A method, device and medium for correcting the track of clustered flying targets

Abstract

The embodiments of the present specification disclose a method, device and medium for correcting the track of cluster flying targets, which relate to the technical field of cluster target tracking and are used to solve the problem that it is currently difficult to achieve high-precision tracking of cluster targets. The method comprises: obtaining multi-target point tracks detected by a preset radar on cluster flying targets to preprocess them to obtain an initial cluster target track; constructing a graph representation of the initial cluster target track to extract target interaction information and missed frame correction information according to a preset graph convolutional neural network, and updating the initial node features of the graph representation; fusing the initial node features with the updated node features to smooth the fused node features according to a bidirectional long short-term memory network to obtain a node hidden state; obtaining the node hidden states of the same target at different times to assign feature weights to them to obtain an updated node hidden state, and converting the updated node hidden state to obtain a corrected cluster target track.

Description

Technical Field

[0001] The present invention relates to the technical field of cluster target tracking, and in particular to a method, device and medium for correcting the track of cluster flying targets. Background Art

[0002] The current detection of cluster flying targets is of great significance to the study of biological migration mechanisms and the maintenance of aviation safety. Radar has become one of the effective means of monitoring the detection of cluster flying targets in the air because of its advantages of all-weather and all-day. However, the current cluster has collaborative interactions between individuals within the group, and its motion model also has complex and maneuverable problems. The filter model also has problems of mismatching or delayed matching with the actual motion. In addition, radar also has problems such as missed detection and noise. These problems will aggravate the track estimation error, making it difficult to achieve high-precision tracking in the current cluster flying target detection.

[0003] At present, in order to realize the research on group target motion, accurate modeling has been carried out, such as distributed group model, group target dynamic model based on graph theory, group model based on continuous time stochastic differential equation, etc. However, due to the diversity and changeability of cluster motion models, the above-mentioned methods based on model assumptions have the disadvantages of unreasonable assumptions and uncertain model parameters. In order to correct the motion track, the existing method adopts a track correction algorithm based on LSTM network. Although this method can learn the time series characteristics of the target track in a data-driven way and improve the target tracking accuracy in complex and unknown maneuvering scenarios, it is difficult to achieve high-precision tracking of cluster targets because it does not take into account the interaction between cluster targets. Summary of the invention

[0004] In order to solve the above problems, one or more embodiments of the present specification provide a method, device and medium for correcting the track of cluster flying targets, which are used to achieve high-precision tracking of cluster flying targets.

[0005] One or more embodiments of this specification adopt the following technical solutions:

[0006] One or more embodiments of the present specification provide a method for correcting the track of a cluster flying target, the method comprising:

[0007] Acquire multi-target point traces detected by a preset radar on cluster flying targets, so as to pre-process the multi-target point traces and obtain initial cluster target tracks of the cluster flying targets;

[0008] Constructing a graph representation of the initial cluster target track, processing the graph representation according to a preset graph convolutional neural network, updating the initial node features of the graph representation according to the processed target interaction information and missed frame correction information, and obtaining updated node features;

[0009] The initial node feature and the updated node feature are integrated to smooth the integrated node feature according to a bidirectional long short-term memory network to obtain a node hidden state;

[0010] The node hidden states belonging to the same target at different times are obtained to assign feature weights to them, to obtain updated node hidden states, and the updated node hidden states are transformed based on the fully connected layer to obtain the corrected cluster target track.

[0011] Optionally, in one or more embodiments of the present specification, the acquiring of multi-target point tracks detected by a preset radar on cluster flying targets to pre-process the multi-target point tracks to obtain initial cluster target tracks of the cluster flying targets specifically includes:

[0012] Acquire detection point traces of cluster flying targets detected by a preset radar, and perform multi-target tracking on the detection point traces based on a multi-hypothesis tracking method to obtain multi-target point traces;

[0013] Performing filtering processing on the multi-target point traces based on a preset Kalman filter to obtain filtered multi-target point traces;

[0014] The multi-target point track is intercepted based on a preset fixed length to obtain the cluster multi-target track segment of the cluster flying target as the initial cluster target track.

