An ellipse multi-extent target tracking method and system based on a dynamic heterogeneous graph neural network

Abstract

This invention discloses an elliptical multi-extended target tracking method and system based on a dynamic heterogeneous graph neural network. The method includes: acquiring target information to be tracked; constructing a multi-extended target tracking model; inputting the target information to be tracked into the multi-extended target tracking model for processing to obtain tracking results; wherein, the multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and guides the global optimization of the model by designing an end-to-end multi-task loss function. This invention is the first to propose a dynamic heterogeneous graph representation for modeling the multi-extended target tracking process, which adaptively characterizes the complex evolution of heterogeneous information such as measurement, extended target state, and association in time and space through graph structure, overcoming the limitations of traditional vectorized representation.

Description

Technical Field

[0001] This invention relates to the field of radar target tracking technology, and in particular to an elliptical multi-extended target tracking method and system based on a dynamic heterogeneous graph neural network. Background Technology

[0002] Target tracking is one of the key tasks of modern radar systems. It aims to estimate the state (such as position and velocity) of an unknown number of targets in real time based on measurement data (which includes noise and false alarms) obtained from the radar receiver, and to continuously track their motion trajectories. Traditional multi-target tracking methods are mainly based on Bayesian filtering theory, such as Probability Hypothesis Density (PHD) filtering and Multiple Hypothesis Tracking (MHT). These methods model the target state as a random finite set (RFS) and propagate the posterior probability density through recursive Bayesian estimation to characterize the random appearance and disappearance of targets.

[0003] With the rapid development of modern sensing technology, another more challenging task is multiple extended target tracking (METT). Unlike idealized point targets, high-resolution radar can acquire the fine structure and scattering characteristics of targets. For targets with large volumes and complex structures, multiple resolution cells are typically covered, and a single target generates multiple radar measurement information. In addition to estimating its motion parameters, it is also necessary to estimate its shape and appearance (size, orientation, etc.). METT is an important research direction in the field of radar target tracking, aiming to solve the problems of detection, association, and state estimation when multiple extended targets with different geometric shapes exist simultaneously in real-world environments. For the problem of multiple extended target tracking, researchers have proposed a variety of solutions based on different approaches. Traditional methods are mostly based on the RFS framework, introducing shape parameters to characterize the extended target characteristics. The Gaussian Inverse Wishart PHD (GIW-PHD) filter was proposed, using the Gaussian Inverse Wishart distribution to model the kinematics and shape expansion of extended targets. Beard et al. proposed a labeled GIW-PHD filter that can output the trajectory of extended targets. They also extended the GLMB filter to track elliptical extended targets, modeling the joint distribution of target existence probability, motion state, and shape parameters. Xia et al. proposed a novel variant of the PMBM filter, modeling the target's scattering points as a random finite set, which can flexibly adapt to extended targets of different shapes. Although these extended target tracking methods based on RFS theory have solved the problem of representing and estimating multiple extended targets to some extent, they still inevitably face limitations such as high state dimension, large computational cost, and model dependence. Another approach is to treat the extended target as a set of interconnected scattering points, such as the method based on hierarchical Bayesian models proposed by Daniyan et al. This method first clusters the scattering points and then uses elliptic curves to fit the shape of the extended target. These methods simplify the problem to some extent, but they do not make sufficient use of the spatial correlation of scattering points.

[0004] In recent years, deep learning methods have achieved great success in the field of computer vision target tracking, inspiring researchers to explore their application in end-to-end solutions for multi-target tracking. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are used to detect and track targets end-to-end from image sequences. Furthermore, some researchers have proposed combining deep learning with random finite set filtering to learn density functions representing the states of multiple targets. While these works demonstrate the potential of a data-driven paradigm, they primarily focus on visual tracking and are difficult to directly apply to extended target tracking problems. In addition, most existing deep tracking architectures adopt a "black box" design, lacking interpretability and consideration of radar data characteristics and task structure, resulting in insufficient generalization ability.

[0005] In summary, multi-target tracking has long relied primarily on statistical modeling and Bayesian inference, which suffers from inherent limitations such as model dependence, computational high cost, and insufficient ability to handle dynamic changes. Deep learning, with its powerful feature extraction and function fitting capabilities, offers a new approach to this problem, but targeted and interpretable end-to-end methods are still lacking. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes an elliptical multi-extended target tracking method and system based on dynamic heterogeneous graph neural networks, which establishes a new end-to-end multi-extended target tracking paradigm based on dynamic heterogeneous graphs and integrates task structure and data representation.

[0007] On the one hand, to achieve the above objectives, this invention provides an elliptical multi-extended target tracking method based on a dynamic heterogeneous graph neural network, comprising:

[0008] Obtain information about the target to be tracked;

[0009] Construct a multi-extended target tracking model;

[0010] The target information to be tracked is input into the multi-extended target tracking model for processing to obtain the tracking result;

[0011] The multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and to use a graph neural network to perform end-to-end computation on the discrete-time dynamic heterogeneous graph to guide the global optimization of the model.

[0012] Preferably, constructing the multi-extended target tracking model includes:

[0013] The multi-extended target tracking process is modeled as a discrete-time dynamic heterogeneous graph, and a heterogeneous graph slice containing measurement nodes and target nodes is constructed in each time step. The heterogeneous graph slice includes measurement-measurement edges, measurement-target edges, and target-target edges.

[0014] The heterogeneous graph slices at each time step are encoded, and the embedding representations of different types of nodes are aggregated through a multi-head attention mechanism to generate updated node embeddings.

[0015] Temporal modeling is performed on the embedding sequence of the target node, and a target node representation containing long-term spatiotemporal dependencies is generated by combining the historical state memory matrix;

[0016] The measurement-target correlation probability matrix, target quantity prediction, and target state estimation are jointly decoded by the correlation prediction decoding module, wherein the correlation prediction decoding module includes a correlation calculation layer, a target quantity prediction layer, and a target state prediction layer.

[0017] Design a multi-task loss function to jointly optimize data association, target quantity estimation, and state filtering tasks, guide global model optimization, and obtain the multi-extended target tracking model.

[0018] Preferably, the weights of the measurement-measurement edges are calculated using a Gaussian kernel function to reflect the spatial proximity of the measurement nodes; the weights of the measurement-target edges are calculated using normalized Mahalanobis distance to measure the association probability between the measurement and the target node; and the weights of the target-target edges are calculated using normalized Gaussian Wasserstein distance to measure the similarity of the target nodes in terms of position and shape.

