Graph feature transfer method, target tracking method and graph feature transfer device

CN117953233BActive Publication Date: 2024-11-26BEIJING GUOSU INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410014082.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-03
Publication Date
2024-11-26
Estimated Expiration
2044-01-03

AI Technical Summary

Technical Problem

The existing graph attention neural network cannot effectively suppress interference between false neighbors and similar neighbors in target tracking, resulting in node characteristics assimilation and is not suitable for tracking scenarios of high-speed moving targets such as table tennis.

Method used

The graph message delivery neural network model is adopted to generate a feature set of nodes and edges, and perform nodes and edge message delivery steps in the graph message delivery layer, and use feature mapping functions and real-number mapping functions to calculate the message delivery weight, and output the updated node and edge features, thereby improving the accuracy and reliability of feature transmission.

Benefits of technology

Without pre-distinguishing the authenticity of candidate targets, the accuracy and reliability of image and target feature message transmission in video frames are improved, and the effectiveness and accuracy of target tracking are improved. It is especially suitable for tracking scenarios of high-speed sports targets such as table tennis. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117953233B_ABST
    Figure CN117953233B_ABST
Patent Text Reader

Abstract

The present application provides a graph feature transfer method, a target tracking method and a graph feature transfer device, the method comprising: generating a node feature set and a corresponding edge feature set of N graphs according to feature vectors of each candidate target corresponding to the current target video stream; inputting the node feature set and the edge feature set of the N graphs into a preset graph message transfer neural network model, so that the graph message transfer neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message transfer layer, so as to output the node feature set and the edge feature set after the feature update corresponding to the N graphs. The present application can effectively improve the accuracy and reliability of image and target feature message transfer in video frames without having to distinguish in advance whether the candidate targets are real, and thus can effectively improve the effectiveness and accuracy of target tracking based on the feature transfer results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of image data processing technology, and in particular to a graph feature transfer method, a target tracking method and a graph feature transfer device. Background Art

[0002] At present, some studies have focused on solving the target tracking problem based on graph neural networks; they mainly use mature graph convolutional networks (Graph Convolutional Networks) and graph attention networks (Graph Attention Networks) for feature aggregation and message transmission, and judge and classify the features of nodes, edges or the entire graph after the transmission. The results of these feature judgments are used to complete the tracking of potential candidate targets, that is, to form a target chain.

[0003] The graph attention neural network adds an attention mechanism on the basis of the graph convolutional network, and screens the neighbor nodes of the graph through self-learning. However, it cannot completely suppress the interference of many false neighbors and / or similar neighbors, which will cause the original features of the node to be assimilated by the features of the surrounding nodes, limiting its use in node and edge classification problems and graphs with a large number of false and generalized connections; in addition, the original graph attention neural network does not set special features for the edges, and uses a simple weighted average aggregation and update of node information, which is easily affected by false interference, so it is not effective in tracking problems. At the same time, the graph attention neural network requires that the graph attention mechanism at the beginning has a clear understanding of the surrounding nodes and makes a judgment. In scenarios such as table tennis as the target, there are usually a large number of false, suspected targets and interference items, which will further affect the accuracy of feature transfer and target tracking, making the graph attention neural network unsuitable for tracking scenarios of high-speed moving targets such as table tennis. Summary of the invention

[0004] In view of this, embodiments of the present application provide a graph feature transfer method, a target tracking method, and a graph feature transfer device to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present application provides a graph feature transfer method, including:

[0006] Generate a node feature set and a corresponding edge feature set of an N-part graph according to the feature vectors of each candidate target corresponding to the current target video stream;

[0007] The node feature sets and edge feature sets of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature sets and edge feature sets after feature updates corresponding to the N graphs.

[0008] In some embodiments of the present application, the node feature set and the edge feature set of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after the feature update corresponding to the N graphs, including:

[0009] The node feature sets and edge feature sets of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model selects different nodes as current target nodes in the N graphs, and executes a preset node message passing step based on the graph message passing layer for each selected target node, so as to obtain node feature update vectors corresponding to each of the nodes in the N graphs and form a node feature set after feature update; and the graph message passing neural network model also selects different edges as current target edges in the N graphs, and executes a preset edge message passing step based on the graph message passing layer for each selected target edge, so as to obtain edge feature update vectors corresponding to each of the edges in the N graphs and form an edge feature set after feature update.

[0010] In some embodiments of the present application, the graph message delivery layer includes: a node feature delivery unit for performing the node message delivery step;

[0011] Wherein, the node message transmission step includes:

[0012] The node feature vectors of all neighbor nodes corresponding to the target node are traversed in the node feature set of the N-part graph, and the edge feature vectors of all neighbor edges between the target node and the neighbor nodes are traversed from the edge feature set of the N-part graph, so as to obtain different result sets corresponding to the target node, wherein each of the result sets contains the correspondence between the node feature vector of the target node, the node feature vector of one of the neighbor nodes, and the edge feature vector of one of the neighbor edges;

[0013] Splicing the result sets of the target node to obtain a corresponding spliced ​​tensor;

[0014] Inputting the concatenated tensor into a preset feature mapping function, so that the feature mapping function outputs a message tensor corresponding to the concatenated tensor based on a preset first parameter to be learned;

[0015] Selecting the maximum value in each batch direction of the message tensor to obtain the maximum message tensor of the target node;

[0016] The maximum message tensor and the node feature vector of the target node are input into a preset node feature update function, so that the node feature update function outputs a node feature update vector corresponding to the target node based on a preset third parameter to be learned.