[0015] Optionally, in one or more embodiments of the present specification, constructing a graphical representation of the initial cluster target track specifically includes:

[0016] Determine the state matrix of each target of the initial cluster target track according to the cluster multi-target track segment, obtain the state vector of the initial cluster target track at each moment according to the state matrix, and use the state vector as the initial node feature of the graph representation;

[0017] Determine the state of each initial node in the graph representation and the relationship between each initial node based on the characteristics of each initial node;

[0018] The type of each edge in the graph representation is determined according to the relationship between each initial node, and the initial edge feature of the graph representation is determined according to the feature of each initial node; wherein the types include: spatial edge and temporal edge.

[0019] Optionally, in one or more embodiments of the present specification, the spatial edge connects the tracks of adjacent targets at the same time, and the temporal edge connects the missed detection frame and other frames of the same target track; the missed detection frame does not transmit information to other nodes.

[0020] Optionally, in one or more embodiments of the present specification, the graph representation is processed according to a preset graph convolutional neural network, so as to update the initial node features of the graph representation according to the processed target interaction information and missed frame correction information, and obtain the updated node features, specifically including:

[0021] According to the encoding layer of the preset graph convolutional neural network, encoding the initial edge features of the graph representation to obtain processed edge features;

[0022] Respectively averaging and fusing the processed edge features corresponding to the spatial edge and the processed edge features corresponding to the temporal edge to obtain spatial edge aggregation features corresponding to the spatial edge and temporal edge aggregation features corresponding to the temporal edge;

[0023] The spatial edge aggregation feature and the temporal edge aggregation feature are concatenated, and the concatenated edge aggregation feature is encoded to extract the target interaction information and missed frame correction information corresponding to the edge aggregation feature;

[0024] The initial node feature is updated according to the target interaction information and the missed frame correction information to obtain an updated node feature.

[0025] Optionally, in one or more embodiments of the present specification, the initial node feature and the updated node feature are fused to perform smoothing on the fused node feature according to a bidirectional long short-term memory network to obtain a node hidden state, specifically including:

[0026] Processing the initial node features based on the fully connected layer of the preset graph convolutional neural network to obtain track embedding features;

[0027] Adding the track embedding feature to the updated node feature to obtain a fused node feature;

[0028] The bidirectional long short-term memory network is used to extract the temporal correlation of the nodes belonging to the same track at each time step in the fused node features, so as to splice the bidirectional temporal correlation of each time step, realize smoothing processing of the node features, and obtain the processed node hidden state.

[0029] Optionally, in one or more embodiments of the present specification, obtaining the node hidden states belonging to the same target at different times to perform feature weight assignment thereto to obtain updated node hidden states specifically includes:

[0030] Based on a preset formula, feature weights are assigned to the node hidden states at different times belonging to the same target to obtain updated node hidden states;

[0031] Wherein, the preset formula is:

[0032] , represents the hidden state of the i-th target at time k after the temporal attention layer ta, , i represents the target number, k represents the time number, is the weight coefficient of the i-th target at time k, are the training parameters of the network, represents the node hidden state of the i-th target at time k after passing through the bidirectional long short-term memory network l;

[0033] After obtaining the updated node hidden state, the method further includes:

[0034] According to the updated node hidden state, an updated node hidden state matrix is ​​obtained.

[0035] Optionally, in one or more embodiments of the present specification, the step of converting the updated node hidden state based on the fully connected layer to obtain a corrected cluster target track specifically includes:

[0036] The updated node hidden state is converted based on a preset conversion formula of the fully connected layer to obtain a corrected track segment; wherein the preset conversion formula is: , represents the fully connected layer, represents the updated node hidden state, is the corrected track segment, o indicates that the vector x is the network output;

[0037] The corrected track segments are spliced ​​in sequence to obtain a corrected cluster target track.

[0038] One or more embodiments of this specification provide a track correction device for cluster flying targets, the device comprising:

[0039] at least one processor; and,

[0040] a memory communicatively connected to the at least one processor; wherein,

[0041] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform any of the above methods.