[0019] Preferably, encoding the heterogeneous map slice at each time step includes:

[0020] The measurement node and the target node are subjected to type-specific linear transformations through a heterogeneous graph attention network, and the embedded representations of heterogeneous neighbor nodes are aggregated through a multi-head attention mechanism.

[0021] The multi-head attention mechanism uses edge-type specific attention vectors to calculate attention coefficients between nodes and generates updated node embeddings through weighted aggregation.

[0022] Preferably, the processing procedure of the multi-head attention mechanism is as follows:

[0023]

[0024] In the formula, For the embedding of the i-th node in the l-th layer, || represents concatenation. A set of edge types, Let i be the set of neighbors of node i under edge class r. For attention weights, For node type φ j The transformation matrix, where K is the number of attention heads. Embed the j-th neighbor node in the (l-1)-th layer.

[0025] Preferably, temporal modeling of the embedding sequence of the target node includes:

[0026] A self-attention temporal aggregation module is used to perform temporal modeling on the embedding sequence of the target node. The self-attention temporal aggregation module is used to add sinusoidal positional encoding to the target node embedding to generate a positional encoding vector. The positional encoding vector interacts with the historical state memory matrix through a dot product attention mechanism to dynamically fuse historical temporal information and generate the target node representation containing long-term spatiotemporal dependencies. The historical state memory matrix is ​​used to store the historical embeddings of the target node and is updated through a sliding window mechanism to limit the computational complexity.

[0027] Preferably, the correlation calculation layer is used to embed the measurement node and the target node into a common space through linear transformation, and to generate the measurement-target correlation probability matrix through matrix multiplication and the sigmoid function;

[0028] The target quantity prediction layer is used to predict the total number of targets at the next moment based on the current measurement information and historical target information.

[0029] The target state prediction layer is used to combine associated probability weighted measurement information, target node embedding, and time series aggregation embedding to obtain the predicted state vector through multilayer perceptron mapping.

[0030] Preferably, the multi-task loss function includes binary cross-entropy loss, class cross-entropy loss, mean absolute error loss, and Gauss-Wasesterstein distance loss;

[0031] The calculation of the Gauss-Wasestein distance loss is as follows:

[0032] The discrete measurement point set is transformed into a probability distribution using the Dirac measure;

[0033] For each discrete point, calculate the Wasserstein distance between the discrete point as a Gaussian distribution with mean and variance equal to an identity matrix and the target elliptical Gaussian distribution;

[0034] The average Wasserstein distance of all points is used as a measure of the relationship between the point set and the elliptic distribution:

[0035] Based on the Gauss-Wasestein distance formula, the distribution difference between each measurement point and the target ellipse is calculated, and the mean value is taken as the final loss term.

[0036] On the other hand, to achieve the above objectives, the present invention also provides an elliptical multi-extended target tracking system based on a dynamic heterogeneous graph neural network, comprising:

[0037] Information acquisition unit: used to acquire information about the target to be tracked;

[0038] Model building unit: used to build multi-extended target tracking models;

[0039] Tracking processing unit: used to input the target information to be tracked into the multi-extended target tracking model for processing to obtain tracking results; wherein, the multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and guides the global optimization of the model by designing an end-to-end multi-task loss function.

[0040] Preferably, the model building unit includes:

[0041] Dynamic heterogeneous graph modeling module: used to construct heterogeneous graph slices containing measurement nodes and target nodes in real time, wherein the heterogeneous graph slices include measurement-measurement edges, measurement-target edges, and target-target edges;

[0042] Heterogeneous graph encoding module: used to encode the heterogeneous graph slices at each time step, and aggregate the embedding representations of different types of nodes through a multi-head attention mechanism to generate updated node embeddings;

[0043] Temporal aggregation module: used to perform temporal modeling on the embedding sequence of the target node, and generate a target node representation containing long-term spatiotemporal dependencies by combining the historical state memory matrix;

[0044] Association prediction decoding module: used to jointly output measurement-target association probability, target quantity, and state estimation;

[0045] Multi-task optimization module: Used to design multi-task loss functions, jointly optimize data association, target quantity estimation and state filtering tasks, guide global model optimization, and obtain the multi-extended target tracking model.

[0046] Compared with the prior art, the present invention has the following advantages and technical effects:

[0047] (1) This invention proposes for the first time a dynamic heterogeneous graph representation for modeling the multi-extended target tracking process. It adaptively characterizes the complex evolution of heterogeneous information such as measurement, extended target state, and association in time and space through graph structure, overcoming the limitations of traditional vectorized representation.

[0048] (2) This invention is the first to apply graph attention neural networks to the field of target tracking, achieving end-to-end multi-target tracking. It jointly optimizes target detection, data association, and state filtering, fully exploiting the coupling dependencies between different tasks and improving tracking performance. In particular, heterogeneous graph encoding utilizes type-specific attention aggregation to flexibly model the interaction patterns of measurements and targets under different semantics. The self-attention temporal aggregation mechanism enhanced by external memory enables the current target state estimation to pay attention to historical state information, making up for the shortcomings of existing deep tracking methods based on temporal sequential processing.

[0049] (3) This invention provides a novel technical approach for modeling, solving, and evaluating multi-target tracking problems, and is expected to promote theoretical innovation and engineering practice in the field of radar target tracking. The developed end-to-end tracking algorithm is universal and can be applied not only to ground surveillance radars but also to a wider range of sensor platforms and application scenarios such as airborne early warning and sea surface search. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a schematic diagram of dynamic heterogeneous graph modeling for multi-extended target tracking according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the graph coding module of the multi-extended target tracking model according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the temporal aggregation module of the multi-extended target tracking model in an embodiment of the present invention;

[0054] Figure 4This is a schematic diagram of the association prediction and decoding module of the multi-extended target tracking model in an embodiment of the present invention;

[0055] Figure 5 This is an experimental comparison diagram of the GWD index for matching between a set of measurement points and an elliptical target in an elliptical multi-extended target tracking algorithm according to an embodiment of the present invention. In the diagram, (a) is a schematic diagram of a set of measurement points distributed outside the elliptical target with 100 measurement points, (b) is a schematic diagram of a set of measurement points distributed both inside and outside the elliptical target with 100 measurement points, (c) is a schematic diagram of a set of measurement points only inside the elliptical target with 100 measurement points, (d) is a schematic diagram of a set of measurement points distributed inside the elliptical target with 1000 measurement points, (e) is the GWD index of a set of measurement points with 100 measurement points under three distribution methods, and (f) is a schematic diagram of the numerical change of the index when the set of measurement points is distributed inside the elliptical target and the number of measurement points is continuously increased.

