Compact ground wave radar track association method and system

Abstract

The present application belongs to the technical field of radar data processing, and in particular to a compact ground wave radar track association method and system, aiming to solve the association ambiguity problem of T / R and T-R dual radars in a multi-target dense scene, by acquiring and preprocessing the track data of the two radars to generate time and space aligned short track segments; for each segment, a local topological graph is constructed with the main target as the center and the neighborhood targets as the nodes, and a graph sequence pair to be compared is formed in time sequence; the graph sequence pair is input into a twin GAT-LSTM network, spatial features are extracted through a graph attention network (GAT), time sequence features are fused through a long short-term memory network (LSTM), and a double-branch high-dimensional space-time feature vector is output, and the segment-level association score is obtained through splicing and similarity estimation; based on the score set of all segments in the track overlap area, a consistency index is calculated and the track association is determined. By jointly modeling the time and space features of the target, the accuracy and robustness of track association in a target dense environment are improved.

Description

Technical Field

[0001] This invention belongs to the field of radar data processing technology, specifically a method and system for correlation of tracks in compact ground wave radar. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] High-frequency ground-wave radar is an important means of achieving large-scale, over-the-horizon detection of maritime targets. The T / RR compact ground-wave radar system works collaboratively through a T / R monostation radar subsystem and a TR bistation radar subsystem. Its core characteristic lies in sharing a single transmitting station (T) and using two geographically separated receiving stations (R) to collaboratively observe the same target from different geometric perspectives. Ultimately, track fusion compensates for the shortcomings of a single compact ground-wave radar in terms of detection accuracy, resolution, and anti-jamming capabilities. However, to achieve effective target track fusion, it is essential to accurately determine whether two overlapping tracks from the T / R and TR stations correspond to the same maritime target; this process is called "track association."

[0004] Traditional track association methods primarily rely on statistical or logical comparisons of target kinematic parameters (such as position, velocity, and heading). Their core assumption is that kinematic parameter observations of the same target under different sensors should be statistically consistent. Therefore, while traditional statistical methods are computationally efficient, in multi-target scenarios, tracks of multiple targets with similar kinematic parameters are prone to erroneous associations. Meanwhile, fuzzy mathematics methods, which rely on expert rules to design membership functions, are poorly adaptable to complex and variable marine environments. Furthermore, while existing deep learning methods can extract nonlinear features from tracks, they typically only process sequences of motion states of a single target, resulting in insufficient representational power of the extracted track features.

[0005] In fact, in dense target scenarios, the spatial topological relationships and dynamic changes among neighboring targets contain important spatiotemporal characteristics. These characteristics play a crucial role in distinguishing neighboring targets with similar motion patterns. Existing technologies lack track association methods that can simultaneously describe the spatial topological relationships and temporal changes of multiple targets and adapt to the complex observation characteristics of T / RR compact ground wave radars. Summary of the Invention

[0006] This invention provides a method and system for track association of compact ground wave radar, which solves the problem of high track association error rate caused by similar target movements and close spatial positions in dense multi-target scenarios of T / RR compact ground wave radar, and improves the accuracy and robustness of cross-radar target association.

[0007] The first aspect of this invention discloses a compact ground-wave radar track association method, comprising the following steps:

[0008] The track data of T / R monostatic radar and TR bistatic radar are acquired and preprocessed to obtain time-synchronized and spatially unified track pairs to be associated. The two tracks in the track pairs are then split into multiple short track segments of equal length to form multiple track segment pairs.

[0009] For each time step in the track segment, the target corresponding to the current track point is the master node, and other targets in its spatial neighborhood are the neighbor nodes. A local star topology graph is constructed, and multiple local graphs arranged in time sequence within the same track segment are combined into a graph sequence to obtain a pair of "graph sequences to be compared".

[0010] The graph sequence to be compared is input into a Siamese GAT-LSTM network, which has at least two GAT-LSTM encoders with shared parameters and identical structures. Each GAT-LSTM encoder includes a graph attention network unit and a long short-term memory network unit. The graph spatial features of each time step are extracted by the graph attention network unit, and the temporal features of the graph sequence from multiple time steps are fused by the long short-term memory network unit. Each branch encoder outputs a high-dimensional spatiotemporal feature vector representing the main target of the track segment. The high-dimensional spatiotemporal feature vectors output by the Siamese GAT-LSTM network are concatenated, and similarity estimation is performed to obtain the segment-level association similarity score.

[0011] The original track pairs to be associated are split into multiple short track fragment pairs with overlapping time. All fragment pairs are traversed, and the above steps are performed on each fragment pair in sequence to obtain a set of similarity scores for all fragment pairs. Based on this set, a track-level association consistency index is calculated. If the index exceeds a preset threshold, the two tracks are determined to be associated and correspond to the same target.

[0012] Further preprocessing includes:

[0013] Interpolation algorithms are used for resampling to synchronize track data to a unified timeline;

[0014] The original polar coordinate measurement data of the radar were uniformly converted to the geographic coordinate system;

[0015] The target's speed and heading information are calculated based on the coordinate difference between adjacent time steps, forming a standardized track state sequence.

[0016] Furthermore, when constructing a local star topology, the total number of nodes is... It includes 1 master node and a maximum of The number of neighboring nodes; if the number of neighboring targets in the neighborhood exceeds [number]... At that time, the targets are sorted according to their Euclidean distance, and only those closest to the main target are retained. If there are not enough neighbors, the node positions are filled with zero vectors.

[0017] Furthermore, the graph attention network unit employs a multi-head attention mechanism; for a master node in the graph, its spatial feature vector is concatenated by weighted aggregation results calculated independently by each attention head, as shown in the following equation:

[0018] ;

[0019] in, Indicates the first Attention weights calculated from each attention head Represents the activation function of the exponential linear unit. express Time Neighbor Nodes The feature vectors, the number of attention heads are .

[0020] Furthermore, the Long Short-Term Memory (LSTM) network unit and the Graph Attention (GAT) network unit are embedded and coupled to form the GAT-LSTM unit. At each time step, the GAT-LSTM unit simultaneously receives the graph structure information at the current time step and the hidden state and unit state at the previous time step, and performs joint updates on the spatial topology information and temporal recursive state through its internal gating mechanism.