[0042] One or more embodiments of the present specification provide a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute any of the above-described methods.

[0043] At least one of the above technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects:

[0044] Preprocessing the multi-target point tracks to obtain the initial cluster target track helps to remove interference factors such as noise and outliers in the original data, and improves the reliability of the entire subsequent track correction process. By constructing a graph representation of the initial cluster target track and processing it with the help of a preset graph convolutional neural network, the complex interaction information between targets can be implicitly mined, solving the problem that traditional methods often find it difficult to effectively capture the collaborative interaction relationship between cluster targets, resulting in inaccurate track tracking. In addition, using the interaction information between targets and the correction information for missed frames to update the initial node features not only considers the dynamic relationship between targets, but also corrects possible missed detections, so that the updated node features can more comprehensively reflect the true state of the target. Further smoothing the track through a bidirectional long short-term memory network and a temporal attention mechanism solves the problem of reduced tracking accuracy caused by the mismatch between the traditional single target filter model and the group target model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0046] Figure 1 A schematic flow chart of a method for correcting the track of a cluster of flying targets provided in an embodiment of this specification;

[0047] Figure 2 A schematic diagram of a processing process of a method for correcting the track of a cluster flying target provided in an embodiment of this specification;

[0048] Figure 3 A schematic diagram of a conventional track correction result of a cluster flying target provided in an embodiment of this specification;

[0049] Figure 4 A schematic diagram of a track correction result of a cluster flying target provided in an embodiment of this specification;

[0050] Figure 5 A graph showing the comparison of tracking accuracy of a cluster flying target track correction method provided in an embodiment of this specification;

[0051] Figure 6 A schematic diagram of the structure of a track correction device for cluster flying targets provided in an embodiment of this specification;

[0052] Figure 7 A schematic diagram of the structure of a non-volatile storage medium provided in an embodiment of this specification. DETAILED DESCRIPTION

[0053] The embodiments of this specification provide a method, device and medium for correcting the track of clustered flying targets.

[0054] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0055] like Figure 1 As shown, the present specification embodiment provides a flow chart of a method for correcting the track of a cluster flying target. Figure 1 It can be seen that in one or more embodiments of this specification, a method for correcting the track of a cluster flying target includes the following steps:

[0056] S101: Acquire multi-target point tracks detected by a preset radar on cluster flying targets, pre-process the multi-target point tracks, and obtain initial cluster target tracks of the cluster flying targets.

[0057] like Figure 2 In order to obtain the group target track, in the embodiment of this specification, the multi-target point track obtained by the preset radar detecting the cluster flying target is first obtained, and then the multi-target point track is preprocessed to improve the tracking accuracy of the cluster flying target, so as to obtain the initial cluster target track of the cluster flying target. In this process, specifically, in one or more embodiments of this specification, the multi-target point track detected by the preset radar on the cluster flying target is obtained, and the multi-target point track is preprocessed to obtain the initial cluster target track of the cluster flying target, which specifically includes the following process:

[0058] First, the detection point traces of the cluster flying targets detected by the preset radar are obtained. Since it is a cluster flying target, there are multiple targets flying in the air at the same time. It is difficult to determine the complete motion trajectory of each target by relying on a single detection point trace. At this time, a multi-hypothesis tracking method is adopted. The core idea of ​​this method is to consider multiple possible target association assumptions for each detection point trace. Because the detection point traces obtained by the radar each scan may come from different targets, and due to the influence of factors such as noise and target maneuverability, it is difficult to directly and accurately determine which point traces belong to the same target, so in the embodiment of this specification, the detection point traces will be multi-target tracked according to the multi-hypothesis tracking method to obtain multi-target point traces, so as to achieve tracking of multiple targets in the cluster. Then, in order to remove the interference of factors such as noise in the multi-target point traces, the multi-target point traces will be filtered through a preset Kalman filter to obtain the filtered multi-target point traces. After that, the multi-target point traces are intercepted according to the preset fixed length to obtain the cluster multi-target track segment of the cluster flying target as the initial cluster target track. For example, the track segment of the fixed length K forward from the new moment is intercepted to obtain the cluster multi-target track segment set. ,in, represents the state matrix of target i, and Represents the state vector at each moment, which contains the position of the target in the Cartesian coordinate system, that is, and speed status information .