[0056] Figure 6 The following is a comparison chart of the actual tracking results and evaluation indicators of the multi-extended target tracking method according to the embodiments of the present invention, wherein (a) is a schematic diagram of the trajectory, (b) is a schematic diagram of the GWD indicator, and (c) is a schematic diagram of the IoU indicator;

[0057] Figure 7 This is a flowchart of an elliptical multi-extended target tracking method based on a dynamic heterogeneous graph neural network according to an embodiment of the present invention. Detailed Implementation

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0060] This embodiment proposes an elliptical multi-extended target tracking method based on a dynamic heterogeneous graph neural network, such as... Figure 7 ,include:

[0061] Obtain information about the target to be tracked;

[0062] Construct a multi-extended target tracking model;

[0063] The target information to be tracked is input into the multi-extended target tracking model for processing to obtain the tracking result;

[0064] Among them, the multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and to use a graph neural network to perform end-to-end computation on the discrete-time dynamic heterogeneous graph to guide the global optimization of the model.

[0065] Specifically, in this embodiment, we consider a tracking scenario containing M extended targets. The kinematic state and shape parameters of the j-th target can be represented by a state vector. To indicate:

[0066]

[0067] in, and This represents the position coordinates of target j at time t. and For the corresponding velocity components, The direction angle of the target, and and These are the major and minor axes of the ellipse, respectively.

[0068] At each time t, the sensor returns a set of measurements. in Let n represent the position coordinates of the i-th measurement. (t) Let be the total number of measurements at time t. Measurement set. Typically, it includes real measurements generated by M real targets (each target may generate multiple measurements) and spurious measurements (i.e., false alarms) generated by environmental clutter, but these cannot be directly distinguished. Multi-extended target tracking aims to base its analysis on a set of measurements over a series of time steps. Real-time estimation of the state of each target at every moment And establish the correspondence between measurements and targets. Among them, Let t be the estimated total number of targets. Let be the state estimate of the j-th target at time t. This embodiment uses a graph neural network to solve this problem, therefore, the tracking process first needs to be modeled as a graph.

[0069] Furthermore, constructing a multi-extended target tracking model includes:

[0070] The multi-extended target tracking process is modeled as a discrete-time dynamic heterogeneous graph, and a heterogeneous graph slice containing measurement nodes and target nodes is constructed in each time step. The heterogeneous graph slice includes measurement-measurement edges, measurement-target edges, and target-target edges.

[0071] The heterogeneous graph slices at each time step are encoded, and the embedding representations of different types of nodes are aggregated through a multi-head attention mechanism to generate updated node embeddings.

[0072] Temporal modeling is performed on the embedding sequence of the target node, and a target node representation containing long-term spatiotemporal dependencies is generated by combining the historical state memory matrix;

[0073] The measurement-target correlation probability matrix, target quantity prediction, and target state estimation are jointly decoded by the correlation prediction decoding module, which includes a correlation calculation layer, a target quantity prediction layer, and a target state prediction layer.

[0074] Design a multi-task loss function to jointly optimize data association, target quantity estimation, and state filtering tasks, guide global model optimization, and obtain a multi-extended target tracking model.

[0075] Specifically, such as Figure 1 The entire tracking process is modeled as a discrete-time dynamic heterogeneous graph. Where T is the total number of tracking time steps. For each time step t, the dynamic graph... Contains a heterogeneous subgraph This is called a graph slice, representing the current tracking state. Each graph slice consists of two types of nodes: measurement nodes. and state nodes These correspond to the target state measured and estimated by the sensors at the current moment, respectively; they also include three types of edges: measurement-measurement edges, measurement-state edges, and state-state edges, which respectively characterize the interaction relationships between different nodes.

[0076] For image slices Its node set is It includes two types of heterogeneous nodes, defined as follows:

[0077] Measurement node set in, Let n represent the i-th measurement node. (t) Let t be the total number of measurements taken at time t.

[0078] Each measurement node Carry a feature vector Initialize to the corresponding measurement position coordinates:

[0079]

[0080] target node set in Let m represent the state estimate of the j-th target node at time t. (t) The total number of targets estimated at time t. Each target node... Carry a feature vector Initialize to the corresponding target state estimate:

[0081]

[0082] In the formula, and Let represent the estimated position coordinates of target j at time t. and These are the estimated velocity components in the x-axis and y-axis directions, respectively. These represent the orientation angle, major semi-axis, and minor semi-axis of the elliptical target. Let be the state estimate of the j-th target at time t.

[0083] Image slice It contains three types of semantic edges, each connecting a different subset of nodes:

[0084] Measurement - Measurement Edge Set in, Indicates measurement node and measurement nodes Undirected edges between them.

[0085] Measurement - Measurement edge weights Defined as a Gaussian kernel function:

[0086]

[0087] Here, ||·||² represents the Euclidean distance. The physical meaning of this weight design is that the closer two measurements are in space, the greater the probability that they come from the same target.

[0088] Measurement - Target Edge Set in; Indicates measurement node and target node Undirected edges between them.

[0089] Measurement - Weight of the target edge Defined as the normalized Mahalanobis distance:

[0090]

[0091] Among them, Mahalanobis distance The definition of is:

[0092]

[0093] in, To estimate the position covariance matrix of the j-th target at time t, it can be estimated based on its shape parameters. Perform the calculation:

[0094]

[0095] in, Representing a two-dimensional rotation matrix:

[0096]

[0097] Mahalanobis distance is a weighted Euclidean distance that takes into account the uncertainty of the target, when the measurement position... With target state estimation The closer the positional components are in the Markovian sense, the greater the edge weight between them, indicating a higher degree of correlation between them.

[0098] Target-target edge set in Represents the target node and Undirected edges between them.

[0099] Weight of the target-target edge Defined as the normalized Gaussian Wasserstein Distance (GWD):

[0100]

[0101] Wherein, GWD(·,·) is defined as:

[0102]

[0103] Here, s1 and s2 are two target state vectors, s1[[x,y]] and s2[[x,y]] represent their position components, and Σ1 and Σ2 are the corresponding position covariance matrices.

[0104] The Gauss-Wasesterstein distance takes into account the differences in position and shape between two targets, and is a robust similarity measure. Intuitively, the closer two targets are in spatial location and the more similar their shapes are, the stronger their interaction is.