[0017] In some embodiments of the present application, after outputting the message tensor corresponding to the concatenated tensor, the node message transmission step further includes:

[0018] Inputting the message tensor into a preset real number mapping function, so that the real number mapping function outputs a message transmission weight in real number form corresponding to the message tensor based on a preset second parameter to be learned;

[0019] The message passing weight corresponding to the message tensor is associated with each edge of the N-part graph to obtain the message passing weight of two nodes corresponding to each edge in the N-part graph.

[0020] In some embodiments of the present application, the graph message passing layer includes: an edge feature passing unit for performing the edge message passing step;

[0021] The edge message transmission step includes:

[0022] Performing SoftMax mapping on the message transfer weights of the two nodes associated with the target edge obtained in advance to obtain a threshold value regressed to (0,1);

[0023] The node feature update vectors corresponding to the two nodes associated with the target edge, the pre-acquired threshold and the edge feature vector of the target edge are input into a preset edge feature update function, so that the edge feature update function outputs the edge feature update vector corresponding to the target edge based on a preset fourth parameter to be learned.

[0024] In some embodiments of the present application, generating a node feature set and a corresponding edge feature set of an N-part graph according to feature vectors of each candidate target corresponding to the current target video stream includes:

[0025] Using a preset target detection method to detect the current target video stream frame by frame to obtain a feature vector of the candidate target corresponding to each video frame of the target video stream;

[0026] Setting a unique identifier for each candidate target and storing it in a preset candidate target data set;

[0027] Based on the preset total number of video frames and the starting frame of the N graphs to be constructed, the feature vectors of the corresponding candidate targets are selected from the target data set as the node features of each node, so as to form a node feature set of the N graphs;

[0028] An edge feature set corresponding to the node feature set of the N-part graph is generated, and the edge features corresponding to each edge in the edge feature set and the adjacency matrix of the N-part graph are initialized.

[0029] Another aspect of the present application provides a target tracking method, comprising:

[0030] Scoring the node feature update vectors in the node feature set and the edge feature update vectors in the edge feature set after the feature update corresponding to the N graphs obtained by the graph feature transfer method, so as to obtain the confidence of each node and each edge in the N graphs;

[0031] The confidence of each of the nodes and each of the edges is subjected to noise filtering to obtain nodes and edges for constituting a trajectory, and to generate corresponding target tracking result data.

[0032] A third aspect of the present application provides a graph feature transmission device, comprising:

[0033] The target detection module is used to generate a node feature set and a corresponding edge feature set of an N-part graph according to the feature vectors of each candidate target corresponding to the current target video stream;

[0034] A feature transfer module is used to input the node feature set and edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and edge feature set after feature update corresponding to the N graphs.

[0035] The fourth aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the graph feature transfer method or the target tracking method when executing the computer program.

[0036] The fifth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the graph feature transfer method, or implements the target tracking method.

[0037] The graph feature transfer method provided by the present application generates a node feature set and a corresponding edge feature set of N graphs according to the feature vectors of each candidate target corresponding to the current target video stream; the node feature set and the edge feature set of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after the feature update corresponding to the N graphs; it can effectively improve the accuracy and reliability of image and target feature message transfer in video frames without having to distinguish in advance whether the candidate targets are real, and thus can effectively improve the effectiveness and accuracy of target tracking based on the feature transfer results, and is particularly suitable for scenarios such as table tennis tracking that need to focus on motion features and main motion patterns.

[0038] Additional advantages, purposes, and features of the present application will be partially described in the following description, and will become partially apparent to those skilled in the art after studying the following, or may be learned from the practice of the present application. The purposes and other advantages of the present application can be achieved and obtained by the structures specifically pointed out in the specification and the drawings.

[0039] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present application are not limited to the above specific description, and the above and other purposes that can be achieved by the present application will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present application, constitute a part of the present application, and do not constitute a limitation of the present application. The components in the drawings are not drawn to scale, but are only for the purpose of illustrating the principles of the present application. In order to facilitate the illustration and description of some parts of the present application, the corresponding parts in the drawings may be enlarged, that is, they may become larger relative to other components in the exemplary device actually manufactured according to the present application. In the drawings:

[0041] Figure 1 This is a schematic diagram of a first flow chart of a method for transferring graph features in an embodiment of the present application.

[0042] Figure 2 This is a second flow chart of the method for transferring graph features in one embodiment of the present application.

[0043] Figure 3 This is a schematic diagram showing an example of the specific implementation process of g(x,θ3) in one embodiment of the present application.

[0044] Figure 4This is a schematic diagram showing an example of the specific implementation process of h(x,θ4) in one embodiment of the present application.

[0045] Figure 5 A flowchart of the node message transmission step in the graph feature transmission method in one embodiment of the present application.

[0046] Figure 6 This is a schematic diagram of an example of a GMM update algorithm in an embodiment of the present application.

[0047] Figure 7 The present invention is a flowchart of the edge message transmission step in the graph feature transmission method in one embodiment of the present application.

[0048] Figure 8 Schematic diagram of a target tracking method in one embodiment of the present application.

[0049] Fig. 9 This is a schematic diagram of an example flow of a target tracking method in an application example of the present application.

[0050] Fig.10 This is a schematic diagram of an example of the construction process of a graph in an embodiment of the present application.

[0051] Fig.11 It is a schematic diagram of an example of a graph structure during detection (a "shingled" structure to improve detection performance) in one embodiment of the present application.

[0052] Fig.12 This is a schematic diagram of an example of the graph structure before graph sampling provided in this application.

[0053] Fig.13 This is a schematic diagram of an example of the graph structure after graph sampling provided in this application.

