Intelligent track initiation method

Abstract

The application provides a kind of intelligent track initiation method, utilizes early warning detection equipment m consecutive period measurement data, constructs directed graph, and sets node initial feature and edge feature;Track initiation network is composed of node generation network and node decision network, node generation network is based on the construction of Transform network, 1-hop subgraph is handled to the center of node, and the feature representation of node is output, node decision network is feedforward network, whether the connection between the true target corresponding to node and true target node exists is judged;Using historical measurement data, track initiation training data set is established, and loss function is determined, and track initiation network is optimized and trained;The track initiation network that has been trained is deployed, and the initial track of target is obtained by judging the size of output probability. The application provides a kind of unified, efficient target track initiation method, which can realize stable, rapid and accurate track initiation of target.

Description

TECHNICAL FIELD

[0001] The present application relates to a track initiation technique, and more particularly, to an intelligent track initiation method and device suitable for single target or multi-target tracking problems involved in radar, optical, infrared and other early warning detection equipment. BACKGROUND

[0002] Target detection is the process of discovering, locating and tracking targets by using radar, optical, infrared and other early warning detection equipment. The results can further support high-level processing needs such as attribute recognition, threat assessment and combat decision-making. Target tracking is an important part of target detection, and its purpose is to connect the detection information of the same target at different times, and through filtering estimation, form time-series target state information with the same identity, i.e. target track, to achieve real-time, continuous and accurate grasp of target individuals. Track initiation is the first step of target tracking, which combines measurement information of adjacent periods according to certain rules, and judges whether the combined sequence corresponds to a real target. If so, an initial track is further generated to provide subsequent processing for point track association and track filtering.

[0003] The track initiation method should be able to correctly initiate a real track while effectively suppressing false tracks. Its performance has an important impact on fast and stable multi-target tracking. In view of the important role of track initiation, scholars have conducted extensive and in-depth research on it. The existing track initiation methods can be broadly divided into four categories: 1) The first category is to transform the track initiation problem into a detection problem for processing, mainly including Bayesian detection method and sequential probability ratio test method, etc. This kind of method introduces filter residuals into the statistical decision-making process, and achieves good detection effect. The disadvantage is that the calculation of the posterior probability of the algorithm depends heavily on the statistical model, making it difficult to apply to engineering practice; 2) The second category is sequential processing algorithm, mainly including intuitive method and logical method. This kind of algorithm is a heuristic algorithm, which mainly realizes track initiation by manually designing initiation rules. It has the advantages of simple algorithm structure and strong engineering practicability. The disadvantage is that it has limited ability to cope with complex scenes, and its efficiency decreases significantly in strong clutter and target dense scenes; 3) The third category is batch processing method, mainly including Hough transform method and various improved methods. This kind of algorithm can effectively initiate straight-line moving targets, but has problems such as large amount of calculation and inability to initiate quickly; 4) The fourth category is intelligent track initiation method. At present, the true target recognition problem in track initiation is regarded as a classification problem, and methods such as support vector machine, random forest and convolutional neural network in deep learning in machine learning are used to solve it. It has the advantages of fast initiation speed and strong ability to cope with complex scenes.

[0004] In summary, intelligent track initiation methods can effectively utilize the powerful learning and representation capabilities of machine learning and deep learning, possessing enormous performance potential and representing an important development direction for track initiation method research. However, existing methods primarily employ classification methods such as support vector machines and neural networks to distinguish between real and false targets in the pre-selected sequence during the true target confirmation step. The pre-selected sequence generation step still relies on traditional methods, i.e., searching according to predefined rules. Intelligent track initiation is not achieved across the entire process and all stages. Therefore, the level of intelligence in existing track initiation methods remains quite limited, and their potential performance has not been fully realized. To further improve the performance of existing track initiation methods and solve track initiation problems in complex scenarios such as dense targets, target formations, and strong clutter, it is urgent to conduct in-depth research on intelligent track initiation methods to achieve intelligent track initiation across the entire process and all stages. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent track initiation method, offering an effective track initiation approach for target tracking problems. The specific technical solution is as follows:

[0006] Using measurement data from m consecutive detection cycles of the early warning detection equipment, a directed graph is constructed, and based on the graph connectivity, it is further decomposed into multiple disconnected sub-directed graphs G, which are then processed separately. The nodes of the directed graph are composed of measurement points, and the directed edges are composed of the lines connecting nodes of two adjacent detection cycles. The direction of the edges is from the node of the previous cycle as the starting point to the node of the next cycle as the ending point.