[0105] In summary, slicing The set of edges can be represented as:

[0106]

[0107] Accordingly, three adjacency matrices are defined. and To represent the weights of different types of edges:

[0108]

[0109] Here's a slice of the image. Since it is a fully connected graph, the adjacency matrix is... and n (t) ×n(t) n (t) ×m (t) and m (t) ×m (t) A dense matrix.

[0110] Furthermore, encoding the heterogeneous graph slice at each time step includes:

[0111] The measurement node and the target node are subjected to type-specific linear transformations through a heterogeneous graph attention network, and the embedded representations of heterogeneous neighbor nodes are aggregated through a multi-head attention mechanism.

[0112] The multi-head attention mechanism uses edge-type specific attention vectors to calculate attention coefficients between nodes and generates updated node embeddings through weighted aggregation.

[0113] Specifically, this embodiment designs a Heterogeneous Graph Attention Network (HGAT) as the main tool for graph encoding. Compared with traditional graph neural networks, HGAT has the advantage of being able to distinguish different types of nodes and edges, and model their feature evolution and interaction patterns separately, thereby better adapting to complex data structures in multi-target tracking scenarios.

[0114] Construction of heterogeneous graphs:

[0115] Definition diagram The set of nodes is Among them, the set of measurement nodes Each node corresponds to a measurement vector. target node set Each node corresponds to a target state estimate. m t The total number of targets estimated at time t-1.

[0116] Definition diagram The set of edges is in, For measurement-to-measurement edges, connect all measurement node pairs, with edge weights. This is a Gaussian kernel function that measures the geometric similarity between nodes. For the measurement-target edge, connect all measurement-target node pairs, and assign edge weights. The normalized Mahalanobis distance measures the strength of the association between nodes. For each target-target edge, connect all target node pairs, and assign edge weights. The normalized Gauss-Wasestein distance measures the interaction patterns between nodes.

[0117] Heterogeneous graph attention layer:

[0118] The graph encoding module employs a stacked L-layer heterogeneous graph multi-head attention layer (HGMAL) to learn node embeddings. The input to the l-th HGMAL layer is the node embedding matrix from the previous layer. The output is the updated node embedding matrix. Initial embedding matrix H (t,0) These are the original features of the node.

[0119] The calculation process for HGMAL is as follows:

[0120] Type-specific transformation: For nodes i of different types φ∈{z,s}, apply a type-specific transformation matrix. Embed it Mapped to K d l 3D feature space:

[0121]

[0122] Heterogeneous graph attention coefficients: For each type of edge r∈{zz,zs,ss}, use a type-specific attention vector. Calculate the attention coefficient between nodes i and j, and then normalize it using softmax:

[0123]

[0124] in, Let i be the set of neighboring nodes of node i under edge class r. Let be the attention scores of nodes i and j of edge type r in the l-th layer of feature space k. Let j be the feature node j in the l-th layer of feature space k at time t that participates in the attention score calculation. Let be all feature nodes connected by edges of type r at time t. Let i be the attention score between node i in the l-th layer of feature space k with edge type r and its neighbor node n.

[0125] Heterogeneous neighbor aggregation: Node i is aggregated through attention weights Aggregate the neighbor information of all K heads and all r-type edges to obtain the updated node representation:

[0126]

[0127] Here, ‖ represents vector concatenation.

[0128] In summary, the core of the graph coding module is the Heterogeneous Graph Multi-head Attention Layer (HGMAL), which fuses local structural and semantic information by distinguishing the interaction patterns of different types of nodes and edges. Figure 2 .

[0129] The calculation of HGMAL can be summarized as follows:

[0130]

[0131] in, For the embedding of the i-th node in the l-th layer, || represents concatenation. A set of edge types, Let i be the set of neighbors of node i under edge class r. For attention weights, For node type φ j The transformation matrix, where K is the number of attention heads. Embed the j-th neighbor node in the (l-1)-th layer.

[0132] After stacking L layers of HGMAL, the final node embedding matrix can be obtained:

[0133]

[0134] in, Embed a submatrix for the measurement nodes. H is a submatrix embedded for the target node. (t,L) Let be the node embedding matrix of layer L at time t. These embedding vectors integrate the structural information and heterogeneous interaction patterns of the graph. The output of the graph encoding will be fed into the subsequent temporal aggregation and association prediction decoding modules for subsequent temporal modeling and tracking tasks.

[0135] Furthermore, temporal modeling of the embedding sequence of the target node includes:

[0136] A self-attention temporal aggregation module is used to perform temporal modeling on the embedding sequence of the target node. The self-attention temporal aggregation module is used to add sinusoidal positional encoding to the target node embedding to generate a positional encoding vector. The positional encoding vector interacts with the historical state memory matrix through a dot product attention mechanism to dynamically fuse historical temporal information and generate the target node representation containing long-term spatiotemporal dependencies. The historical state memory matrix is ​​used to store the historical embeddings of the target node and is updated through a sliding window mechanism to limit the computational complexity.

[0137] Specifically, such as Figure 3In multi-target tracking tasks, target states often exhibit complex dependencies and evolution patterns over time. To better model and utilize this temporal information, the self-attention temporal aggregation module, based on the graph coding module, dynamically aggregates target state node embeddings from different times using a self-attention mechanism, generating node representations that contain historical contextual information.

[0138] The self-attention temporal aggregation module mainly consists of three parts: position encoding, memory matrix update, and self-attention aggregation.

[0139] Location coding:

[0140] To incorporate time step information into node embedding, the Sinusoidal Positional Encoding (SPE) function SPE(t,m,d) is first used to generate a unique d-dimensional positional encoding vector for the m-th target node at time t:

[0141] SPE(t,m,2i)=sin(t m / 10000 2i / d );

[0142] SPE(t,m,2i+1)=cos(t m / 10000 2i / d );

[0143] Among them, t m =t+m / M t Let be the timestamp of node m, and i be the dimension index. The position encoding vector is added point-by-point to the node embedding to obtain a node representation matrix incorporating temporal information.

[0144]

[0145] Memory matrix update:

[0146] Before performing attention aggregation, the node at the current moment needs to be embedded. Added to the memory matrix to construct the complete temporal context:

[0147]

[0148] Here, [·; ·] represents a concatenation operation on the first dimension (time dimension). There is no X here. (t-1) Instead of embedding historical nodes in the calculation, it directly reuses previous results to improve computational efficiency.

[0149] Self-attention convergence:

[0150] The goal of self-attention aggregation is to enable each node at the current moment to adaptively focus on information about related nodes in the history, forming a context-aware node representation.