[0054] Fig.14 A schematic diagram of the structure of the graph feature transmission device provided in this application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the implementation modes and the accompanying drawings. Here, the illustrative implementation modes and descriptions of the present application are used to explain the present application, but are not intended to limit the present application.

[0056] It should also be noted here that in order to avoid obscuring the present application due to unnecessary details, only the structures and / or processing steps closely related to the scheme according to the present application are shown in the accompanying drawings, while other details that are not very relevant to the present application are omitted.

[0057] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0058] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0059] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0060] Traditional graph neural networks receive input data of a graph structure consisting of nodes and edges, where nodes and edges have their own features, and edge features can be ignored in some graph neural networks. Based on the graph structure, the graph neural network will aggregate the features of its neighbor nodes and neighbor edges node by node, and use this to update its own node or edge features. The neighbors here are usually defined by the adjacency matrix of the graph structure or a custom adjacency matrix. In the process of feature transfer, the graph can be pooled, the graph structure can be changed, and some nodes and edges can be deleted. Usually, one feature aggregation / transfer in a graph neural network is called a layer of the graph neural network. After passing through multiple layers of the graph neural network, the features of each node or edge can be processed separately according to the specific task (such as aggregation, judgment, classification, etc.).

[0061] However, the traditional graph attention neural network cannot completely suppress the interference of many false neighbors and / or similar neighbors, which will cause the original features of the node to be assimilated by the features of the surrounding nodes, limiting its use in node and edge classification problems and graphs with a large number of false and generalized connections; in addition, the original graph attention neural network does not set special features for edges, and uses a simple weighted average to aggregate and update node information, which is easily affected by false interference, so it is not effective in tracking problems. At the same time, the graph attention neural network requires the graph attention mechanism to have a clear understanding of the surrounding nodes and make a judgment at the beginning. In scenarios such as table tennis, there are usually a large number of false, suspected targets and interference items, which will further affect the accuracy of feature transfer and target tracking, making the graph attention neural network unsuitable for tracking scenarios of high-speed moving targets such as table tennis.

[0062] Based on this, the embodiments of the present application respectively provide a graph feature transmission method, a target tracking method, a graph feature transmission device for executing the graph feature transmission method, a physical device and a computer-readable storage medium, which can effectively improve the accuracy and reliability of image and target feature message transmission in video frames without having to pre-distinguish whether the candidate targets are real, and thus can effectively improve the effectiveness and accuracy of target tracking based on the feature transmission results.

[0063] The details are described in detail through the following examples.

[0064] Based on this, the present application embodiment provides a graph feature transfer method that can be implemented by a graph feature transfer device, see Figure 1 , the graph feature transfer method specifically includes the following contents:

[0065] Step 100: Generate a node feature set and a corresponding edge feature set of N graphs according to the feature vectors of each candidate target corresponding to the current target video stream.

[0066] In one or more embodiments of the present application, candidate targets include but are not limited to real targets and false targets. The node feature set includes a one-to-one correspondence between the unique identifier of each node and the feature vector of each node, and the node is a candidate target selected from the candidate targets corresponding to the current target video stream for forming an N-part graph. The edge feature set includes a one-to-one correspondence between the unique identifier of each edge and the feature vector of each edge.

[0067] It can be understood that an N-partite graph is a special graph structure in which the vertex set of the N-partite graph can be divided into N mutually non-intersecting subsets, so that the two endpoints of each edge in the N-partite graph come from different subsets. In other words, an N-partite graph is a graph with N independent vertex sets, where there are no edges connecting the vertices in each vertex set. Where N is a positive integer equal to or greater than 1.

[0068] Take the bipartite graph as an example, it is a special N-partite graph (N=2), whose vertex set can be divided into two non-intersecting subsets, so that the two endpoints of each edge in the graph come from the two subsets. Bipartite graphs are widely used in many fields, such as network flow problems, matching problems, scheduling problems, etc.

[0069] It should be noted that N-partite graph and complete N-partite graph are different concepts. In a complete N-partite graph, there is an edge connecting any two vertices in each vertex set. In a general N-partite graph, the edges between vertex sets may or may not exist.

[0070] Step 200: Input the node feature set and edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and edge feature set after feature update corresponding to the N graphs.

[0071] In one or more embodiments of the present application, the graph message passing neural network model can be pre-trained based on the node feature set and edge feature set of N graphs corresponding to historical video streams and the corresponding labels, and is used for graph message passing, that is, feature transfer of video frame images.

[0072] The graph message passing neural network model can be abbreviated as a GMM (Graph Max Message Neural Network) model.

[0073] From the above description, it can be seen that the graph feature transmission method provided in the embodiment of the present application can effectively improve the accuracy and reliability of the image and target feature message transmission in the video frame without having to pre-distinguish whether the candidate target is real, thereby effectively improving the effectiveness and accuracy of target tracking based on the feature transmission results, and is particularly suitable for scenarios such as table tennis tracking that need to focus on motion features and main motion patterns.

[0074] In order to further improve the effectiveness and reliability of node feature update and edge feature update implemented by using the GMM model, in a graph feature transfer method provided in an embodiment of the present application, see Figure 2 , step 200 in the graph feature transfer method specifically includes the following contents:

[0075] Step 210: Input the node feature set and edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model selects different nodes in the N graphs as current target nodes, and executes the preset node message passing steps based on the graph message passing layer for each selected target node, so as to obtain the node feature update vectors corresponding to each of the nodes in the N graphs and form a node feature set after feature update; and the graph message passing neural network model also selects different edges in the N graphs as current target edges, and executes the preset edge message passing steps based on the graph message passing layer for each selected target edge, so as to obtain the edge feature update vectors corresponding to each of the edges in the N graphs and form an edge feature set after feature update.