[0021] Furthermore, each branch encoder of the twin GAT-LSTM network includes GAT-LSTM units are stacked in a hierarchical order, and each layer recursively models the graph sequence.

[0022] Among them, the Layer in Hidden state output at all times The input features of the next layer at the same time; the hidden state of the top-level GAT-LSTM unit at the last time step. This serves as a comprehensive spatiotemporal representation of the flight path segment.

[0023] Furthermore, the two high-dimensional spatiotemporal feature vectors are concatenated and their similarity is estimated. Specifically, the two high-dimensional spatiotemporal feature vectors are concatenated along their feature dimensions to form a joint feature vector. The joint feature vector is then input into a similarity estimation module, which is a feedforward neural network containing fully connected layers and activation functions. Its output is a segment-level association similarity score between 0 and 1.

[0024] Furthermore, the twin GAT-LSTM network and similarity estimation module are pre-trained using an intra-group contrastive learning strategy. The training strategy uses real associated track pairs as positive samples and constructs negative samples by replacing the main target with other targets in the spatial neighborhood. Each positive sample pair is combined with multiple negative sample pairs into a training group for optimization training. Among them, the sample with the highest segment-level association similarity score is judged as an associated sample by the network.

[0025] Furthermore, a track-level correlation consistency index is calculated based on this set, specifically:

[0026] Two paths to be linked are generated together in their overlapping region. There are 10 track segment pairs, of which the number of track segment pairs identified as related by the twin GAT-LSTM network is 10. The track-level correlation consistency index is ;

[0027] When the consistency index is satisfied , If the preset trajectory-level consistency association threshold is met, the two trajectories are considered to originate from the same target and are therefore associated; otherwise, they are considered not associated.

[0028] A second aspect of the present invention discloses a compact ground-wave radar track association system, comprising:

[0029] The data acquisition and preprocessing module is configured to: acquire and preprocess track data from T / R monostatic radar and TR bistatic radar to obtain time-synchronized and spatially unified track pairs to be associated, and split the two tracks in the track pairs into multiple short track segments of equal length to form multiple track segment pairs;

[0030] The track map sequence module is configured to: for each time step in the track segment, take the target corresponding to the current track point as the master node and other targets in its spatial neighborhood as neighbor nodes, construct a local star topology map, and form a map sequence by arranging multiple local maps in the same track segment in time order, to obtain a pair of "map sequences to be compared";

[0031] The spatiotemporal feature encoding and fusion module is configured as follows: the graph sequence to be compared is input into a twin GAT-LSTM network, which has at least two GAT-LSTM encoders with shared parameters and identical structures. Each GAT-LSTM encoder includes a graph attention network unit and a long short-term memory network unit. The graph spatial features of each time step are extracted through the graph attention network unit, and the temporal features of the graph sequence of multiple time steps are fused through the long short-term memory network unit. Each branch encoder outputs a high-dimensional spatiotemporal feature vector representing the main target of the track segment.

[0032] The track fragment association module is configured to: concatenate the high-dimensional spatiotemporal feature vectors output by the twin GAT-LSTM network and perform similarity estimation to obtain a fragment-level association similarity score; split the original track pairs to be associated into multiple short-term track fragment pairs with overlapping time; traverse all fragment pairs and perform the above steps sequentially for each fragment pair to obtain a set of similarity scores for all fragment pairs; calculate the track-level association consistency index based on this set; if the index exceeds a preset threshold, it is determined that the two tracks are associated and correspond to the same target.

[0033] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects:

[0034] 1. By constructing a temporal neighborhood graph sequence centered on the main target, the spatial topological relationships and temporal variation patterns among multiple targets are used as the core basis for association discrimination. The "track association" problem is reconstructed into a "spatiotemporal graph similarity discrimination" problem. Combining the target-centered temporal neighborhood graph sequence, the twin GAT-LSTM network extracts discriminative spatiotemporal features, and mechanisms such as segment-level matching and track-level consistency judgment, the system can effectively distinguish targets with similar motion parameters but different spatial contexts by utilizing the "neighborhood environment" as an information source. This significantly reduces the false association and missed association rates in dense target environments, and improves the accuracy and robustness of cross-radar target track association.

[0035] 2. To address the heterogeneity of track data between T / R and TR radars due to observation perspective and system errors, a parameter-sharing twin GAT-LSTM network is employed. This architecture maps track image sequences from different radars to a unified latent feature space for similarity measurement. During training, the network is optimized to learn feature representations that are insensitive to changes in perspective and system errors but sensitive to the intrinsic spatiotemporal patterns of the target. This eliminates the reliance on complex manual corrections for data heterogeneity, improving the method's engineering practicality and generalization performance.

[0036] 3. A GAT-LSTM spatiotemporal feature encoder was designed for graph sequence data. The Graph Attention Network (GAT) uses a multi-head attention mechanism to adaptively assign differentiated aggregation weights to different neighboring nodes, focusing on key neighborhood information and suppressing interference. The Long Short-Term Memory Network (LSTM) recursively fuses the spatial structure features extracted at each time step with historical states under a gating mechanism, resulting in a final extracted track feature vector that simultaneously contains detailed spatial relationships and long-term temporal dependencies, exhibiting stronger adaptability to observation noise, target maneuvers, and partial occlusion.