[0151] Specifically, for the m-th node at time t, the query vector key vector Sum value vector The calculation is as follows:

[0152]

[0153] in, All are learnable weight matrices.

[0154] Then, node m aggregates information from historical nodes through an attention mechanism:

[0155]

[0156]

[0157] in, and They are memory matrices X (t) The key sequence and value sequence, Let be the attention weights between node m and historical node n. This ultimately yields the aggregated embedding of the m-th node.

[0158] All M t The aggregation and concatenation of the node embeddings yields the target state node embedding matrix at time t, which incorporates historical information.

[0159]

[0160] In the formula, For time t, the Mth... t Aggregation embedding of nodes.

[0161] Based on the above calculations, the output is... This implies the spatiotemporal dependency between the current moment and historical moments, which can better guide subsequent correlation and prediction tasks. At the same time, the calculation process is differentiable, allowing end-to-end gradient propagation and optimization.

[0162] Assume that at time t, the average number of targets per time step is... Then the memory matrix X (t) The time complexity is The space complexity is The time complexity of self-attention computation is . The space complexity is Output matrix Both have a time and space complexity of O(M). t d v ).

[0163] Therefore, the total time complexity is The space complexity is Although the computational overhead increases linearly with time step t, it can still meet real-time requirements through parallelization and optimized coding. Furthermore, d q ,d k ,d v The number of heads is generally much smaller than d, and the number of heads can also be controlled within a small constant range, so the overall complexity is acceptable.

[0164] Furthermore, the correlation computation layer is used to embed the measurement node and the target node into a common space through linear transformation, and to generate the measurement-target correlation probability matrix through matrix multiplication and the sigmoid function;

[0165] The target quantity prediction layer is used to predict the total number of targets at the next moment based on current measurement information and historical target information;

[0166] The target state prediction layer combines associated probability weighted measurement information, target node embedding, and time series aggregation embedding to obtain the predicted state vector through multilayer perceptron mapping.

[0167] Specifically, such as Figure 4 The Association Prediction and Decoding Module (APDM) consists of three sub-layers: Association Computation Layer (ACL), Target Number Prediction Layer (TNPL), and Target State Prediction Layer (TSPL).

[0168] Association Computation Layer (ACL):

[0169] The purpose of the association computation layer is to calculate the association probability between each pair of measurement-target nodes at the current time, and obtain a two-dimensional association probability matrix. Intuitively, the (i,j)th element of the matrix This indicates the probability that the i-th measurement is generated by the j-th target. This information is crucial for subsequent target state updates, because only by correctly associating the measurement with the target can the measurement information be used to correct and improve the target state estimate.

[0170] ACL embeds the measurement nodes output by the graph coding module into the matrix. and target node embedding matrix As input, they are first mapped to a common d through a linear transformation. a In a 3D space, the association probability is calculated using matrix multiplication and the sigmoid function:

[0171]

[0172] in, and These are the learnable mapping parameters for measurement and target embedding, respectively. The sigmoid function compresses the original association scores to the (0,1) interval, yielding normalized probability values.

[0173] This calculation process can be viewed as applying attention weights to the target node using measurement nodes, with the association probability matrix representing these attention weights. Unlike prior association methods such as Global Nearest Neighbor (GNN), the association matrix learned by ACL is data-driven, capable of modeling complex association patterns, and possesses greater adaptability.

[0174] Target Quantity Prediction Layer (TNPL):

[0175] The task of the target quantity prediction layer is to predict the total number of targets at the next moment based on current measurement information and historical target information. Accurately estimating the number of targets is crucial for multi-target tracking, as it determines how many target states the tracker needs to maintain and how to handle newly appearing and disappearing targets.

[0176] TNPL uses a classification approach to predict the number of targets. Specifically, in this embodiment, a prior maximum number of targets M is first set. max Discretize the possible number of targets into M max +1 category (including 0). Then, embed the measurement node at time t. Target node embedding Embedded with target nodes after time-series aggregation The nodes are averaged and concatenated into a 3D feature vector. Then, a multilayer perceptron (MLP) and a softmax function are used to calculate the probability distribution of each category.

[0177]

[0178] Where ColMean(·) represents column mean aggregation of the input matrix, || represents vector concatenation, and MLP M (·) denotes a multilayer perceptron with learnable parameters. and These are the parameters of the quantity classification header. Let be the probability distribution of the number and category of targets at time t. The probability of the number of targets at time t+1.

[0179] Finally, the category with the highest probability is selected as the number of targets to be predicted:

[0180]

[0181] The input to TNPL includes not only the local features extracted by the graph coding module. and It also incorporates global context information learned by the time-series aggregation module. This allows the model to consider not only observational evidence at the current moment but also information from the entire tracking history when predicting the number of targets, thus exhibiting a certain degree of predictability and stability. Furthermore, transforming the continuous quantity estimation problem into a classification problem reduces the learning difficulty of the model.

[0182] Target State Prediction Layer (TSPL):

[0183] The purpose of the target state prediction layer is to estimate the state vector of each target at the next time step. in Compared to classic tracking methods, TSPL is characterized by the fact that it does not predict each target in isolation, but rather makes full use of the target-measurement associations and target-target interactions learned by the graph coding module, as well as the cross-temporal dependencies captured by the temporal aggregation module, and comprehensively considers the all-to-all spatiotemporal context to obtain more accurate and consistent prediction results.

[0184] Specifically, for the j-th target, TSPL first determines the target based on the j-th column of the association probability matrix. Embedding matrix of measurement nodes By performing a weighted summation, the aggregated measurement information of the target is obtained.

[0185]

[0186] Then, Graph embedding with target node j and temporal aggregation embedding The data is concatenated and mapped to the state space using another multilayer perceptron to obtain the predicted state vector.

[0187]

[0188] Among them, MLP s (·) represents another multilayer perceptron with learnable parameters W. s ,b s .

[0189] It can be seen that the prediction of the target state comprehensively utilizes three aspects of information: through Weighted measurement information This reflects the current observational evidence related to target j; the graph embedding of target j itself. It encodes its interaction patterns with other targets and measurements at time t; the temporal aggregation embedding of target j. It contains its state evolution and global dependencies in past moments.

[0190] The convergence of these three information streams allows TSPL to achieve a balance between current observations and past trajectories, enabling more robust and consistent predictions. Furthermore, the multilayer perceptron endows the model with the ability to model nonlinearities, allowing it to fit complex state transition functions.