[0076] Specifically, similar to the graph convolutional network (GCN), the GMM network can customize the number of layers as needed. Its input is an N-part graph structure G(V, E), which can be abbreviated as graph G, where N allows customization. Assuming that graph G has m nodes, it can at least refer to a 1-part graph and an m-part graph, m is a positive integer greater than 1, and graph G is a simple graph. Graph G contains the features of the nodes, and graph G can also contain the features of the edges that may be missing. The output of the GMM model is the feature vectors of each node and edge obtained after the feature transfer of the GMM model of num layers, where num layers are just the number of layers equal to or greater than 1.

[0077] In the current case, there are no residual links between the layers in the GMM model, and the layers are independent, that is, there are no shared parameters between the layers. When the num-layer GMM model is running, the input of the i-th layer GMM model is the node feature vectors corresponding to the n nodes output by the i-1th layer. The edge features corresponding to the n edges The output of the i-th layer GMM model is the node feature vector corresponding to the n nodes in the i-th layer. The edge features corresponding to the n edges i is a positive integer greater than 1.

[0078] In a specific implementation, the parameter group to be learned in each layer of the GMM model may include a first parameter to be learned θ1, a second parameter to be learned θ2, a third parameter to be learned θ3 and a fourth parameter to be learned θ4. In the four main operations f, l, g, h respectively;

[0079] Example: f(x, θ1) represents the feature mapping function, θ1 is the parameter to be learned during inference, and x represents the input; the function value of f(x, θ1) is the output; (Note: a layer of a deep learning model such as a convolutional neural network can also be abstracted as f).

[0080] Assume that the existing node and edge features p∈R i , q∈R j For input data:

[0081] f(x, θ1): Input feature is In∈R 2i+j , where In = (p, q, p) and the output feature is Mes∈R k ;

[0082] l(x, θ2): Input feature is Mes∈R k The output feature is a real number Gra∈R, and this function is optional;

[0083] g(x,θ3): Input feature is (Mes,p)∈R k+i The output feature is the new node feature p 1∈R ii , the specific implementation process of g(x, θ3) is as follows Figure 3 As shown;

[0084] h(x,θ4): Input is (p 1 ,Gra,p 1 , Gra, q) is the new edge feature q 1 ∈R jj , the specific implementation process of h(x, θ4) is as follows Figure 4 shown.

[0085] On this basis, in order to further improve the application effectiveness and reliability of the graph message passing layer in performing node feature transmission, in a graph feature transmission method provided in an embodiment of the present application, the graph message transmission layer includes: a node feature transmission unit for executing the node message transmission step.

[0086] For the corresponding Figure 5 , the node message transmission step specifically includes the following contents:

[0087] Step 211: traverse the node feature vectors of all neighbor nodes corresponding to the target node in the node feature set of the N-part graph, and traverse the edge feature vectors of all neighbor edges between the target node and the neighbor nodes from the edge feature set of the N-part graph to splice and obtain different result sets corresponding to the target node. v , wherein each of the result sets includes a correspondence between a node feature vector of the target node, a node feature vector of a neighboring node, and an edge feature vector of a neighboring edge.

[0088] Step 212: Concatenate the result sets of the target node to obtain a corresponding concatenated tensor In A .

[0089] Step 213: In the spliced ​​tensor A Input the preset feature mapping function f(In A , θ1), so that the feature mapping function outputs the splicing tensor In based on the preset first parameter to be learned θ1 A The corresponding message tensor Mes A .

[0090] Step 231: In the message tensor Mes A The maximum value is selected in each batch direction to obtain the maximum message tensor Mes of the target node.

[0091] Step 232: Input the maximum message tensor Mes and the node feature vector of the target node into a preset node feature update function g((Mes, p u ), θ3), so that the node feature update function outputs the node feature update vector corresponding to the target node based on the preset third parameter to be learned

[0092] In addition, the maximum message tensor Mes and the node feature vector of the neighboring node can also be input into a preset node feature update function g((Mes, p v ), θ3), so that the node feature update function outputs the node feature update vector corresponding to the neighbor node based on the preset third parameter to be learned

[0093] In order to further improve the effectiveness and reliability of edge message transmission in the process of graph feature transmission, in a graph feature transmission method provided in an embodiment of the present application, see Figure 5 The following contents are also specifically included between step 213 and step 231:

[0094] Step 221: Transform the message tensor Mes A Input the preset real number mapping function l(Mes A , θ2), so that the real number mapping function outputs the message tensor Mes based on the preset second parameter to be learned θ2 A The corresponding real-number message passing weight Gra A .

[0095] Step 222: The message tensor Mes A The corresponding message passing weight is associated with each edge of the N-part graph to obtain the message passing weight Gra of the two nodes corresponding to each edge in the N-part graph A .

[0096] That is, the “message tensor” Mes and the “message transmission strength” Gra are calculated, and all node features are updated, such as Figure 6 As shown, for each node u∈V, the node message passing step is performed:

[0097] (1) Traverse all adjacent nodes v and their corresponding edges e v ∈E, concatenate u, v, e v The eigenvector p of u ,qe v , p v , each traversal gets the result set In v ;

[0098] (2) Splice all In vGet the concatenated tensor In a ∈R m×(2i+j) (Non-adjacent node tensors are replaced by 0);

[0099] (3)Mes A =f(In A ,θ1),Mes a ∈R m×k , get each u, v, e v A "message" tensor on triples;

[0100] (4)Gra A = l(Mes A ,θ2),Gra A ∈R m , get each u, v, e v The “message” weight on the triplet and compare it with the edge e v , association;

[0101] (5) Maximum message tensor Mes = max(Mes A , dim=0), get the "maximum message" of node u;

[0102] (6) Update the feature vector of node u.