[0007] Based on the feature information contained in the measurement data, the initial node features of the directed graph G are constructed, and based on the spatiotemporal information contained in the measurement data, the edge features of the directed graph G are constructed.

[0008] The trajectory initiation network TINET consists of a node generation network GNET and a node decision network DNET. The process is as follows: First, the node generation network GNET is used to perform l full node iterations on each node of the directed graph G to obtain the feature representation of each node in the directed graph G. Then, the node features are further processed by the node decision network DNET to output the probability of true and false target measurement nodes and the probability of whether there is a connection between true target measurement nodes. Finally, by judging the magnitude of the output probability, the connection between true target measurement points and true target measurement points is obtained. The connected true target measurement points are the initial target trajectory.

[0009] The node generation network GNET is built upon the Transformer network and supports nodes v i Process the 1-hop subgraph centered on the node v, and output the node v. i Feature representation Input is divided into inflow outflow and this node Three parts Including node v i Information about directed edges ending at a destination and information about the nodes they connect to. Including node v i The GNET network processing consists of three parts: Inflow Information Extraction (IE), Outflow Information Extraction (OE), and Feature Generation (FG). IE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, residual connection, layer normalization, and feedforward network modules. OE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. FG includes preprocessing, position encoding, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules.

[0010] The node decision network (DNET) is a feedforward network that determines whether a node corresponds to a true target and whether there is a connection between true target nodes. The input is the feature representation of each node in the directed graph G. The network GNET output is generated by nodes, and the output is the decision of true and false target measurement nodes. Determining whether a connection exists between the target node and the target node. As the final output of the TINET track initiation network, The subscripts i and j represent the node numbers, respectively;

[0011] Using historical measurement data from early warning detection equipment, the input part of the trajectory starting sample is constructed according to the directed graph G construction method. Based on the output of the decision network DNET, the expected output part of the trajectory starting sample is constructed. This is used as a single trajectory starting sample to establish the trajectory starting training dataset D. At the same time, the loss function Loss is determined. Then, the gradient backpropagation algorithm is used to optimize and train the TINET network by minimizing the set loss function Loss.

[0012] Deploy the pre-trained TINET trajectory initiation network, capture early warning detection data in real time, generate a directed graph G, and input it into the TINET trajectory initiation network. TINET outputs the probability that the measured node corresponds to the real target. The probability that a connection exists between the target node and the true target node. Finally, by judging the magnitude of the output probability, the connection between the true target measurement points is obtained, and the connected true target measurement points are the initial target track. The subscripts i and j represent the node numbers, respectively.

[0013] The intelligent trajectory initiation method proposed in this invention provides a target trajectory initiation method applicable to early warning and detection equipment such as radar, optics, and infrared. It requires no manual debugging and has advantages such as wide applicability, suitability for complex scenarios, rapid debugging, convenient deployment, and good practical effect. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating an intelligent trajectory initiation method provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To provide a target trajectory initiation method that is structurally unified, quick to debug, easy to deploy, and has superior performance, embodiments of this invention provide an intelligent trajectory initiation method, such as... Figure 1 As shown, the method includes the following steps:

[0018] Step 1: Using the measurement data of m consecutive detection cycles of the early warning detection equipment, construct a directed graph, and further decompose it into multiple disconnected sub-directed graphs G according to the graph connectivity, and perform subsequent processing on each of them. The nodes of the directed graph are composed of measurement points, and the directed edges are composed of the lines connecting the nodes of two adjacent detection cycles. The direction of the edge is from the node of the previous cycle as the starting point to the node of the next cycle as the ending point.