[0191] Ultimately, the output of APDM includes: the association probability matrix P. (t) , the predicted number of targets and state estimation for each target These results will be fed back to the graph encoding module at the next time step to construct a new heterogeneous graph, forming a cyclical tracking process.

[0192] Furthermore, the multi-task loss function includes binary cross-entropy loss, class cross-entropy loss, mean absolute error loss, and Gauss-Wasesterstein distance loss;

[0193] The calculation of the Gauss-Wasestein distance loss is as follows:

[0194] The discrete measurement point set is transformed into a probability distribution using the Dirac measure;

[0195] For each discrete point, calculate the Wasserstein distance between the discrete point as a Gaussian distribution with mean and variance equal to an identity matrix and the target elliptical Gaussian distribution;

[0196] The average Wasserstein distance of all points is used as a measure of the relationship between the point set and the elliptic distribution:

[0197] Based on the Gauss-Wasestein distance formula, the distribution difference between each measurement point and the target ellipse is calculated, and the mean value is taken as the final loss term.

[0198] Specifically, associated losses:

[0199] For the association probability matrix output by the Association Computation Layer (ACL) This embodiment uses binary cross-entropy loss to measure its correlation with the true correlation matrix. Differences between them:

[0200]

[0201] in, This indicates that the i-th measurement is truly associated with the j-th target. This loss function, which indicates no correlation, encourages the model to learn a soft probability matrix consistent with the true correlation.

[0202] Target quantity loss:

[0203] The target quantity probability distribution output by the Target Quantity Prediction Layer (TNPL) Cross-entropy loss is used to measure its correlation with the one-hot encoding of the true number of targets. Differences between them:

[0204]

[0205] in, It is a one-hot vector, where only the element corresponding to the actual number of targets is 1, and the rest are 0. This loss function enables the model to learn to accurately predict the total number of targets in the next time step.

[0206] Loss in target state:

[0207] For each target state estimate output by the Target State Prediction Layer (TSPL) The mean absolute error (MAE) and Gaussian Wasserstein distance (GWD) are combined to measure its resemblance to the true target state. Differences between them:

[0208]

[0209] in, For MAE loss:

[0210]

[0211] For GWD losses:

[0212]

[0213] in, These represent the position components of the predicted and actual target states, respectively. Let λ represent the position covariance matrices of the predicted and actual target states, respectively, tr(·) denotes the trace of the matrix, and λ mae and λgwd Weighting coefficients to balance the two types of losses.

[0214] Multitasking loss:

[0215] The weighted sum of the three loss functions above yields the multi-task loss of the association prediction and decoding module at time t:

[0216]

[0217] Where, λ ac ,λ num ,λ state The weights for each loss term can be adjusted based on the importance of the task and the numerical scale. During training, the losses over the entire time series are accumulated, and an L2 regularization term for all learnable parameters (denoted as Θ) is added to obtain the final optimization objective:

[0218]

[0219] Where, λ Θ is the regularization coefficient, and T is the total time step.

[0220] This embodiment also provides an elliptical multi-extended target tracking system based on a dynamic heterogeneous graph neural network, including:

[0221] Information acquisition unit: used to acquire information about the target to be tracked;

[0222] Model building unit: used to build multi-extended target tracking models;

[0223] Tracking processing unit: used to input the target information to be tracked into the multi-extended target tracking model for processing to obtain tracking results; wherein, the multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and guides the global optimization of the model by designing an end-to-end multi-task loss function.

[0224] Specifically, the model building unit includes:

[0225] Dynamic heterogeneous graph modeling module: used to construct heterogeneous graph slices containing measurement nodes and target nodes in real time, wherein the heterogeneous graph slices include measurement-measurement edges, measurement-target edges, and target-target edges;

[0226] Heterogeneous graph encoding module: used to encode the heterogeneous graph slices at each time step, and aggregate the embedding representations of different types of nodes through a multi-head attention mechanism to generate updated node embeddings;

[0227] Temporal aggregation module: used to perform temporal modeling on the embedding sequence of the target node, and generate a target node representation containing long-term spatiotemporal dependencies by combining the historical state memory matrix;

[0228] The time-series aggregation module also includes an external storage submodule, used to dynamically maintain the embedding matrix of historical target nodes and to achieve cross-time step information interaction through a self-attention mechanism.

[0229] Association prediction decoding module: used to jointly output measurement-target association probability, target quantity, and state estimation;

[0230] Multi-task optimization module: Used to design multi-task loss functions, jointly optimize data association, target quantity estimation and state filtering tasks, guide global model optimization, and obtain multi-extended target tracking model.

[0231] To more clearly illustrate the technical solution of the present invention, specific embodiments are provided below for description:

[0232] Input: Measurement sequence Initial state estimation S (0) .

[0233] Step 1: Initialize the memory matrix Initial hidden state

[0234] Step 2: for t = 1 to T;

[0235] Step 2.1: The graph encoding module processes the measurement sequence at time t. and state estimation S (t-1) Output the target measurement node embedding at that moment. and state node embedding

[0236] Step 2.2: The time-series aggregation module processes the embedding of target state nodes. Embedded memory matrix X of historical moment nodes (t-1) Output the target state node embedding matrix at time t, which incorporates historical information. And update the memory matrix X (t) .

[0237] Step 2.3: Correlation prediction decoding module processing Output the correlation probability matrix P (t) Target quantity prediction and the state estimate S for each target (t+1) . , S (t+1) .

[0238] Step 3: Output tracking results

[0239] Calculation process of the image encoding module:

[0240] Input: Measurement set State estimation S (t-1) .

[0241] Step 1: Construct a heterogeneous graph

[0242] Step 1.1: Add Measurement Nodes and target node

[0243] Step 1.2: Add Measurement - Measurement Edge Calculate weights

[0244] Step 1.3: Add Measurement - Target Edge Calculate weights

[0245] Step 1.4: Add target - target edge Calculate weights

[0246] Step 2: Initialize node features

[0247] Step 3: for l = 1 to L;

[0248] Step 3.1: Type-Specific Transformation In the formula, Let be the transformation matrix. and The node embeddings before and after the transformation are respectively. For node set type, K is the number of attention channels.

[0249] Step 3.2: Edge type-specific attention:

[0250] In the formula, For attention score, For attention weights, For edge type, Where K is the node type and K is the number of attention channels.

[0251] Step 3.3: Heterogeneous Neighbor Aggregation Based on attention score To embed the aggregation node For the aggregated embedding, For node type.

[0252] Output: Measurement node embedding Target node embedding

[0253] The calculation process of the time series aggregation module:

[0254] Input: target node embedding Memory Matrix X (t-1) .