[0059] S102: Construct a graph representation of the initial cluster target track, process the graph representation according to a preset graph convolutional neural network, update the initial node features of the graph representation according to the processed target interaction information and missed frame correction information, and obtain updated node features.

[0060] Since there may be complex interactive behaviors such as coordinated movement and mutual avoidance between different target tracks in the cluster target tracking scenario, in order to more effectively mine these relationship information and better understand the interaction between targets. In the embodiments of this specification, a graph representation of the initial cluster target track is constructed, and the initial node features of the graph representation are updated according to the processed target interaction information and missed frame correction information, so as to obtain updated node features.

[0061] Specifically, in one or more embodiments of the present specification, constructing a graph representation of the initial cluster target track specifically includes the following process:

[0062] First, according to the cluster multi-target track segment, the state matrix of each target of the initial cluster target track is determined, so as to obtain the state vector of the initial cluster target track at each moment according to the state matrix, and use the state vector as the initial node feature of the graph representation. At the same time, the state of each initial node in the graph representation and the relationship between each initial node are determined according to the characteristics of each initial node. The type of each edge in the graph representation is determined according to the relationship between each initial node, and the initial edge feature of the graph representation is determined according to the characteristics of each initial node; it should be noted that the types include: spatial edges and temporal edges, both of which are directed edges. Spatial edges connect the tracks of adjacent targets at the same time, and temporal edges connect the missed frames of the same target track with other frames; missed frames do not transmit information to other nodes.

[0063] That is, in a certain scenario, a graph representation of the cluster trajectory G = (V, E) is first constructed, and then the state vector of the cluster multi-target trajectory is used as the initial node feature of the network , each node and edge It is calculated based on the node status and the relationship between nodes. And the time difference, the initial edge features can be determined as:

[0064] , is the initial node feature and initial node features The characteristic difference between is the initial node feature and initial node features The time difference between them is obtained by obtaining the initial edge feature, which helps to input the edge feature into the graph neural network later, so as to extract the interaction information between targets and the missed frame correction information through the graph neural network, and obtain the updated track node feature to realize the correction of the track.

[0065] Specifically, in one or more embodiments of the present specification, the graph representation is processed according to the preset graph convolutional neural network, so as to update the initial node features of the graph representation according to the processed target interaction information and missed frame correction information, and obtain the updated node features, which specifically includes the following process:

[0066] According to the encoding layer of the preset graph convolutional neural network, the initial edge features of the graph representation are encoded to obtain the processed edge features. Figure 2 The graph neural network shown in the figure includes an edge feature encoding layer and a node feature encoding layer. Both layers of the network are multi-layer perceptrons. In the embodiment of this specification, the initial edge feature is input into the encoding layer. , and get the processed edge features:

[0067] ;

[0068] in, is the processed edge feature, is the initial edge feature.

[0069] In this process, since the spatial edge connects the tracks of neighboring targets at the same time, the edge feature encoding layer can mine more complex spatial relationships between targets, such as relative position change trends, relative speeds and other information. For the time edge, it connects the missed frame node and other frame nodes of the same target track. At this time, the encoding layer can extract relationship features such as the strength of association between different frames in the time series and the importance of information transmission. Therefore, based on this process, the interactive information between targets and the time association information related to the missed frames can be obtained, which is helpful for the subsequent update of track node features. Then, the processed edge features corresponding to the above-mentioned spatial edges and the processed edge features corresponding to the time edges are averaged and fused respectively to obtain the spatial edge aggregation features corresponding to the spatial edges and the temporal edge aggregation features corresponding to the temporal edges. That is, the features of the two types of edges are averaged and fused to obtain the edge aggregation features:

[0070] ;

[0071] ;

[0072] in, represents the set of spatial edges connecting node i, represents the set of time edges connecting node i, represents the edge aggregation characteristics of the spatial edge, The edge aggregation features of the time edge are represented. The spatial edge aggregation features and the temporal edge aggregation features are concatenated, and the concatenated edge aggregation features are encoded to extract the target interaction information and missed frame correction information corresponding to the edge aggregation features. Then, according to the target interaction information and missed frame correction information, the initial node features are updated to obtain the updated node features. That is, the aggregation features of the two types of edges are concatenated and passed through the node feature encoding layer. Get the updated node features:

[0073] ;

[0074] in, Indicates that it has passed the node feature encoding layer The updated node features are obtained.