[0037] 4. By decomposing long tracks into multiple short segments, high-precision feature matching is first performed at the segment level using a deep learning model, and then consistent aggregation and decision are made on the matching results of multiple segments at the track level. This can effectively smooth out local misjudgments caused by instantaneous measurement errors, short-term target maneuvers, or single missed detections, and ensure the reliability and stability of the correlation results from a global perspective. It is particularly suitable for complex and non-stationary marine monitoring environments. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0039] Figure 1 A flowchart illustrating a compact ground wave radar track association method provided in one or more embodiments of the present invention;

[0040] Figure 2 A schematic diagram of the spatial feature extraction process based on GAT provided for one or more embodiments of the present invention;

[0041] Figure 3 A structural diagram of a GAT-LSTM spatiotemporal feature coding unit provided for one or more embodiments of the present invention;

[0042] Figure 4 This is an overall architecture diagram of the trajectory association method based on twin GAT-LSTM network provided in one or more embodiments of the present invention;

[0043] Figure 5 A comparison chart of the correlation accuracy of different methods provided in one or more embodiments of the present invention;

[0044] Figure 6 A trend chart of association accuracy variation under different target densities provided for one or more embodiments of the present invention;

[0045] Figure 7 A comparison chart of the association accuracy of different feature extractors provided for one or more embodiments of the present invention;

[0046] Figure 8 This is a schematic diagram of a typical multi-target trajectory scenario based on radar measured data, provided for one or more embodiments of the present invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] Track association involves measuring the similarity and classifying the target motion state sequences (tracks) reported by different sensors in complex environments containing observation errors, missed target detections, false alarms, and multi-target interference.

[0050] The T / RR compact ground wave radar is a bistatic cooperative detection system consisting of two compact ground wave radars. Based on a "one transmitter, two receivers" configuration, it comprises a co-located T / R subsystem and a long-range R receiving station. Its core feature is the shared use of a single transmitting station (T) and the coordinated observation of targets through two geographically separated receiving stations (R), significantly improving detection accuracy and anti-jamming capabilities. For ease of representation, the system consisting of the other long-range R receiving station and the shared single transmitting station T is called the TR subsystem.

[0051] As introduced in the background section, the core assumption of traditional track association methods is that the observed motion parameters of the same target under different sensors should be statistically consistent. Therefore, methods based on statistical theory are represented by classic nearest neighbor methods. The principle is to calculate the statistical distance (such as Mahalanobis distance or Euclidean distance) between the state vectors (usually position, velocity, and heading) of two tracks at the same time point and compare it with a preset threshold. These methods are computationally simple and have high real-time performance, but their discrimination criteria rely entirely on the target's own instantaneous or short-term kinematic parameters. In scenarios with multiple dense targets, especially those with spatial proximity and similar motion patterns (such as formation sailing or parallel channels), the tracks of multiple different targets are prone to exhibit high similarity in kinematic parameters, making it impossible to effectively distinguish them using statistical distance, thus generating a large number of erroneous and missed associations.

[0052] Some fuzzy mathematics-based methods utilize fuzzy set theory to address uncertainties in association decision-making. By designing membership functions, differences in parameters such as position, speed, and heading between tracks are transformed into fuzzy metrics of "association probability," and the association result is then derived through comprehensive evaluation. Such methods alleviate the rigidity of hard-threshold decisions to some extent. However, their performance is highly dependent on membership functions and weighting rules designed with expert experience. These rules are typically static and difficult to adapt to complex and changing marine environments (such as changes in observation characteristics caused by sea state and electromagnetic interference) and the diverse maneuvering behaviors of targets, resulting in poor generalization ability and robustness.

[0053] Meanwhile, deep learning-based methods, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), can automatically extract nonlinear time-dependent features from the historical state sequences of flight tracks, avoiding the need for manually designed features. However, existing deep learning methods typically treat each flight track as an independent time series, with the network model only learning the motion patterns of that target itself. This "isolated" feature extraction approach ignores the relative spatial relationships between the main target and its surrounding targets in dense target scenes, and their temporal changes. Different targets with similar kinematic parameters often have different surrounding "neighborhood environments" (such as the distribution of neighboring targets and their relative motion states). Existing deep learning association models fail to model this highly discriminative multi-target spatiotemporal topology, resulting in limited ability to distinguish between fuzzy association scenes where "motions are similar but targets are different."

[0054] Therefore, this solution provides a compact ground wave radar track association method and system, which reconstructs track association into a "spatiotemporal map similarity discrimination" problem: constructing a target-centered temporal neighborhood map sequence, extracting discriminative spatiotemporal features through a twin GAT-LSTM network, and achieving accurate association by combining segment-level matching and track-level consistency judgment.

[0055] like Figure 1 As shown, the compact ground wave radar track association method includes the following steps:

[0056] Track data preprocessing: The track data of T / R monostatic radar and TR bistatic radar are spatiotemporally aligned and sliced ​​preprocessed to generate short track segments with time synchronization and unified spatial reference.

[0057] Track map sequence construction: For each track segment, at each time step, a local topological map centered on the main target and including other targets in its spatial neighborhood is constructed, and a map sequence is formed in chronological order to represent the spatiotemporal relationship between multiple targets;

[0058] Spatial feature extraction based on GAT: A graph attention network is used to extract the spatial dependency features between nodes in the graph at each time step, and neighborhood information is adaptively aggregated through a multi-head attention mechanism;

[0059] Spatiotemporal feature encoding based on GAT-LSTM: A graph attention network is embedded into a long short-term memory network unit to construct a GAT-LSTM fusion encoder, which recursively encodes the graph sequence, integrates spatial topology and temporal change information, and outputs a high-dimensional spatiotemporal feature vector of the track segment.

[0060] Track segment association based on twin GAT-LSTM network: The track segment map sequences generated by the T / R monostatic radar and the TR bistatic radar to be associated are respectively input into two GAT-LSTM encoders with shared parameters, and the feature vectors are extracted and their similarity scores are calculated.

[0061] Network training and implementation: During the training phase, an intra-group contrastive learning strategy is adopted to construct spatially adjacent negative sample pairs to optimize the network's discriminative ability;

[0062] Track-level association decision: During the implementation phase, the similarity of all segments within the overlapping track interval is aggregated for consistency, and track-level association decision is completed based on the preset consistency threshold.

[0063] (1) Track data preprocessing.

[0064] Due to inherent differences in spatial location and observation angle between T / R monostatic radar and TR bistatic radar, track data of the same target generated by different radar systems suffer from problems such as asynchronous sampling rates, misaligned start times, and inconsistent spatial coordinate systems, making direct feature-level comparison impossible. Therefore, a unified spatiotemporal reference needs to be constructed to eliminate heterogeneity.

