature vectors of candidate measurement points and the graph feature vectors of the last three frames of the current track history, and in the measurement-track association gate, calculate the similarity between the graph feature vectors of the candidate measurement points and the graph feature vectors of the last three frames of the current track history, and select the candidate measurement point with the highest similarity as the measurement point associated with the current track.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps in the vessel target measurement-track association method based on graph representation learning according to any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the vessel target measurement-track association method based on graph representation learning according to any one of claims 1-7.

Images