[0255] Step 1: Location Encoding In the formula, For embedding m state nodes at time t, SPE(t,m,d) represents the d-dimensional temporal position encoding of the m state nodes at time t.

[0256] Step 2: Memory Storage

[0257] Step 3: Self-attention convergence

[0258] Step 3.1: W is the learnable weight matrix, K (t) and V (t) They are memory matrices X (t) The key sequence and value sequence.

[0259] Step 3.2: for m = 1 to m t ;

[0260] Step 3.2.1: In the formula, q is the query vector and W is the weight matrix.

[0261] Step 3.2.2:

[0262] Step 3.2.3: In the formula, This represents the aggregate embedding of the m-th node.

[0263] Step 3.3: In the formula, Let t be the embedding matrix of the target state node that incorporates historical information.

[0264] Output: Aggregate Embedding Updated memory matrix X (t) .

[0265] The calculation process of the correlation prediction decoding module:

[0266] Input: Measurement node embedding Target node embedding Aggregation Embedding

[0267] Step 1: Connect the computation layer:

[0268] Step 1.1: W az and b az These are the learnable mapping parameters of the measurement embedding.

[0269] Step 1.2: W as and b as These are the learnable mapping parameters of the target embedding.

[0270] Step 1.3: P (t) For the correlation probability value

[0271] Step 2: Target Quantity Prediction Layer:

[0272] Step 2.1: Let be the probability distribution of the number and category of targets at time t.

[0273] Step 2.2: W represents the probability of the number of targets at time t+1. pM and b pM These are learnable parameters.

[0274] Step 2.3: The target quantity is predicted.

[0275] Step 3: Target State Prediction Layer

[0276] Step 3.1: for j = 1 to m t ;

[0277] Step 3.1.1: This represents the aggregated measurement information of the j-th target. is the j-th column of the correlation probability matrix.

[0278] Step 3.1.2:

[0279] Step 3.1.3:

[0280] In the formula, MLP(·) represents another multilayer perceptron, || represents the stitching operation, and W s and b s The learnable parameters are, This is an intermediate result. This is the predicted state vector.

[0281] Step 3.2:

[0282] Output: Association probability matrix P (t) Target quantity estimation Target state estimation S (t+1) .

[0283] Model hyperparameter configuration:

[0284] Graph encoding module: Original feature dimension d of measurement nodes z =2, the original feature dimension d of the target node s =7, embedding dimension d=256, number of attention heads K=8, number of heterogeneous graph attention layers L=3;

[0285] Time-series aggregation module: Location encoding dimension d en =256, query / key / value matrix dimension d q ,d k ,d v =64,64,256, Maximum time step T of the memory matrix mem =100;

[0286] Association prediction decoding module: mapping matrix dimension d az =256,d as =256, maximum target quantity M max =20, the target quantity prediction MLP structure is 256→128→64→M max +1, the target state prediction MLP structure is 256→128→64→d s The activation function of the MLP is ReLU;

[0287] Model training hyperparameters: Optimizer is Adam, initial learning rate is 0.001, batch size is 32, number of training epochs is 1000, and the loss function weights are λ. ac =1.0,λ num =0.1,λ state =1.0, regularization coefficient λ Θ =0.0001, λ mae =0.5,λ gwd =0.5.

[0288] This invention also provides a metric for elliptical multi-extended target tracking algorithms; please refer to [reference needed]. Figure 5 It is used for the matching metric between the measurement point set and the elliptical target.

[0289] Specifically:

[0290] Let the point set be The ellipse is E: The goal is to compute the Wasserstein distance between a discrete point set P and a continuous Gaussian (elliptical) distribution N(μ,Σ). Let N(μ,Σ) represent n points in a two-dimensional space, and let N(μ,Σ) represent a two-dimensional Gaussian distribution with mean μ and covariance matrix Σ, which geometrically corresponds to an ellipse.

[0291] First, the discrete point set P is represented as a probability distribution using the Dirac measure:

[0292]

[0293] in, Indicates that it is located at point x i The Dirac measure at point n. This means that the probability mass of each point is uniformly distributed across all points, with each point having a probability mass of 1 / n.

[0294] The Gaussian distribution of an ellipse is represented as:

[0295]

[0296] Where |Σ| represents the determinant of the covariance matrix Σ.

[0297] The Wasserstein distance between two Gaussian distributions is defined as:

[0298]

[0299] Where N1~N(μ1,Σ1) and N2~N(μ2,Σ2) are two Gaussian distributions, and tr(·) represents the trace of the matrix.

[0300] Since the point set P is not a truly continuous distribution, the Wasserstein distance between it and the Gaussian distribution N(μ,Σ) cannot be directly calculated using the above formula. Therefore, for each point x... i Consider it as a mean x i The covariance matrix is ​​a Gaussian distribution N(x) ∈ I. i Let , ∈ I), where ∈ is a very small positive number, and I is the identity matrix. Then, the Wasserstein distance between each such Gaussian distribution and the target Gaussian distribution N(μ,Σ) can be calculated, and their average can be taken as an approximation of the Wasserstein distance between the point set P and N(μ,Σ):

[0301]

[0302] For each point x i The above formula can be used to calculate

[0303]

[0304] Further simplification yields:

[0305]

[0306] Adding the contributions of each point, we get:

[0307]

[0308] As ∈ approaches zero, we obtain the final approximate expression:

[0309]

[0310] The simulation conditions for this embodiment are:

[0311] System: 64-bit Windows 11 operating system;

[0312] CPU: 12th Gen Intel(R)Core(TM)i5-12400 2.50GHz;

[0313] GPU: NVIDIA GeForce RTX 3090Ti;

[0314] Python: 3.7.4;

[0315] PyTorch: 1.6.0;

[0316] Torchvision: 0.7.0.

[0317] Figure 5 This invention provides a Gauss-Wasesterstein distance index for a point set and an elliptical target, used to measure the matching degree between the point set and the elliptical target. Figure 5 There are six sub-graphs in total, among which Figure 5 (a), (b), and (c) are schematic diagrams showing the following scenarios: the point set is outside the elliptical target; the point set is distributed both inside and outside the elliptical target; and the point set is only inside the elliptical target with 100 measurement points, respectively. Figure 5 (d) is a schematic diagram when the point set is distributed inside the elliptical target and has a quantity of 1000. Figure 5 (e) is Figure 5 The index values ​​corresponding to (a), (b), and (c) Figure 5 (f) represents the change in the index value when the point set is distributed inside the elliptical target and the number of measurement points is continuously increased. It can be seen that if the point set is distributed inside the elliptical target (actual measurement), the influence on the index value is minimal, providing an excellent basis for measurement-target matching.