[0065] In the time dimension, to address the differences in update rates, an interpolation algorithm is used to resample the heterogeneous track states to a unified time axis;

[0066] In the spatial dimension, the original polar coordinate measurements of the radar are uniformly mapped to the geographic coordinate system, and the target's speed and heading are calculated using the coordinate difference between adjacent time points, thereby providing standardized input data for subsequent feature modeling.

[0067] After completing the spatiotemporal alignment, considering that the complete trajectory usually has a large time span and includes non-stationary maneuvers such as turning and acceleration, directly modeling the entire trajectory not only increases the difficulty of model learning, but also makes it easier for local subtle temporal features to be masked by global features.

[0068] Based on this, this scheme introduces a sliding window strategy, dividing the overlapping portion of the track pairs to be associated generated by the T / R monostatic radar and the TR bistatic radar into several short-time track segments of fixed length, as shown in the following formula, the first... Each flight track segment is defined as:

[0069] ;

[0070] in, This is the sequence of track states for the overlapping portion. For window length, The total length of the track state sequence in the overlapping part is the number of time steps it contains.

[0071] Through the above segmentation process, the complex full-track association problem is decomposed into a matching sub-problem of multiple short-time track segments. Since the motion characteristics within short-time track segments are approximately stable, this strategy is beneficial for the model to accurately capture the local motion consistency features of the target. At the same time, by aggregating the segment-level discrimination results, the sensitivity of local features and the consistency of overall association can be effectively balanced.

[0072] (2) Construction of track map sequence.

[0073] In multi-target, dense scenarios, the ambiguity of track association often stems from the spatial proximity and highly similar motion characteristics between targets. Traditional methods neglect the constraints of the relative spatial relationships between targets. To address this, our proposed scheme constructs a local neighborhood graph with a fixed radius centered on the target at each moment of a track segment. This graph describes the spatial topological relationships between the main target and other targets within its neighborhood.

[0074] Specifically, for any pair of candidate T / R and TR track segments, for each time step within it... A graph structure is constructed for each target. This graph structure treats the candidate target to be associated as the master node, and the remaining targets in the spatial neighborhood as candidate neighbor nodes. To maintain the uniformity of the input dimension of the deep learning model, the total number of nodes in the graph structure is fixed. (i.e., consisting of 1 master node and at most...) (composed of neighboring nodes). When the number of neighboring targets in the neighborhood exceeds... At that time, the targets are sorted according to their Euclidean distance, and only those closest to the main target are retained. There are 1000 neighbors; conversely, when there are not enough neighbors, zero vectors are used to fill the node positions to ensure that the graph structure dimension remains consistent at all times.

[0075] For each moment in the flight track segment , define the first The target state vector corresponding to each node is: ,in and These represent longitude and latitude, respectively. and These represent speed and heading, respectively. This forms the node feature matrix. ,right Perform Min-Max normalization on each column to obtain the normalized node feature matrix. .

[0076] To obtain a computationally efficient and structurally stable topological representation, this scheme adopts a star topology centered on the main target, establishing edge connections only between the main node and each neighboring node, without considering direct connections between neighboring nodes. With nodes exist Edge weight at time step Calculated using the Gaussian kernel function:

[0077] ;

[0078] in, Let be the spherical distance between the two. A scaling factor used to control the sensitivity of edge weights to changes in spatial distance.

[0079] To preserve node information during neighborhood aggregation, a matrix composed of all edge weights is used. By introducing self-loops and then performing symmetric normalization, a normalized adjacency matrix is ​​obtained:

[0080] ;

[0081] in, It is the identity matrix. This is the corresponding degree matrix. Therefore... The graph structure at time points can be represented as follows .

[0082] To extract the spatial topological relationships of multiple targets and their temporal variations, this scheme organizes the graph structures constructed at multiple consecutive moments within a track segment into a graph sequence in chronological order:

[0083] ;

[0084] The above-mentioned graph structure construction and graph sequence organization process remains completely consistent in the track data processing of T / R radar and TR radar to ensure the comparability of the two types of heterogeneous data within the twin network framework.

[0085] (3) Spatial feature extraction based on GAT.

[0086] After constructing the graph sequence of the flight path segments, it is necessary to further extract discriminative spatial features from the constructed local neighborhood graph to characterize the nonlinear spatial relationship between the main target and its neighboring targets. Considering that traditional Graph Convolutional Networks (GCNs) typically assign fixed or equal weights to neighboring nodes during neighborhood aggregation, making it difficult to distinguish the differences in the impact of different interference sources on the main target, this scheme introduces a Graph Attention Network (GAT) as a spatial feature extractor. GAT dynamically learns the importance weights of different neighboring nodes to the main target through an adaptive attention mechanism, thereby effectively avoiding the problem of key information dilution caused by equal-weight aggregation.

[0087] like Figure 2 As shown, for the figure First, through a learnable weight matrix For the Feature vectors of each node Perform a linear transformation to map it to a high-dimensional latent feature space:

[0088] ;

[0089] Based on this, in order to characterize the nodes with neighboring nodes The correlation between them is calculated using the following formula, which is the attention coefficient between them:

[0090] ;

[0091] in, Let be a learnable attention weight vector, and || denote the vector concatenation operation.

[0092] Subsequently, the attention coefficients are subjected to LeakyReLU nonlinear activation and softmax normalization to obtain the final normalized attention weights:

[0093] ;

[0094] This weight quantitatively describes the neighboring nodes. During spatial feature aggregation, the master node The extent of their contribution.

[0095] like Figure 2 As shown, to further enhance the model's ability to capture complex spatial relationships and improve training stability, this scheme employs a multi-head attention mechanism. Let the number of attention heads be... Under this mechanism, nodes exist The spatial feature vector at time step is composed of the weighted aggregation results calculated independently by each attention head:

[0096] ;

[0097] in, Indicates the first Attention weights calculated from each attention head Represents the activation function of the exponential linear unit. Represents a node The model uses a neighborhood set. Through a multi-head attention mechanism, the model can model the multi-scale spatial dependencies between neighborhood nodes in parallel from different feature subspaces, significantly improving the robustness and discriminative power of feature representation.