[0075] S103: Fusing the initial node features with the updated node features, and smoothing the fused node features according to a bidirectional long short-term memory network to obtain a node hidden state.

[0076] After obtaining the updated node features based on the above steps, the updated node features are combined with the nodes after passing through a fully connected layer. The initial node features are fused, that is, the initial node features are fused with the updated node features as the input of the next layer, which is expressed as:

[0077] ;

[0078] in, The fused node. In this process, the initial node features and the updated node features are fused, and the original features and the new features obtained after correction and mining are combined, so that the features finally used for subsequent processing contain both the basic state information of the target at the beginning and the key information after considering the target interaction and missed detection correction, which enriches the feature connotation and improves the comprehensiveness of the target description.

[0079] Then, in order to solve the problem of decreased tracking accuracy caused by the mismatch between the traditional single target filtering model and the group target model, the fused node features will be further smoothed according to the bidirectional long short-term memory network and the temporal attention mechanism to obtain the node hidden state.

[0080] Specifically, in one or more embodiments of the present specification, the initial node features and the updated node features are fused, and the fused node features are smoothed according to the bidirectional long short-term memory network to obtain the node hidden state, which specifically includes the following process:

[0081] Then, the initial node features are processed based on the fully connected layer of the preset graph convolutional neural network to obtain the track embedding features. The track embedding features are added to the updated node features to obtain the fused node features. At the same time, the bidirectional long short-term memory network is used to extract the temporal correlation of the nodes belonging to the same track at each time step in the fused node features, so as to splice the bidirectional temporal correlation of each time step to achieve smooth processing of the node features and obtain the processed node hidden state. In other words, the nodes belonging to the same track are extracted based on the bidirectional long short-term memory network. At each time step The output vector at time k is composed of the output of the forward LSTM and the backward LSTM, so the node hidden state obtained at this time is .

[0082] S104: Acquire the node hidden states belonging to the same target at different times to assign feature weights to them, obtain updated node hidden states, and transform the updated node hidden states based on the fully connected layer to obtain a corrected cluster target track.

[0083] After the node hidden state is obtained based on the above step S103, in order to more accurately capture the dynamic changes of the target over time and more accurately describe the real state of the target, the node hidden state belonging to the same target at different times will be obtained, so as to assign feature weights to the node hidden state belonging to the same target at different times, thereby obtaining the updated node hidden state. That is, the node hidden state of the same target at different times is obtained and the feature weights are assigned, which can highlight the key features according to the performance and importance of the target at different time stages, and adaptively adjust the characterization of the target state, so that the updated node hidden state is more in line with the actual state changes of the target.

[0084] Specifically, in one or more embodiments of the present specification, obtaining node hidden states belonging to the same target at different times to assign feature weights to them and obtain updated node hidden states specifically includes the following process:

[0085] Based on the preset formula, the feature weights of the node hidden states belonging to the same target at different times are assigned to obtain the updated node hidden states.

[0086] The preset formula is:

[0087] ;

[0088] ;

[0089] in, represents the updated node hidden state, express, are the training parameters of the network, Represents the node hidden state of node i at time k.

[0090] After obtaining the updated node hidden state, the method further includes:

[0091] According to the updated node hidden state, the updated node hidden state matrix is ​​obtained, that is, the node hidden state matrix belonging to the same target at different times is obtained. Each feature weight is assigned to each, and the updated hidden state matrix is ​​obtained .