[0103] In order to further improve the application effectiveness and reliability of edge feature transmission in the graph message transmission layer, in a graph feature transmission method provided in an embodiment of the present application, the graph message transmission layer includes: an edge feature transmission unit for executing the edge message transmission step.

[0104] For the corresponding Figure 7 , the edge message transmission step specifically includes the following contents:

[0105] Step 241: SoftMax mapping is performed on the message transfer weights of the two nodes associated with the target edge obtained in advance to obtain a threshold value (Gra, Gra) that regresses to (0, 1).

[0106] Step 242: Input the node feature update vectors corresponding to the two nodes associated with the target edge, the pre-acquired thresholds (Gra, Gra), and the edge feature vector q of the target edge into a preset edge feature update function So that the edge feature update function outputs the edge feature update vector corresponding to the target edge based on the preset fourth parameter to be learned θ4

[0107] That is, according to the “message transmission strength” Gra, update the edge features, such as Figure 6As shown, for each edge e∈E, the edge message passing step is performed:

[0108] (1) According to the graph structure of the N-part graph, the two weights obtained in the previous step (one edge has two associated nodes) (Gra e , Gra e ) respectively perform Sigmoid (when N = number of nodes m, this can be regarded as a special case of SoftMax) or SoftMax mapping to obtain the threshold (Gra, Gra) that regresses to (0, 1);

[0109] (2) Update the eigenvector of edge e.

[0110] Here is a set of specific implementation examples of f, l, g, and h. In actual use, their specific implementations can be adjusted to adjust the model performance:

[0111] 1. f, l: y = Linear2(Linear1(x)); for f, x = In A ,y=Mes A ; for l, x = Mes A , y = Gra A ;

[0112] 2.g:

[0113] 3.h:

[0114] In the above statement, Concatenate refers to the tensor (or vector) concatenation operation (since this operation does not require any parameters, there is no subscript to distinguish it), Linear i Refers to linear mapping and activation function. Different subscripts are only used to represent Linear with different parameters. Linear with the same subscript in different formulas has no association.

[0115] In order to further improve the effectiveness and reliability of graph feature transmission, in a graph feature transmission method provided in an embodiment of the present application, see Figure 2 , step 100 in the graph feature transfer method specifically includes the following contents:

[0116] Step 110: Detect the current target video stream frame by frame using a preset target detection method to obtain feature vectors of candidate targets corresponding to respective video frames of the target video stream;

[0117] Step 120: setting a unique identifier for each candidate target and storing it in a preset candidate target data set;

[0118] Step 130: Based on a preset total number of video frames and a starting frame of the N graphs to be constructed, feature vectors of corresponding candidate targets are selected from the target data set as node features of each node, respectively, to form a node feature set of the N graphs;

[0119] Step 140: Generate an edge feature set corresponding to the node feature set of the N-part graph, and initialize the edge features corresponding to each edge in the edge feature set and the adjacency matrix of the N-part graph.

[0120] The feature vector of the candidate target includes: the position, size and candidate region image features of the candidate target, and the candidate region image features include: semantic features, color features and gradient features;

[0121] The preset target detection method includes: a target detection method based on deep learning, a differential detection method or an HSV binary connected domain extraction method;

[0122] Specifically, this task can use general target detection methods based on deep learning, such as YOLO and FAST-RCNN, or detect video frames through traditional computer vision detection methods such as differential detection and HSV binary connected domain extraction; obtain the candidate target (a large degree of false detection is allowed) position, size (i.e., candidate target box), candidate region image features (including but not limited to semantic features, color features, gradient features, etc.), etc.

[0123] In a specific implementation, after the detection results are aggregated, a unique ID is determined for each candidate target and the feature information described above is stored for subsequent use.

[0124] Based on the above-mentioned embodiment of the graph feature transfer method, the present application also provides an embodiment of a target tracking method, see Figure 8 , the target tracking method specifically includes the following contents:

[0125] Step 300: scoring the node feature update vectors in the node feature set and the edge feature update vectors in the edge feature set after the feature update corresponding to the N graphs obtained by the graph feature transfer method, so as to obtain the confidence of each node and each edge in the N graphs;

[0126] Step 400: performing noise filtering on the confidence of each of the nodes and each of the edges to obtain nodes and edges for constituting a trajectory, and generating corresponding target tracking result data.

[0127] Specifically, the video stream is input and the target is tracked in the video stream. The target is currently a binary classification problem (target / noise), so both nodes and edges can output confidence. The entire framework also supports the detection of multiple different categories of targets. In this case, the single confidence of the node in step D can be adjusted to a classification confidence to complete the tracking problem of multiple categories of targets. Fig. 9 In a specific application example, the specific implementation process of the target tracking method is as follows:

[0128] A) Object Detection

[0129] This task can use general target detection methods based on deep learning, such as YOLO and FAST-RCNN, or detect video frames through traditional computer vision detection methods, such as differential detection and HSV binary connected domain extraction; obtain the position and size (i.e., candidate target box) of the candidate target (a large degree of false detection is allowed), image features of the candidate region (including but not limited to semantic features, color features, gradient features, etc.), etc., and after summarizing the detection results, determine a unique ID for each candidate target and store the feature information mentioned above for later use.