[0098] Finally, all nodes in the graph are... Stacking the spatial feature vectors at each time step yields the spatial feature matrix at that time step.

[0099] ;

[0100] For ease of subsequent description, the above spatial feature extraction process can be represented as a graph attention operator:

[0101] ;

[0102] Its complete processing flow is as follows Figure 2 As shown, Figure 2 This demonstrates how to use attention coefficients to perform weighted aggregation of neighbor node features.

[0103] (4) Spatiotemporal feature encoding based on GAT-LSTM.

[0104] Because the target motion process has significant temporal correlation, and its spatial relationship with neighboring targets also changes dynamically over time, relying solely on static graph structures is insufficient to fully describe the complex spatiotemporal interaction patterns in multi-target scenarios. Therefore, this solution embeds the Gaussian Array Atlas (GAT) into a Long Short-Term Memory (LSTM) network, employing a GAT-LSTM spatiotemporal feature fusion mechanism. This aims to leverage the long-term dependency capture capability of LSTM to achieve deep coupling between spatial topological information and temporal recursive states.

[0105] The structural diagram of the GAT-LSTM spatiotemporal feature encoding unit is shown below. Figure 3 As shown, this illustrates the process of embedding the graph attention operator into the input gate, forget gate, and output gate of an LSTM to achieve deep fusion of spatiotemporal features.

[0106] like Figure 3 As shown, in At any given time, the input to the GAT-LSTM spatiotemporal feature encoding unit includes the node feature matrix at the current time step. Adjacency matrix And the hidden state of the previous moment. and unit state To simultaneously encode the spatial structure information of the current input and the spatial structure information of historical moments, this scheme designs two parallel graph attention channels:

[0107] aisle Used to spatially encode the graph structure of the current input to obtain features. ;

[0108] aisle This is used to spatially encode the hidden state of the graph structure from the previous time step, yielding features. ;

[0109] The two feature sets are added element-wise to form the joint input for the LSTM gated computation:

[0110] ;

[0111] Next, The matrix is ​​divided into four sub-matrices, each corresponding to one of the input gates of the LSTM. Forgotten Gate Candidate cell status Pre-activation item of output gate :

[0112] ;

[0113] Based on this, the gate variables of the LSTM are obtained through a nonlinear activation function:

[0114] ;

[0115] in, This represents the sigmoid function. This represents the hyperbolic tangent function.

[0116] Under the control of the gated variables, the GAT-LSTM cells recursively update the cell state and hidden state according to the update rules of the standard LSTM:

[0117] ;

[0118] ;

[0119] in, This indicates that Hadamard multiplies element by element.

[0120] Through the aforementioned mechanism, spatial neighborhood information directly participates in the updating of the recursive temporal state, achieving deep fusion of spatiotemporal features. The essence of the "deep fusion" of the GAT-LSTM unit lies in changing the traditional LSTM mechanism that relies solely on its own state values ​​for temporal evolution, and instead establishing a coupling mode driven by spatial topology and time-gated control. The "forget gate" and "input gate" inside the LSTM are no longer determined solely by the isolated motion state of the target, but are dynamically modulated by the instantaneous spatial relationship between the target and its neighboring targets. At the physical level, this simulates the interactive effect of multi-target motion: when the spatial topology around the target changes drastically (such as neighboring targets moving closer or further away), the aggregated features extracted by GAT will fluctuate significantly, thereby directly intervening in the update strategy of the memory unit through nonlinear gating, forcing the model to adaptively adjust its reliance on historical motion trends. This design ensures that the recursive evolution in the temporal dimension is constrained by the topological constraints of the spatial dimension, thus achieving joint optimization of spatiotemporal information at the lowest level of feature encoding, rather than simple feature concatenation.

[0121] To extract higher-order spatiotemporal abstract features, this scheme further employs a hierarchical stacking approach to construct a GAT-LSTM encoder, such as... Figure 4 As shown. Let the encoder be... The GAT-LSTM units are stacked hierarchically, with each layer recursively modeling the graph sequence. In the hierarchical structure, the first... Layer in Hidden state output at all times It is directly used as the input feature of the next layer at the same time, that is (First layer input) Through this layer-by-layer transmission mechanism, the lower-level units primarily focus on short-term motion changes and local, instantaneous spatial relationships, while the higher-level units gradually aggregate motion consistency features over longer time scales and time-varying neighborhood spatial relationships. Finally, this scheme selects the hidden state of the top-level GAT-LSTM unit at the last moment. This serves as a comprehensive spatiotemporal representation of the flight path segment.

[0122] Considering that the core focus of track association is the main target, the hidden state vector corresponding to the main node is extracted as the final feature representation of the track segment:

[0123] ;

[0124] In this vector, the superscript (0) indicates the primary target node. This vector fully integrates the temporal dependencies and multi-target spatial interaction information within the track in terms of feature dimensions, providing a highly recognizable feature foundation for subsequent association discrimination.

[0125] (5) Track fragment association based on twin GAT-LSTM network.

[0126] After obtaining the high-dimensional spatiotemporal feature representation of the track segments, this scheme introduces a Siamese network architecture based on the idea of ​​metric learning, and maps the T / R and TR radar tracks to a unified latent feature space through a parameter-sharing dual-branch encoder.

[0127] Figure 4 This paper presents the overall architecture diagram of the track association method based on the twin GAT-LSTM network, and introduces the system-level architecture including a parameter-sharing dual-branch encoder, a feature stitching module, and a similarity estimation module.

[0128] Specifically, the sequence of T / R radar track segments A sequence of images of candidate TR radar track segments The inputs are fed into two identical GAT-LSTM encoders that share the same parameters. The encoder is able to automatically learn a consistent representation across radar tracks during training, making track segments from the same target as close as possible in the latent feature space, while effectively distinguishing track segments from different targets.