[0092] Further, in one or more embodiments of the present specification, the updated node hidden state is converted based on the fully connected layer to obtain a corrected cluster target track, specifically including:

[0093] The updated node hidden state is converted based on a preset conversion formula of the fully connected layer to obtain a corrected track segment; wherein the preset conversion formula is: , represents the fully connected layer, represents the updated node hidden state, is the corrected track segment. The corrected track segments are spliced ​​in sequence to obtain the corrected cluster target track. In addition, it should be noted that the loss function of the preset graph convolutional neural network is

[0094] ;

[0095] in, is the filtering correction result output by the network, is the true value of the trajectory. That is, by calculating the sum of the squares of the differences between the network output and the true trajectory, the loss function can intuitively reflect the error size of the network in the process of correcting the target track. This helps to iteratively train and optimize the parameters of the preset graph neural network, so that the network can learn in the direction of reducing errors and correcting the target track more accurately.

[0096] In addition, in order to intuitively demonstrate the effect of the correction method, in an embodiment of this specification, 10 target tracks are generated based on the group target motion model simulation. The distance between each target is about 10m, the overall speed of the group is about 20m / s, the speed noise variance is, and the motion model is first a CV model, which is converted to a CT model after 8s. The initial position of the target is Figure 3 and Figure 4 As shown by the black circle in the figure. The sampling interval is 1s, the track has 20 moments in total, and the length of the track segment sliding window processing is 8. The standard deviation of the radar measurement noise is 0.3m, and the detection probability is set to 0.9. At this time, first, assuming that the target can be correctly associated during the tracking process, the Kalman filter is used to filter the target track to obtain the traditional tracking track. The target real track, radar measurement and traditional tracking track are shown in Figure 3 The present invention is applied to cluster target tracking experiments, and the corrected track results are shown in Figure 4 As shown in Figure 1, it can be seen that the track output by the present invention is closer to the true value, and the error is smaller at the missed frame. The track tracking error at each moment is calculated, and the error curve is shown in Figure 1. Figure 5 As shown, it can be seen that the method of the present invention has a significant improvement compared with the original track, and is superior to the existing long short-term memory network correction method.

[0097] like Figure 6 As shown, the present specification embodiment provides a schematic diagram of the structure of a track correction device for clustered flying targets. Figure 6 It can be seen that in one or more embodiments of this specification, a track correction device for cluster flying targets includes:

[0098] at least one processor; and,

[0099] a memory communicatively connected to the at least one processor; wherein,

[0100] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform any of the above methods.

[0101] like Figure 7 As shown, the present specification provides a schematic diagram of the structure of a non-volatile storage medium. Figure 7 It can be seen that in one or more embodiments of the present specification, a non-volatile storage medium stores computer executable instructions 701, and the computer executable instructions 701 can execute any of the above-mentioned methods.

[0102] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0103] The above specific embodiments of the present specification are described. In some cases, the actions or steps recorded in the specification can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the process depicted in the accompanying drawings does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] The above is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification should be included in the scope of this specification.

Claims

1. A method for correcting the track of clustered flying targets, characterized in that: The method comprises: Acquire multi-target point traces detected by a preset radar on cluster flying targets, so as to pre-process the multi-target point traces and obtain initial cluster target tracks of the cluster flying targets; Constructing a graph representation of the initial cluster target track, processing the graph representation according to a preset graph convolutional neural network, updating the initial node features of the graph representation according to the processed target interaction information and missed frame correction information, and obtaining updated node features; The initial node feature and the updated node feature are integrated to smooth the integrated node feature according to a bidirectional long short-term memory network to obtain a node hidden state; Obtaining the node hidden states belonging to the same target at different times to assign feature weights to them, obtaining updated node hidden states, and converting the updated node hidden states based on the fully connected layer to obtain a corrected cluster target track; The graph representation is processed according to a preset graph convolutional neural network to update the initial node features of the graph representation according to the processed target interaction information and missed frame correction information to obtain updated node features, specifically including: According to the encoding layer of the preset graph convolutional neural network, encoding the initial edge features of the graph representation to obtain processed edge features; The processed edge features corresponding to the spatial edge and the processed edge features corresponding to the temporal edge are respectively averaged and fused to obtain the spatial edge aggregation features corresponding to the spatial edge and the temporal edge aggregation features corresponding to the temporal edge; The spatial edge aggregation feature and the temporal edge aggregation feature are concatenated, and the concatenated edge aggregation feature is encoded to extract the target interaction information and missed frame correction information corresponding to the edge aggregation feature; The initial node feature is updated according to the target interaction information and the missed frame correction information to obtain an updated node feature.