[0019] Step 1.1: Extract measurement data Z from m consecutive detection cycles of the early warning detection equipment. 1:m ={Z1,Z2,...,Z m}, Z i This represents the set of detection data acquired in the i-th detection cycle. m represents the number of detection cycles extracted, the specific value of which is set manually, and a reasonable range is 4 ≤ m ≤ 6, n i z represents the number of probe data acquired in the i-th scan cycle. i,1This represents a single detection data point, which includes at least the detection time and location information [t,x,y], and may also include target velocity, target altitude information, and feature information. For radar, the feature information can be the target echo intensity; for optical and infrared, the feature information can be the target appearance features. The target location can be described in a local rectangular coordinate system, a geodetic rectangular coordinate system, or a geographic coordinate system. For optical and infrared early warning detection equipment, it can also be described as pixel coordinates with a unified reference.

[0020] Step 1.2: Using the measurement dataset Z 1:m Construct a directed graph G, with nodes v consisting of measurement points, and directed edges e consisting of lines connecting nodes in two adjacent detection cycles. The direction of each directed edge points from the node in the preceding cycle (starting point) to the node in the following cycle (ending point). The directed edges must also consider the maximum measurement error ε. max Under the condition of satisfying the target maximum speed V max and the target minimum speed V min The constraint is that the spatial distance between the two measurement points corresponding to a directed edge node should be greater than V. min Multiply by the time interval between the two measurement points and subtract ε max It should be less than V max Multiply by the time interval between the two measurement points and add ε max The maximum measurement error ε of the detection equipment max The target maximum speed V is obtained from experience or statistics. max and the target minimum speed V min The settings are determined based on the object being tracked by the detection equipment.

[0021] Step 1.3: Treat the directed graph G as an undirected graph. According to the definition of undirected graph connectivity, use graph connectivity judgment methods such as disjoint set, DFS, and BFS to decompose the directed graph G constructed in Step 1.2 into multiple unconnected sub-directed graphs for subsequent processing. Here, G(V,E) is still used to represent the decomposed sub-directed graph.

[0022] Step 2: Based on the feature information contained in the measurement data, construct the initial node features of the directed graph G, and based on the spatiotemporal information contained in the measurement data, construct the edge features of the directed graph G.

[0023] Step 2.1: Construct the initial node features of the directed graph G based on the feature information contained in the measurement data. A reasonable method for constructing initial node features is: Initial node features The initial features of a node are constructed from both measured feature data and randomly distributed data. Given the common scenario where measured data does not contain feature data, the initial features can be entirely constructed from randomly distributed data. This random distribution can be generated by a multidimensional Gaussian distribution with a mean of 0 and a variance of 1, or by a uniform distribution within the interval (0,1). The subscript i represents the node number;

[0024] Step 2.2: Construct the directed edge features of the directed graph G based on the spatiotemporal information contained in the measurement data. A reasonable method for constructing directed edge features is to obtain the directed edge features from the time difference, position difference, and direction information of the corresponding measurements of two nodes. Assume that the spatiotemporal measurement data corresponding to nodes i and j are [t...]. i ,x i ,y i ] and [t j ,x j ,y j Then the characteristics of the directed edge can be obtained as follows: in In this context, the subscript i represents the starting node number of the directed edge, and the subscript j represents the ending node number of the directed edge.

[0025] Step 2.3: Obtain the initial features of all nodes in the directed graph G. and all directed edge features Then, the Min-Max or Z-Score methods are used to standardize them in each dimension. Alternatively, standardization can be performed without initial standardization, relying on layer normalization in the TINET track origination network.

[0026] Step 3: The trajectory initiation network TINET consists of a node generation network GNET and a node decision network DNET. The process is as follows: First, the node generation network GNET is used to perform l full node iterations on each node of the directed graph G to obtain the feature representation of each node in the directed graph G. Then, the node features are further processed by the node decision network DNET to output the probability of true and false target measurement nodes and the probability of whether there is a connection between true target measurement nodes. Finally, by judging the magnitude of the output probability, the connection between true target measurement points and true target measurement points is obtained. The connected true target measurement points are the initial trajectory of the target.

[0027] Step 3.1: The track initiation network TINET consists of a node generation network GNET and a node decision network DNET. The node generation network GNET is used to generate the feature representation of the nodes, and the node decision network DNET uses the node feature representation to determine whether a node corresponds to a real target measurement and whether there is a connection between the nodes of the real target measurement.