[0129] The feature vectors output by the two branch encoders are as follows:

[0130] ;

[0131] To quantify the correlation between the two, the two feature vectors are concatenated along their feature dimensions to form a joint feature vector:

[0132] ;

[0133] Subsequently, a similarity estimation module was introduced. This module is constructed from a multi-layer fully connected network. The first two layers combine the ReLU activation function to enhance non-linear expressive power, and the last layer uses the Sigmoid activation function to output a normalized similarity score. .in, This represents the probability estimate that T / R radar track segments and TR radar track segments originate from the same target. The larger the value, the higher the probability that the two are a correct correlation pair.

[0134] (6) Network training and implementation.

[0135] To effectively enhance the network's discrimination capability in multi-target dense scenarios, this scheme does not employ the traditional training method of random positive and negative sample pairs during network training and implementation. Instead, it introduces an intra-group comparison training strategy. This strategy uses real-world correlated track pairs as positive samples and constructs highly similar negative samples within their local spatial neighborhoods, thereby enabling the network to learn more refined spatiotemporal discrimination features. Specifically, for any pair of known real-world correlated T / R radar tracks and TR radar tracks, after completing temporal and spatial alignment, a track segment is extracted from the overlapping region to form a graph sequence positive sample pair. Subsequently, in the graph sequence corresponding to this track segment, the primary target node on one side (T / R or TR) is replaced with any other neighboring target node within its spatial neighborhood, while keeping the graph sequence on the other side unchanged, thereby constructing several negative sample pairs. or .

[0136] Because these negative samples are highly similar to the positive samples in terms of spatial location, velocity magnitude, and heading changes, but originate from different targets, they can effectively simulate the most challenging fuzzy scenarios in real-world association tasks. A positive sample pair is combined with multiple negative sample pairs to form a training group, where each training group contains... For each pair of samples, the corresponding network-predicted association similarity score is: The real label is ,in:

[0137] ;

[0138] The network training uses the cross-entropy loss function:

[0139] ;

[0140] This loss function prompts the network to maximize the similarity difference between positive and negative samples, so that positive samples obtain the highest association similarity score within the same training group, thereby effectively alleviating the association ambiguity problem in multi-object scenarios.

[0141] During the network implementation phase, for any T / R radar track segment to be associated... This was compared with all candidate TR radar track fragments. The trajectory segments are paired and fed into the trained Siamese GAT-LSTM network one by one. The similarity estimation module will output the corresponding set of similarity scores:

[0142] ;

[0143] Each similarity score Indicates the current T / R radar track segment and the first The probability estimate that the TR radar track segments originate from the same target.

[0144] (7) Track-level correlation decision.

[0145] Since the similarity results output by the network only reflect the feature matching degree at the short-term track segment level, directly relying on the similarity score of a single segment for association determination is highly susceptible to interference from random factors such as short-term abnormal measurements, target maneuvers, or missed detections. Therefore, this scheme further introduces consistency constraints at the full track level to construct a robust track-level association determination strategy. Suppose a pair of candidate tracks is divided into [number] segments within their overlapping interval. There are [number] track segment pairs, of which the number of track segment pairs determined to be "related" by the network is [number]. The track-level correlation consistency index is defined as follows:

[0146] ;

[0147] The final association decision is made based on this indicator: when the consistency indicator meets the following criteria... If the T / R radar track and the TR radar track originate from the same target, then they are considered unrelated. The preset trajectory-level consistency association threshold is used. This two-stage association strategy of "fragment-level feature matching + trajectory-level consistency decision" can effectively smooth the impact of local anomalies on the overall association results, and significantly improve the stability and reliability of the system in multi-object scenarios.

[0148] Thus, this solution completes track association, resolving the track association problem in multi-target scenarios. Specifically, for a maritime vessel target of interest or focus, the T / R radar tracks it and obtains a track. However, due to the large positioning error of compact ground-wave radar, this track is inaccurate. The TR radar can then track it to obtain another corresponding track, facilitating subsequent track fusion. However, multiple interfering targets exist in the local neighborhood of this target, easily leading to ambiguity in the association and making it difficult to determine the relationship between multiple targets. Therefore, this solution utilizes a graph structure to transform the mutual interference between multiple targets into usable spatial relationships.

[0149] Suppose there are four targets in a multi-target scenario. The four target tracks obtained by T / R radar independent tracking are labeled T1, T2, T3, and T4, respectively, and the four target tracks obtained by TR radar independent tracking are labeled R1, R2, R3, and R4, respectively. The actual correlation is T1-R1, T2-R2, T3-R3, and T4-R4. However, because these four targets are close to each other and have similar kinematic parameters, it is difficult to accurately determine the correspondence between the T / R radar tracks and the TR radar tracks. It is possible that R2 is closer to T1 (due to a larger radar positioning error), resulting in an incorrect correlation result such as T1-R2.

[0150] The original track data is filtered using simple spatiotemporal physical constraints. Each T / R radar track and TR radar track in the multi-target scenario is traversed through permutations and combinations to determine if there is an overlap in their time intervals. If overlap exists, this pair of T / R and TR radar tracks is grouped into a "candidate track pair to be associated" and added to the candidate set. This scheme is then used to evaluate each pair of tracks in the candidate set to obtain the final track association result. Therefore, in this embodiment, the "track to be associated" refers to a T / R radar track and a TR radar track awaiting association determination.

[0151] This approach overcomes the limitations of traditional methods that rely solely on kinematic parameters for correlation. By constructing a temporal local neighborhood graph centered on the main target, it creatively introduces the relative spatial topological relationships between targets into track correlation. This strategy utilizes the differences in neighborhood structure over time as a discriminative constraint, effectively solving the correlation ambiguity problem that arises in extreme scenarios where targets are spatially close and have highly similar motion characteristics, and significantly reducing the false correlation rate in dense target environments.

[0152] This scheme designs a GAT-LSTM spatiotemporal feature fusion mechanism, embedding GAT within an LSTM unit to achieve deep coupling between spatial topological information and temporal recursive states. During feature extraction, a multi-head attention mechanism adaptively assigns weights to different neighboring nodes, automatically focusing on high-value neighborhood information and suppressing interference from irrelevant targets. This mechanism ensures that the generated trajectory representation vector not only possesses spatiotemporal integrity but also exhibits strong robustness to complex multi-target scenarios.