[0318] Figure 6This is the tracking result of the model when the number of targets is 3, as provided in this embodiment of the invention. Figure 6 (a) is a schematic diagram of the trajectory. Figure 6 (b) is the GWD indicator. Figure 6 (c) represents the IoU index. The total tracking time is 60 time steps. It can be seen that the tracking error decreases with time. Correspondingly, GWD converges, IoU increases and eventually the change tends to level off. It is not difficult to see from the results that the elliptical multi-extended target tracking algorithm based on discrete-time dynamic heterogeneous graph neural network proposed in this embodiment has good tracking accuracy and can generate target trajectories in real time.

[0319] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An elliptical multi-extended target tracking method based on a dynamic heterogeneous graph neural network, characterized in that, include: Obtain information about the target to be tracked; Construct a multi-extended target tracking model; The target information to be tracked is input into the multi-extended target tracking model for processing to obtain the tracking result; The multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and to use a graph neural network to perform end-to-end computation on the discrete-time dynamic heterogeneous graph to guide the global optimization of the model. Constructing the multi-extended target tracking model includes: The multi-extended target tracking process is modeled as a discrete-time dynamic heterogeneous graph, and a heterogeneous graph slice containing measurement nodes and target nodes is constructed in each time step. The heterogeneous graph slice includes measurement-measurement edges, measurement-target edges, and target-target edges. The heterogeneous graph slices at each time step are encoded, and the embedding representations of different types of nodes are aggregated through a multi-head attention mechanism to generate updated node embeddings. Temporal modeling is performed on the embedding sequence of the target node, and a target node representation containing long-term spatiotemporal dependencies is generated by combining the historical state memory matrix; The measurement-target correlation probability matrix, target quantity prediction, and target state estimation are jointly decoded by the correlation prediction decoding module, wherein the correlation prediction decoding module includes a correlation calculation layer, a target quantity prediction layer, and a target state prediction layer. Design a multi-task loss function to jointly optimize data association, target quantity estimation, and state filtering tasks, guide global model optimization, and obtain the multi-extended target tracking model.

2. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The weights of the measurement-measurement edges are calculated using a Gaussian kernel function to reflect the spatial proximity of the measurement nodes; the weights of the measurement-target edges are calculated using normalized Mahalanobis distance to measure the association probability between the measurement and target nodes; and the weights of the target-target edges are calculated using normalized Gaussian Wasserstein distance to measure the similarity of the target nodes in terms of position and shape.

3. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, Encoding the heterogeneous graph slice at each time step includes: The measurement node and the target node are subjected to type-specific linear transformations through a heterogeneous graph attention network, and the embedded representations of heterogeneous neighbor nodes are aggregated through a multi-head attention mechanism. The multi-head attention mechanism uses edge-type specific attention vectors to calculate attention coefficients between nodes and generates updated node embeddings through weighted aggregation.

4. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 3, characterized in that, The processing procedure of the multi-head attention mechanism is as follows: ； In the formula, For the first Layer Embedding of individual nodes, Indicates splicing, A set of edge types, For nodes exist The set of neighbors under the edge class, For attention weights, For node type The transformation matrix, For the number of attention heads, For the first Layer Each neighbor node is embedded.

5. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, Temporal modeling of the embedding sequence of the target node includes: A self-attention temporal aggregation module is used to perform temporal modeling on the embedding sequence of the target node. The self-attention temporal aggregation module is used to add sinusoidal positional encoding to the target node embedding to generate a positional encoding vector. The positional encoding vector interacts with the historical state memory matrix through a dot product attention mechanism to dynamically fuse historical temporal information and generate the target node representation containing long-term spatiotemporal dependencies. The historical state memory matrix is ​​used to store the historical embeddings of the target node and is updated through a sliding window mechanism to limit the computational complexity.

6. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The associated computation layer is used to embed measurement nodes and target nodes into a common space through linear transformation, and through matrix multiplication and... The function generates the measurement-target correlation probability matrix; The target quantity prediction layer is used to predict the total number of targets at the next moment based on the current measurement information and historical target information. The target state prediction layer is used to combine associated probability weighted measurement information, target node embedding, and time series aggregation embedding to obtain the predicted state vector through multilayer perceptron mapping.

7. The elliptical multi-extended target tracking method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The multi-task loss function includes binary cross-entropy loss, class cross-entropy loss, mean absolute error loss, and Gauss-Wasesterstein distance loss. The calculation of the Gauss-Wasestein distance loss is as follows: The discrete measurement point set is transformed into a probability distribution using the Dirac measure; For each discrete point, calculate the Wasserstein distance between the discrete point as a Gaussian distribution with mean and variance equal to an identity matrix and the target elliptical Gaussian distribution; The average Wasserstein distance of all points is used as a measure of the relationship between the point set and the elliptic distribution: Based on the Gauss-Wasestein distance formula, the distribution difference between each measurement point and the target ellipse is calculated, and the mean value is taken as the final loss term.

8. An elliptical multi-extended target tracking system based on a dynamic heterogeneous graph neural network, characterized in that, include: Information acquisition unit: used to acquire information about the target to be tracked; Model building unit: used to build multi-extended target tracking models; Tracking processing unit: used to input the target information to be tracked into the multi-extended target tracking model for processing to obtain tracking results; wherein, the multi-extended target tracking model is used to model the multi-extended target tracking process as a discrete-time dynamic heterogeneous graph, and guide the global optimization of the model by designing an end-to-end multi-task loss function; The model building unit includes: Dynamic heterogeneous graph modeling module: used to construct heterogeneous graph slices containing measurement nodes and target nodes in real time, wherein the heterogeneous graph slices include measurement-measurement edges, measurement-target edges, and target-target edges; Heterogeneous graph encoding module: used to encode the heterogeneous graph slices at each time step, and aggregate the embedding representations of different types of nodes through a multi-head attention mechanism to generate updated node embeddings; Temporal aggregation module: used to perform temporal modeling on the embedding sequence of the target node, and generate a target node representation containing long-term spatiotemporal dependencies by combining the historical state memory matrix; Association prediction decoding module: used to jointly output measurement-target association probability, target quantity, and state estimation; Multi-task optimization module: Used to design multi-task loss functions, jointly optimize data association, target quantity estimation and state filtering tasks, guide global model optimization, and obtain the multi-extended target tracking model.

Images