[0028] Step 3.2: The TINET network processing flow is as follows: First, the node generation network GNET is used to perform l full node iterations on each node of the directed graph G to obtain the feature representation of each node of the directed graph G. In each iteration, a 1-hop subgraph centered on each node of the directed graph G is extracted sequentially and processed by the GNET network. Then, the node features are further processed by the node decision network DNET to output the probability of true and false target measurement nodes and the probability of whether there is a connection between true target measurement nodes. Finally, by judging the magnitude of the output probability, the connection between true target measurement points and true target measurement points is obtained, and the connected true target measurement points are the initial target trajectory.

[0029] Step 4: Node Generation Network GNET is built on the basis of Transformer networks, and generates nodes v i Process the 1-hop subgraph centered on the node v, and output the node v. i Feature representation Input is divided into inflow outflow and this node Three parts Including node v i Information about directed edges ending at a destination and information about the nodes they connect to. Including node v i The GNET network processing consists of three parts: Inflow Information Extraction (IE), Outflow Information Extraction (OE), and Feature Generation (FG). IE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, residual connection, layer normalization, and feedforward network modules. OE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. FG includes preprocessing, position encoding, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules.

[0030] Step 4.1: Node Generation Network GNET is built on the basis of Transformer networks, and generates nodes v i Process the 1-hop subgraph centered on the node v, and output the node v. i Feature representation Input is divided into inflow outflow and this node Three parts, inflow part Including node v i Directed edge information with endpoint and the information of the connected nodes Outflow portion Including node v i Directed edge information starting from and the information of the connected nodes Where the subscripts i or u represent the node number, and k represents the total number of iterations;

[0031] Step 4.2: In step 3.1, if the inflow portion of GNET is empty, i.e., there are no directed edges ending at node v and no connected nodes, a specific vector is used. Edge information and specific vectors The initial feature information of the nodes is represented. If the outflow input of GNET is empty, that is, there is no directed edge originating from node v and no connected nodes, a special vector is used. Edge information and special vectors The initial feature information of the nodes is represented. and Size and Directed Edge Characteristics Consistent, and Size and initial node features To be consistent, a reasonable setting method is to put Set it to a vector of all zeros, and... Set as a vector of all 1s;

[0032] Step 4.3: GNET network processing consists of three parts: Inflow Information Extraction (IE), Outflow Information Extraction (OE), and Feature Generation (FG). IE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, residual connection, layer normalization, and feedforward network modules. OE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. FG includes preprocessing, position encoding, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. Except for the preprocessing, position encoding, and feature concatenation modules, the functions of the other modules are consistent with the corresponding modules in the Transformer network.

[0033] Step 4.4: The preprocessing module only functions during the first iteration of node processing. It consists of feedforward network layers, and its input is the initial features of the node. The output is the node features of the 0th iteration.

[0034] Step 4.5: Position encoding involves encoding the positions of nodes based on the directed graph structure, and then introducing this encoding into the graph through an addition operation. or In this process, the Weisfeiler-Lehman algorithm, Laplacian feature vectors, or sine and cosine functions based on the relative detection period can be used for position encoding, either individually or in combination.

[0035] Step 4.6: Feature concatenation consists of a feedforward network layer and a vector concatenation operation, specifically involving directed edge information. or After processing by the feedforward network layer, it is then combined with the directed node information. or Simply assemble them;

[0036] Step 4.7: The GNET network consists of a two-layer Transformer network architecture. In the first layer of the Transformer network architecture, firstly... After processing by the IE network part, its output vector is used as the input to the Query vector and Key vector of the multi-head attention mechanism in the OE network part, and then... After preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, residual connection, and layer normalization in the OE network part, the result is used as the Value vector input for the multi-head attention mechanism in the OE network part. Finally, the output of the multi-head attention mechanism in the OE network part is further processed through residual connections, layer normalization, and a feedforward network, and used as the output of the first layer Transformer network architecture. In the second layer Transformer network architecture, the output of the first layer Transformer network architecture, i.e., the output of the OE network part, is first used as the Query vector and Key vector input for the multi-head attention mechanism in the FG network part, and then the input of this node... After preprocessing, position encoding, and feedforward network processing, the vector is used as the value input for the multi-head attention mechanism in the FG network. Finally, the output of the multi-head attention mechanism in the FG network is further processed through residual connections, layer normalization, and feedforward network to finally output node features. As a GNET network output.