[0153] To address the challenge of comparing heterogeneous T / R single-station and TR dual-station radar data, this solution employs a parameter-sharing twin network architecture. This maps heterogeneous tracks to a unified latent feature space for measurement, effectively mitigating the impact of observational geometric differences and systematic errors. Furthermore, by combining a two-stage strategy of "segment-level feature matching" and "track-level consistency decision," interference from local anomalies is smoothed from a global perspective, ensuring the stability of the final correlation results and demonstrating high engineering application value.

[0154] The overall network structure is as follows Figure 4 As shown, the specific steps include:

[0155] First, track data preprocessing is performed. A unified spatiotemporal reference is constructed for T / R single-station and TR dual-station radar data: in the time dimension, an interpolation algorithm is used to resample heterogeneous tracks to a unified time axis; in the spatial dimension, measurements are mapped to a geographic coordinate system and speed and heading are calculated. Subsequently, a sliding window strategy is introduced to divide the overlapping portion of the aligned T / R and TR radar tracks into several fixed-length short-time track segments, decomposing the full track correlation problem into a segment matching problem.

[0156] Then, a sequence of flight path graphs is constructed. For each moment in a flight path segment, a local star-shaped topology graph is constructed with the main target as the center and other targets in the spatial neighborhood as neighboring nodes. The edge weights between nodes are calculated using a Gaussian kernel function to characterize the influence of spatial distance, and the graph structure of consecutive moments within the segment is organized in chronological order to form a graph sequence describing the spatial topological relationships of multiple targets and their changing patterns.

[0157] Secondly, spatiotemporal feature encoding based on GAT-LSTM is performed. The constructed graph sequence is input into the GAT-LSTM encoder, and the GAT's adaptive attention mechanism is used to dynamically learn the importance weights of neighboring nodes to the main target, extracting spatial features. At the same time, by embedding GAT into the LSTM unit, the LSTM's gating mechanism is used to capture temporal dependencies, achieving deep coupling between spatial topological information and temporal recursive states, and finally extracting the high-dimensional spatiotemporal feature vector of the main target.

[0158] Then, similarity measurement of track fragment associations is performed based on Siamese networks. A parameter-sharing Siamese GAT-LSTM network architecture is used to extract features from the graph sequences of T / R radar track fragments and candidate TR radar track fragments, respectively, and map them to a unified latent feature space. The feature vectors output from the two branches are concatenated, and a similarity estimation module containing fully connected layers and a sigmoid activation function is used to calculate the probability score that both originate from the same target.

[0159] Finally, track-level association decisions are performed. The association results of all segment pairs of candidate track pairs within the overlapping interval are statistically analyzed, and a track-level association consistency index is calculated. When this index exceeds a preset consistency threshold, the T / R radar track and the TR radar track are determined to be the same target, thereby effectively smoothing out local anomaly interference and achieving stable and reliable cross-radar track association.

[0160] Through the above implementation methods, this scheme can efficiently and accurately complete the task of associating the tracks of maritime vessels, significantly improving the track association accuracy of compact ground-wave radar systems in dense target environments. Experimental results show that this method outperforms traditional statistical, fuzzy mathematics, and existing deep learning methods in terms of association accuracy.

[0161] To evaluate the effectiveness and robustness of the proposed compact ground wave radar track association method based on twin GAT-LSTM network, a systematic verification experiment covering simulation datasets and measured compact ground wave radar data was designed and implemented. The proposed method was also compared with several existing typical track association methods in multiple dimensions.

[0162] Figure 5 This chart compares the correlation accuracy of our proposed method with existing mainstream methods, demonstrating the performance advantages of our proposed method over traditional statistical methods and fuzzy mathematics methods. Figure 6 The graph shows the trend of association accuracy under different target densities, demonstrating that as the target density increases, this method consistently outperforms other comparative methods in terms of robustness. Figure 7 The comparison chart shows the correlation accuracy of different feature extractors, verifying the superiority of the GAT-LSTM encoder in this scheme compared to mainstream feature extractors such as CNN and GCN. Figure 8 This diagram illustrates a typical multi-target trajectory scenario based on radar measurement data, showcasing the distribution of multiple spatially adjacent vessels detected by T / R monostation and TR bistation radars in practical applications. Experimental results demonstrate that this scheme exhibits significant advantages in overall correlation accuracy, stability in dense multi-target scenarios, and feasibility for practical engineering applications.

[0163] Specifically, the method proposed in this scheme significantly outperforms traditional statistical methods (NN, GNN) and fuzzy mathematics methods (FSE, FIS) in terms of association accuracy. For example... Figure 5 As shown in the figure. In complex multi-target environments, the proposed method maintains its performance advantage and demonstrates excellent robustness as the target density increases. The trend of its association accuracy is shown in the figure. Figure 6 As shown in the figure. Furthermore, comparative experiments with other mainstream deep learning feature extractors show that the GAT-LSTM feature encoder used in this scheme achieves the highest association accuracy, as shown in the figure. Figure 7As shown.

[0164] For extremely complex scenarios where targets are spatially adjacent and have highly similar motion characteristics (such as...) Figure 8 As shown in the figure, the proposed method, through deep fusion of the spatiotemporal features of the tracks, can accurately identify and correctly distinguish real associated tracks. Compared with traditional methods, this method effectively overcomes the problem of association ambiguity in dense multi-target scenarios, achieves high-precision track association, and further verifies its reliability and application value in practical radar engineering systems.