[0130] B) Construction of graph G

[0131] See also Fig.10 , determine the number of video frames L and the starting frame f of graph G (f increases as the video stream is read in, and L can be flexibly set according to needs);

[0132] Graph G will select all targets u in L frames from [f, f+L-1] from Data.X as the node set V of the graph. The feature vector of each node u is determined by the position of u, the shape features of the candidate area, and the image features;

[0133] Generate an edge set E = {e(u, v)} from the node set V, where u and v are targets that do not belong to the same frame, and the pixel distance or other attributes between u and v meet the preset threshold requirements. The preset rules can greatly reduce the number of edges in the graph G and reduce unnecessary calculations; initialize the features of the edge e as needed, which is determined by the relative position of (u, v), the time sequence, etc.;

[0134] Initialize the graph G adjacency matrix M and other matrices or tensors required for calculation;

[0135] Therefore, since any two objects in the same frame are not adjacent, the graph G is obviously an L-partite graph.

[0136] C) GMM Graph Neural Network

[0137] The graph G mentioned above, the node feature set [p1, p2, ..., p m ], edge feature set [q1, q2, ..., qn ], adjacency matrix M and other necessary tensors are input into the GMM graph neural network model. In this case, the updated node features are obtained through the 3-layer GMM model. and edge features

[0138] Here we emphasize that: since the graph G is an L-partite graph, for the connected node u and the nodes v1, v2, ..., v n The edges e1(u, v1), e2(u, v2), ..., e n (u,v n ) have a competitive relationship, so the SoftMax function is introduced to characterize the competition and obtain the Gra values ​​of these edges;

[0139] D) Feature confidence extraction

[0140] After obtaining the node u feature p and the edge e feature q, we define the target tracking result as a “path” [u1, e 12 ,u2,e 23 ,u3,...,u i , e ij ,u j , ..., u m ]. Use the idea of ​​classification problem to define two classification functions w1(p) and w2(q) to obtain the required probability distribution (that is, the confidence of the final output). For the binary classification (single target tracking) problem in this case, the function is preset with the following prior: if the node u and the edge e belong to a certain target tracking result, then w1(p) = 1, w2(q) = 1; then in use, we get w1(p) = α, w2(q) = β, which is equivalent to the probability that the node u and the edge e are members of the trajectory are α and β respectively (and the confidence we need).

[0141] E) Confidence output

[0142] In the video stream, since the movement of objects such as ping-pong balls is unpredictable (unstable features), to ensure the effectiveness of detection, we define the detection step size step. In this case, Therefore, all nodes u and edges e will obtain at least two confidences. We take the highest value. To filter out a lot of noise (or false targets), we set a confidence threshold. Nodes u and edges e that meet the threshold will be written into Data.Y and wait to be organized into trajectories. See Fig.11 .

[0143] F) Tracking result update

[0144] Select the possible updated trajectories from Data.Z according to the current timestamp (frame_id) and add them to the set Tra;

[0145] Select the unmatched target u from Data.Y, convert it into a "track" with only one target, and add it to the set Tra;

[0146] Merge the trajectories in the set Tra: In the formula, the grade function calculates the correlation between all targets of the two trajectories (i.e., the confidence of edge e, a is the preset threshold), and gather is the operation of merging the two trajectories, as shown in Table 1. During the program operation, the intermediate variable edgeGap is needed to store the minimum value of the correlation of the trajectories in the merged set Tra. If the value is greater than the preset threshold a, it indicates that there are no trajectories that need to be merged in the current merged set Tra, and the merging process ends.

[0147] Table 1

[0148]

[0149]

[0150] Filter out the "trajectories" with only one target in Tra, and update the remaining trajectories in Tra to Data.Z;

[0151] Delete useless information in Data.Y according to frame_id,stp,L.

[0152] In addition, this application also provides a training process of a GMM (Graph Max Message Neural Network) model for N-part graphs, specifically:

[0153] For an N-part graph with k types of nodes, during training, a subgraph can be generated by randomly sampling each part of the node; the generated subgraph will have a much smaller number of nodes and edges than the original graph (for most graph neural networks, its parameters will not change). During the training process, random sampling is performed as the number of training epochs increases. This can significantly reduce the memory usage, improve the training speed, and increase the diversity of the data, which can delay or alleviate the overfitting phenomenon to a certain extent. Fig.12 and Fig.13 As shown; at the same time, random sampling is not absolute. In this embodiment, the target node and the edges belonging to the target trajectory are retained in each sampling, and only non-target (noise nodes) are sampled.

[0154] Note that the complete graph G is still input during the inference process, and experiments show that there is no significant performance difference between the two.

[0155] This embodiment has high target tracking accuracy and running speed; it can track multiple targets at the same time; there is no need to preset a large number of parameters, and the graph neural network can be directly trained using labeled data, and the solution has good compatibility; the graph neural network training process converges relatively quickly, and the application of the training method can reduce the video memory occupancy during solution training.

[0156] That is to say, the embodiments of the present application provide a GMM graph neural network model, a sampling graph neural network training method for N-part graph problems; (which can be combined with a GMM model or other graph neural network models) and the use of a GMM graph neural network model and architecture for target tracking. The graph attention neural network GAT requires that the graph attention mechanism at the beginning has a clear understanding of the surrounding nodes and gives a judgment, but in tracking problems such as table tennis, there are usually a large number of false, suspected targets and interference items, so the convergence of the training parameters is unstable; and the final results obtained by applying the embodiments provided by the present application show that the newly built GMM network has a more stable convergence process than the GAT network. At the same time, the existing target tracking problem based on graph neural networks mainly tracks targets such as pedestrians, so the main learning targets are pedestrians and their image features, and the algorithm provided by the embodiments provided by the present application can be used to track table tennis, and can focus on motion features and main motion patterns.