[0037] Step 5: Node Decision Network DNET is a feedforward network that determines whether a node corresponds to a true target and whether there is a connection between true target nodes. The input is the feature representation of each node in the directed graph G. The network GNET output is generated by nodes, and the output is the decision of true and false target measurement nodes. Determining whether a connection exists between the target node and the target node. As the final output of the TINET track initiation network, The subscripts i and j represent the node numbers, respectively;

[0038] Step 5.1: The node decision network DNET is a feedforward network that determines whether a node corresponds to a true target and whether there is a connection between true target nodes. The input is the feature representation of each node in the directed graph G. After processing by the feedforward network, the signal is split into two paths. One path is processed by a single-layer neural network with Sigmoid as the activation function, which outputs the decision of whether the target measurement node is real or fake. After processing through a single-layer neural network, the node connection representation z is obtained. i Then, the nodes are multiplied together and the result is output as a decision on whether a connection exists between the target nodes.

[0039] Step 5.2: A simple DNET is a two-layer feedforward network, with feature representations for each node in the directed graph G. The process is split into two paths. One path is processed through a feedforward network with a sigmoid activation function, which outputs a decision on whether the target measurement nodes are true or false. After passing through a single layer of feedforward network, the node connectivity representation z is obtained. i Then, the connection representations of each pair of nodes are multiplied, and the result is output as a decision on whether a connection exists between the true target nodes.

[0040] Step 6: Using historical measurement data from early warning detection equipment, construct the input part of the trajectory starting sample according to the directed graph G construction method. Based on the output of the decision network DNET, construct the expected output part of the trajectory starting sample. Use this as a single trajectory starting sample to establish the trajectory starting training dataset D. At the same time, determine the loss function Loss. Then, use the gradient backpropagation algorithm to optimize and train the TINET network by minimizing the set loss function Loss.

[0041] Step 6.1: Collect measurement data obtained by the same type of early warning detection equipment under different environments. According to the directed graph construction method determined in Step 1 and the node and edge feature construction method determined in Step 2, construct the input part of the track starting sample. According to the TINET network output determined in Step 5, construct the expected output part of the track starting sample. Use this as a single track starting sample to establish a track starting training dataset D. The early warning detection equipment can be radar, optical, and infrared, but different training datasets D need to be established for different types and functions of early warning detection equipment. Different environments include all environmental factors that affect the early warning detection data, including but not limited to target density, clutter level, interference, water and weather, season, and time of day. The early warning detection data is the data after target detection processing. For radar, it is the detected point trace and echo. For optical and infrared, it is the detected target center position and target appearance slice.

[0042] Step 6.2: A reasonable loss function is as follows Where V + V represents the set of true target measurement nodes. - Let y represent the set of measurement nodes for the dummy target. i ={0,1}, where 1 indicates that the node's measurement corresponds to a real target, and 0 indicates that the node's measurement corresponds to a false target. The probability of real and fake target measurement nodes output by the TINET network, s i,j = {0, 1}, where 1 indicates that there is a connection between the true target nodes i and j, and 0 indicates that there is no connection. ω is the decision ω makes regarding the existence of a connection between the true target measurement nodes output by the TINET network. V ,ω S The weighting coefficient for sample imbalance;

[0043] Step 6.3: Then, the gradient backpropagation algorithm is used to train TINET by minimizing the set loss function Loss. The network weight optimization methods that can be used include gradient descent, stochastic gradient descent, Momentum, RMSProp, Adam, but are not limited to these.

[0044] Step 7: Deploy the pre-trained Track Initiation Network (TINET), capture early warning detection data in real time, generate a directed graph G, and input it into the TINET network. TINET outputs the probability of each node's measurement corresponding to a real target. The probability that a connection exists between the target node and the true target node. Finally, by judging the magnitude of the output probability, the connection between the true target measurement points is obtained, and the connected true target measurement points are the initial target track. The subscripts i and j represent the node numbers, respectively.