[0165] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A compact ground-wave radar track association method, characterized in that, Includes the following steps: The track data of T / R monostatic radar and TR bistatic radar are acquired and preprocessed to obtain time-synchronized and spatially unified track pairs to be associated. The tracks are then split into multiple short track segments of equal length to form multiple track segment pairs. For each time step in the track segment, a local star topology graph is constructed with the target corresponding to the current track point as the master node and other targets in its spatial neighborhood as neighbor nodes. Multiple local graphs arranged in time sequence within the same track segment are then combined into a graph sequence to obtain a pair of "graph sequences to be compared". The graph sequence to be compared is input into a Siamese GAT-LSTM network, which has at least two GAT-LSTM encoders with shared parameters and identical structures. Each GAT-LSTM encoder includes a graph attention network unit and a long short-term memory network unit. The graph spatial features of each time step are extracted by the graph attention network unit, and the temporal features of the graph sequence from multiple time steps are fused by the long short-term memory network unit. Each branch encoder outputs a high-dimensional spatiotemporal feature vector representing the main target of the track segment. The high-dimensional spatiotemporal feature vectors output by the Siamese GAT-LSTM network are concatenated, and similarity estimation is performed to obtain the segment-level association similarity score. The original track pairs to be associated are split into multiple short track fragment pairs with overlapping time. All fragment pairs are traversed, and the above steps are performed on each fragment pair in sequence to obtain a set of similarity scores for all fragment pairs. Based on this set, a track-level association consistency index is calculated. If the index exceeds a preset threshold, the two tracks are determined to be associated and correspond to the same target.

2. The compact ground wave radar track association method as described in claim 1, characterized in that, The preprocessing includes: Interpolation algorithms are used for resampling to synchronize track data to a unified timeline; The original polar coordinate measurement data of the radar were uniformly converted to the geographic coordinate system; The target's speed and heading information are calculated based on the coordinate difference between adjacent time steps, forming a standardized track state sequence.

3. The compact ground wave radar track association method as described in claim 1, characterized in that, When constructing a local star topology, the total number of nodes is It includes 1 master node and a maximum of One neighboring node; If the number of neighboring targets in the neighborhood exceeds At that time, the targets are sorted according to their Euclidean distance, and only those closest to the main target are retained. If there are not enough neighbors, the node positions are filled with zero vectors.

4. The compact ground wave radar track association method as described in claim 1, characterized in that, The graph attention network unit employs a multi-head attention mechanism; for a master node in the graph, its spatial feature vector is formed by concatenating the weighted aggregation results calculated independently by each attention head, as shown in the following equation: ； in, Indicates the first Attention weights calculated from each attention head Represents the activation function of the exponential linear unit. express Time Neighbor Nodes The feature vectors, the number of attention heads are .

5. The compact ground wave radar track association method as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network unit is embedded and coupled with the Graph Attention (GAT) network unit to form a GAT-LSTM unit. At each time step, the GAT-LSTM unit simultaneously receives the graph structure information at the current time step and the hidden state and unit state at the previous time step, and performs joint updates on the spatial topology information and temporal recursive state through its internal gating mechanism.

6. The compact ground wave radar track association method as described in claim 1, characterized in that, Each branch encoder of the twin GAT-LSTM network includes GAT-LSTM units are stacked in a hierarchical order, and each layer recursively models the graph sequence. Among them, the Layer in Hidden state output at all times The input features of the next layer at the same time; the hidden state of the top-level GAT-LSTM unit at the last time step. This serves as a comprehensive spatiotemporal representation of the flight path segment.

7. The compact ground wave radar track association method as described in claim 1, characterized in that, The two high-dimensional spatiotemporal feature vectors are concatenated and their similarity is estimated. Specifically, the two high-dimensional spatiotemporal feature vectors are concatenated along their feature dimensions to form a joint feature vector. The joint feature vector is then input into a similarity estimation module, which is a feedforward neural network containing fully connected layers and activation functions. Its output is a segment-level association similarity score between 0 and 1.

8. The compact ground wave radar track association method as described in claim 7, characterized in that, The twin GAT-LSTM network and similarity estimation module are pre-trained using an intra-group contrastive learning strategy. The training strategy uses real associated track pairs as positive samples and constructs negative samples by replacing the main target with other targets in the spatial neighborhood. Each positive sample pair is combined with multiple negative sample pairs into a training group for optimization training. The sample with the highest segment-level association similarity score is judged as associated by the network.

9. The compact ground wave radar track association method as described in claim 1, characterized in that, Based on this set, a track-level correlation consistency index is calculated, specifically: Two paths to be linked are generated together in their overlapping region. There are 10 track segment pairs, of which the number of track segment pairs identified as related by the twin GAT-LSTM network is 10. The track-level correlation consistency index is: ; When the consistency index is satisfied , If a preset trajectory-level consistency association threshold is used, two trajectories are considered to originate from the same target and are therefore associated; otherwise, they are considered not associated.

10. A compact ground-wave radar track association system, characterized in that, include: The data acquisition and preprocessing module is configured to: acquire and preprocess track data from T / R monostatic radar and TR bistatic radar to obtain time-synchronized and spatially unified track pairs to be associated, and split the tracks into multiple short track segments of equal length to form multiple track segment pairs; The track map sequence module is configured to: for each time step in the track segment, take the target corresponding to the current track point as the master node and other targets in its spatial neighborhood as neighbor nodes, construct a local star topology map, and form a map sequence by arranging multiple local maps in the same track segment in time order, to obtain a pair of "map sequences to be compared"; The spatiotemporal feature encoding and fusion module is configured as follows: the graph sequence to be compared is input into a twin GAT-LSTM network, which has at least two GAT-LSTM encoders with shared parameters and identical structures. Each GAT-LSTM encoder includes a graph attention network unit and a long short-term memory network unit. The graph spatial features of each time step are extracted through the graph attention network unit, and the temporal features of the graph sequence of multiple time steps are fused through the long short-term memory network unit. Each branch encoder outputs a high-dimensional spatiotemporal feature vector representing the main target of the track segment. The track fragment association module is configured to: concatenate the high-dimensional spatiotemporal feature vectors output by the twin GAT-LSTM network and perform similarity estimation to obtain a fragment-level association similarity score; split the original track pairs to be associated into multiple short-term track fragment pairs with overlapping time; traverse all fragment pairs and perform the above steps sequentially for each fragment pair to obtain a set of similarity scores for all fragment pairs; calculate the track-level association consistency index based on this set; if the index exceeds a preset threshold, it is determined that the two tracks are associated and correspond to the same target.

Images