[0157] From the software level, the present application also provides a graph feature transfer device for executing all or part of the graph feature transfer method, see Fig.14 , the graph feature transfer device specifically includes the following contents:

[0158] The target detection module 10 is used to generate a node feature set and a corresponding edge feature set of N graphs according to the feature vectors of each candidate target corresponding to the current target video stream.

[0159] The feature transfer module 20 is used to input the node feature set and edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and edge feature set after feature update corresponding to the N graphs.

[0160] The embodiment of the graph feature transfer device provided in the present application can be specifically used to execute the processing flow of the embodiment of the graph feature transfer method in the above-mentioned embodiment. Its functions will not be described in detail here, and reference can be made to the detailed description of the above-mentioned graph feature transfer method embodiment.

[0161] The part of the graph feature transmission device that performs graph feature transmission can be executed in the server or completed in the client device. The specific selection can be based on the processing capability of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor for specific processing of graph feature transmission.

[0162] The client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and other implementation scenarios may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, or a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0163] The server and the client device may communicate with each other using any suitable network protocol, including network protocols that have not yet been developed on the date of filing this application. The network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Of course, the network protocols may also include, for example, RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer) protocols used on top of the above protocols.

[0164] From the above description, it can be seen that the image feature transmission device provided in the embodiment of the present application can effectively improve the accuracy and reliability of the image and target feature message transmission in the video frame without having to pre-distinguish whether the candidate target is real, thereby effectively improving the effectiveness and accuracy of target tracking based on the feature transmission results, and is particularly suitable for scenarios such as table tennis tracking that need to focus on motion features and main motion patterns.

[0165] The embodiment of the present application also provides an electronic device, which may include a processor, a memory, a receiver and a transmitter, wherein the processor is used to execute the graph feature transmission method mentioned in the above embodiment, wherein the processor and the memory may be connected via a bus or other means, such as by bus connection. The receiver may be connected to the processor and the memory via wired or wireless means.

[0166] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0167] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the graph feature transfer method in the embodiment of the present application. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implementing the graph feature transfer method in the above method embodiment.

[0168] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0169] The one or more modules are stored in the memory, and when executed by the processor, perform the graph feature transfer method in the embodiment.

[0170] In some embodiments of the present application, the user equipment may include a processor, a memory, and a transceiver unit, which may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected through a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0171] As an implementation method, the functions of the receiver and the transmitter in the present application can be considered to be implemented through a transceiver circuit or a dedicated chip for transceiver, and the processor can be considered to be implemented through a dedicated processing chip, a processing circuit or a general chip.

[0172] As another implementation method, it is possible to use a general-purpose computer to implement the server provided in the embodiment of the present application, that is, to store the program code for implementing the functions of the processor, receiver, and transmitter in a memory, and the general-purpose processor implements the functions of the processor, receiver, and transmitter by executing the code in the memory.

[0173] The embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned graph feature transfer method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0174] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0175] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0176] In the present application, features described and / or illustrated for one embodiment may be used in the same manner or in a similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0177] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A graph feature transfer method, characterized in that: include: Generate a node feature set and a corresponding edge feature set of an N-part graph according to the feature vectors of each candidate target corresponding to the current target video stream; Inputting the node feature set and the edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after feature update corresponding to the N graphs; The step of inputting the node feature set and the edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after feature update corresponding to the N graphs, comprises: The node feature set and edge feature set of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model selects different nodes as current target nodes in the N graphs, and executes a preset node message passing step based on the graph message passing layer for each selected target node, so as to obtain node feature update vectors corresponding to each of the nodes in the N graphs and form a node feature set after feature update; and the graph message passing neural network model also selects different edges as current target edges in the N graphs, and executes a preset edge message passing step based on the graph message passing layer for each selected target edge, so as to obtain edge feature update vectors corresponding to each of the edges in the N graphs and form an edge feature set after feature update; The graph message passing layer comprises: a node feature passing unit for executing the node message passing step; Wherein, the node message transmission step includes: The node feature vectors of all neighbor nodes corresponding to the target node are traversed in the node feature set of the N-part graph, and the edge feature vectors of all neighbor edges between the target node and the neighbor nodes are traversed from the edge feature set of the N-part graph, so as to obtain different result sets corresponding to the target node, wherein each of the result sets contains the correspondence between the node feature vector of the target node, the node feature vector of one of the neighbor nodes, and the edge feature vector of one of the neighbor edges; Splicing the result sets of the target node to obtain a corresponding spliced ​​tensor; Inputting the concatenated tensor into a preset feature mapping function, so that the feature mapping function outputs a message tensor corresponding to the concatenated tensor based on a preset first parameter to be learned; Inputting the message tensor into a preset real number mapping function, so that the real number mapping function outputs a message transmission weight in real number form corresponding to the message tensor based on a preset second parameter to be learned; Associating the message passing weight corresponding to the message tensor with each edge of the N-part graph to obtain the message passing weight of two nodes corresponding to each edge in the N-part graph; Selecting the maximum value in each batch direction of the message tensor to obtain the maximum message tensor of the target node; Inputting the maximum message tensor and the node feature vector of the target node into a preset node feature update function, so that the node feature update function outputs a node feature update vector corresponding to the target node based on a preset third parameter to be learned; The graph message passing layer comprises: an edge feature passing unit for performing the edge message passing step; The edge message transmission step includes: Performing normalized exponential function mapping on the pre-acquired message transmission weights of the two nodes associated with the target edge to obtain a corresponding threshold; The node feature update vectors corresponding to the two nodes associated with the target edge, the pre-acquired threshold and the edge feature vector of the target edge are input into a preset edge feature update function, so that the edge feature update function outputs the edge feature update vector corresponding to the target edge based on a preset fourth parameter to be learned.