[0045] Step 7.1: Deploy the pre-trained trajectory initiation network TINET, capture early warning detection data in real time, generate a directed graph G, input it into the node generation network GNET, and obtain the feature representation of the nodes after l full node iterations. The data is further processed by the Node Decision Network (DNET) to obtain the probability of the node measurement corresponding to the true target. The probability that a connection exists between the target node and the true target node. Finally, by judging the magnitude of the output probability, the connection between the true target measurement points is obtained. One feasible method is if the network output... If the value is greater than or equal to 0.5, the node represents a true target measurement; if it is less than 0.5, it represents a false target measurement. If the network output... A value greater than or equal to 0.5 indicates that the two nodes are connected; a value less than 0.5 indicates that they are not connected.

[0046] Step 7.2: The true target measurements and connections between true targets output by TINET are the final results. The connected true target measurement points of different periods are the initial tracks of the successfully started target.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method of intelligent track initiation, characterized by, The method comprises the following steps: Step 1: using the measurement data of m continuous detection periods of the early warning detection equipment, a directed graph is constructed, and according to the graph connectivity, it is further decomposed into a plurality of mutually disconnected sub-directed graphs G for subsequent processing, wherein the nodes of the directed graph are composed of measurement points, the directed edges are composed of the connection lines between the nodes of two adjacent detection periods, and the direction of the edge is from the node of the previous period as the starting point to the node of the next period as the ending point; Step 2: according to the feature information contained in the measurement data, the initial features of the nodes of the directed graph G are constructed, and according to the space-time information contained in the measurement data, the edge features of the directed graph G are constructed; Step 3: the track initiation network TINET is composed of a node generation network GNET and a node decision network DNET, and the processing process is that firstly, the node generation network GNET is used to perform l times of full node iteration on each node of the directed graph G to obtain the feature representation of each node of the directed graph G, then the node decision network DNET is used to further process the node features, and output the true and false target measurement node probability and the connection probability between the true target measurement nodes, and finally, the true target measurement points and the connection between the true target measurement points are obtained by judging the size of the output probability, and the connected true target measurement points are the initial tracks of the targets; Step 4: Node Generation Network GNET is built on the basis of Transformer networks, and generates nodes v i Process the 1-hop subgraph centered on the node v, and output the node v. i Feature representation Input is divided into inflow outflow and this node Three parts Including node v i Information about directed edges ending at a destination and information about the nodes they connect to. Including node v i The GNET network processing consists of three parts: Inflow Information Extraction (IE), Outflow Information Extraction (OE), and Feature Generation (FG). IE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, residual connection, layer normalization, and feedforward network modules. OE includes preprocessing, position encoding, feature concatenation, multi-head self-attention mechanism, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. FG includes preprocessing, position encoding, multi-head attention mechanism, residual connection, layer normalization, and feedforward network modules. Wherein, the GNET network is composed of two layers of Transformer network architecture, in the first layer of Transformer network architecture, first The output vector is processed through the IE network part, and is input as the Query vector and Key vector of the multi-head attention mechanism in the OE network part, then After preprocessing, position coding, feature splicing, multi-head self-attention mechanism, residual connection, layer normalization processing in the OE network part, the output is input as the Value vector of the multi-head attention mechanism in the OE network part, and finally the output of the multi-head attention mechanism in the OE network part is further processed through residual connection, layer normalization and feedforward network, and the output of the first layer of Transformer network architecture is output, in the second layer of Transformer network architecture, first The output of the first layer of Transformer network architecture, that is, the output of the OE network part, is input as the Query vector and Key vector of the multi-head attention mechanism in the FG network part, then the input of the node After preprocessing, position coding and feedforward network processing, the output is input as the Value vector of the multi-head attention mechanism in the FG network part, and finally the output of the multi-head attention mechanism in the FG network part is further processed through residual connection, layer normalization and feedforward network, and the node feature is finally output As the output of the GNET network; Step 5: Node decision network DNET is a feedforward network, which judges whether the node corresponds to the true target and whether there is a connection between the true target nodes, and the input is the feature representation of each node of the directed graph G The output of the node generation network GNET is the true and false target measurement node decision And the connection decision between the true target nodes As the final output of the track initiation network TINET, wherein The subscripts i and j of respectively represent the node number; Step 6: using the historical measurement data of the early warning detection equipment, the input part of the track initiation sample is constructed according to the directed graph G construction method, the expected output part of the track initiation sample is constructed according to the output result of the decision network DNET, and the single track initiation sample is used as the track initiation training data set D, and the loss function Loss is determined, then the gradient back propagation algorithm is used to optimize and train the TINET network by minimizing the set loss function Loss; Step 7: Deploy the trained track initiation network TINET, intercept the early warning probe data in real time, generate a directed graph G, input it into the track initiation network TINET, and output the probability of the node measurement corresponding to the true target by the TINET and the probability of the connection between the true target nodes Finally, the true target measurement point and the connection between the true target measurement points are obtained by judging the size of the output probability, and the connected true target measurement points are the initial target tracks initiated, wherein the subscripts i and j represent the node serial numbers.