2. The method for correcting the track of cluster flying targets according to claim 1, characterized in that: The step of obtaining multi-target point traces detected by a preset radar on cluster flying targets to pre-process the multi-target point traces and obtain an initial cluster target track of the cluster flying targets specifically includes: Acquire detection point traces of cluster flying targets detected by a preset radar, and perform multi-target tracking on the detection point traces based on a multi-hypothesis tracking method to obtain multi-target point traces; Performing filtering processing on the multi-target point traces based on a preset Kalman filter to obtain filtered multi-target point traces; The multi-target point track is intercepted based on a preset fixed length to obtain the cluster multi-target track segment of the cluster flying target as the initial cluster target track.

3. The method for correcting the track of cluster flying targets according to claim 2, characterized in that: Constructing a graphical representation of the initial cluster target track, specifically including: Determine the state matrix of each target of the initial cluster target track according to the cluster multi-target track segment, obtain the state vector of the initial cluster target track at each moment according to the state matrix, and use the state vector as the initial node feature of the graph representation; Determine the state of each initial node in the graph representation and the relationship between each initial node based on the characteristics of each initial node; The type of each edge in the graph representation is determined according to the relationship between each initial node, and the initial edge feature of the graph representation is determined according to the feature of each initial node; wherein the types include: spatial edge and temporal edge.

4. The method for correcting the track of cluster flying targets according to claim 3, characterized in that: The spatial edge connects the tracks of neighboring targets at the same time, and the temporal edge connects the missed detection frame and other frames of the same target track; the missed detection frame does not transmit information to other nodes.

5. The method for correcting the track of cluster flying targets according to claim 1, characterized in that: The initial node feature and the updated node feature are integrated to smooth the integrated node feature according to the bidirectional long short-term memory network to obtain the node hidden state, specifically including: Processing the initial node features based on the fully connected layer of the preset graph convolutional neural network to obtain track embedding features; Adding the track embedding feature to the updated node feature to obtain a fused node feature; The bidirectional long short-term memory network is used to extract the temporal correlation of the nodes belonging to the same track at each time step in the fused node features, so as to splice the bidirectional temporal correlation of each time step, realize smoothing processing of the node features, and obtain the processed node hidden state.

6. The method for correcting the track of cluster flying targets according to claim 1, characterized in that: The acquiring of the node hidden states belonging to the same target at different times to perform feature weight assignment on the node hidden states to obtain updated node hidden states specifically includes: Based on a preset formula, feature weights are assigned to the node hidden states at different times belonging to the same target to obtain updated node hidden states; Wherein, the preset formula is: , represents the hidden state of the i-th target at time k after the temporal attention layer ta, , i represents the target number, k represents the time number, is the weight coefficient of the i-th target at time k, are the training parameters of the network, represents the node hidden state of the i-th target at time k after passing through the bidirectional long short-term memory network l; After obtaining the updated node hidden state, the method further includes: According to the updated node hidden state, an updated node hidden state matrix is ​​obtained.

7. The method for correcting the track of cluster flying targets according to claim 3, characterized in that: The step of converting the updated node hidden state based on the fully connected layer to obtain a corrected cluster target track specifically includes: The updated node hidden state is converted based on a preset conversion formula of the fully connected layer to obtain a corrected track segment; wherein the preset conversion formula is: , represents the fully connected layer, represents the updated node hidden state, is the corrected track segment, o indicates that the vector x is the network output; The corrected track segments are spliced ​​in sequence to obtain a corrected cluster target track.

8. A track correction device for cluster flying targets, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any one of the methods described in claims 1 to 7.

9. A non-volatile storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 7.

Images