2. The graph feature transfer method according to claim 1, characterized in that: After outputting the message tensor corresponding to the concatenated tensor, the node message transmission step further includes: Inputting the message tensor into a preset real number mapping function, so that the real number mapping function outputs a message transmission weight in real number form corresponding to the message tensor based on a preset second parameter to be learned; The message passing weight corresponding to the message tensor is associated with each edge of the N-part graph to obtain the message passing weight of two nodes corresponding to each edge in the N-part graph.

3. The graph feature transfer method according to claim 1, characterized in that: The step of generating a node feature set and a corresponding edge feature set of N graphs according to the feature vectors of each candidate target corresponding to the current target video stream includes: Using a preset target detection method to detect the current target video stream frame by frame to obtain a feature vector of a candidate target corresponding to each video frame of the target video stream; Setting a unique identifier for each candidate target and storing it in a preset candidate target data set; Based on the preset total number of video frames and the starting frame of the N graphs to be constructed, the feature vectors of the corresponding candidate targets are selected from the target data set as the node features of each node, so as to form a node feature set of the N graphs; An edge feature set corresponding to the node feature set of the N-part graph is generated, and the edge features corresponding to each edge in the edge feature set and the adjacency matrix of the N-part graph are initialized.

4. A target tracking method, characterized in that: include: Scoring the node feature update vectors in the node feature set and the edge feature update vectors in the edge feature set after the feature update corresponding to the N graphs obtained by the graph feature transfer method according to any one of claims 1 to 3, so as to obtain the confidence of each node and each edge in the N graphs; The confidence of each of the nodes and each of the edges is subjected to noise filtering to obtain nodes and edges for constituting a trajectory, and to generate corresponding target tracking result data.

5. A graph feature transfer device, characterized in that: include: The target detection module is used to generate a node feature set and a corresponding edge feature set of an N-part graph according to the feature vectors of each candidate target corresponding to the current target video stream; A feature transfer module, used for inputting the node feature set and the edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after the features corresponding to the N graphs are updated; The step of inputting the node feature set and the edge feature set of the N graphs into a preset graph message passing neural network model, so that the graph message passing neural network model performs feature transfer on each node feature vector in the node feature set and each edge feature vector in the edge feature set based on at least one graph message passing layer, so as to output the node feature set and the edge feature set after feature update corresponding to the N graphs, comprises: The node feature set and edge feature set of the N graphs are input into a preset graph message passing neural network model, so that the graph message passing neural network model selects different nodes as current target nodes in the N graphs, and executes a preset node message passing step based on the graph message passing layer for each selected target node, so as to obtain node feature update vectors corresponding to each of the nodes in the N graphs and form a node feature set after feature update; and the graph message passing neural network model also selects different edges as current target edges in the N graphs, and executes a preset edge message passing step based on the graph message passing layer for each selected target edge, so as to obtain edge feature update vectors corresponding to each of the edges in the N graphs and form an edge feature set after feature update; The graph message passing layer comprises: a node feature passing unit for executing the node message passing step; Wherein, the node message transmission step includes: The node feature vectors of all neighbor nodes corresponding to the target node are traversed in the node feature set of the N-part graph, and the edge feature vectors of all neighbor edges between the target node and the neighbor nodes are traversed from the edge feature set of the N-part graph, so as to obtain different result sets corresponding to the target node, wherein each of the result sets contains the correspondence between the node feature vector of the target node, the node feature vector of one of the neighbor nodes, and the edge feature vector of one of the neighbor edges; Splicing the result sets of the target node to obtain a corresponding spliced ​​tensor; Inputting the concatenated tensor into a preset feature mapping function, so that the feature mapping function outputs a message tensor corresponding to the concatenated tensor based on a preset first parameter to be learned; Inputting the message tensor into a preset real number mapping function, so that the real number mapping function outputs a message transmission weight in real number form corresponding to the message tensor based on a preset second parameter to be learned; Associating the message passing weight corresponding to the message tensor with each edge of the N-part graph to obtain the message passing weight of two nodes corresponding to each edge in the N-part graph; Selecting the maximum value in each batch direction of the message tensor to obtain the maximum message tensor of the target node; Inputting the maximum message tensor and the node feature vector of the target node into a preset node feature update function, so that the node feature update function outputs a node feature update vector corresponding to the target node based on a preset third parameter to be learned; The graph message passing layer comprises: an edge feature passing unit for performing the edge message passing step; The edge message transmission step includes: Performing normalized exponential function mapping on the pre-acquired message transmission weights of the two nodes associated with the target edge to obtain a corresponding threshold; The node feature update vectors corresponding to the two nodes associated with the target edge, the pre-acquired threshold and the edge feature vector of the target edge are input into a preset edge feature update function, so that the edge feature update function outputs the edge feature update vector corresponding to the target edge based on a preset fourth parameter to be learned.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the graph feature transfer method as described in any one of claims 1 to 3, or implements the target tracking method as described in claim 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for transferring graph features as described in any one of claims 1 to 3 is implemented, or the target tracking method as described in claim 4 is implemented.