2. The intelligent track initiation method of claim 1, wherein In step 2, the construction of the directed graph G's edge features is achieved by using the time difference, position difference, and direction information of the measurements corresponding to two nodes. Assuming that the spatiotemporal measurement data corresponding to nodes i and j are [t...],... i ,x i ,y i ] and [t j ,x j ,y j Then the characteristics of the directed edge can be obtained as follows: in In this context, the subscript i represents the starting node number of the directed edge, the subscript j represents the ending node number of the directed edge, and e represents the directed edge.

3. The intelligent track initiation method of claim 1, wherein The position encoding of the step 4 node in the network GNET is position encoding of the node according to a directed graph structure, and then introduced into or The Weisfeiler-Lehman algorithm, Laplacian feature vector or relative detection period-based sine function can be used respectively or in combination for position encoding; u represents the node serial number.

4. The intelligent track initiation method of claim 1, wherein, The feature splicing in the node generation network GNET of the step 4 is composed of a feedforward network layer and a vector splicing operation, specifically, the directed edge information is processed through the feedforward network layer, and then spliced with the directed node information.

5. The intelligent track initiation method of claim 1, wherein, The step 4 node generates a network GNET composed of two layers of Transformer network architecture, in the first layer of Transformer network architecture, first processed by the IE network part, the output vector is input as the Query vector and Key vector of the multi-head attention mechanism in the OE network part, then processed by the OE network part, the output of the multi-head attention mechanism in the OE network part is further processed by residual connection, layer normalization and feedforward network, and the output of the first layer of Transformer network architecture is output, in the second layer of Transformer network architecture, first the output of the first layer of Transformer network architecture, i.e. the output of the OE network part, is input as the Query vector and Key vector of the multi-head attention mechanism in the FG network part, then the input of the node processed by preprocessing, position encoding and feedforward network, and the output of the multi-head attention mechanism in the FG network part is further processed by residual connection, layer normalization and feedforward network, and the node feature is finally output as the output of the GNET network.

6. The intelligent track initiation method of claim 1, wherein, The simple DNET in step 5 is a two-layer feedforward network, and the feature of each node of the directed graph G is represented as It is directly divided into two paths for processing, one path is processed by a feedforward network with a layer of Sigmoid excitation function, and the output is the true or false target measurement node judgment One path is processed by a feedforward network to obtain node connection representation z i Then multiply the node connection representations two by two, and output the judgment of whether there is a connection between the true target nodes 7. The intelligent track initiation method of claim 1, wherein, The loss function in the step 6 is where V + represents the true target measurement node set, V - represents the false target measurement node set, y i = {0, 1}, 1 means the node measurement corresponds to the true target, 0 means the node measurement corresponds to the false target, is the true and false target measurement node probability output by the TINET network, s i,j = {0, 1}, 1 means that there is a connection between the true target nodes i, j, 0 means that there is no connection, is the true target measurement node connection decision output by the TINET network, ω V , ω S is the sample imbalance weighting coefficient.

Images