A multi-clustering structure unmanned aerial vehicle swarm countermeasure method based on motion trajectory mining

Abstract

The present application belongs to the technical field of unmanned aerial vehicle defense, and particularly relates to a multi-cluster structure unmanned aerial vehicle group countermeasure method based on motion trajectory mining. Step 1: obtaining an unmanned aerial vehicle group target image sequence, adopting an unmanned aerial vehicle group sub-cluster detection algorithm, and extracting unmanned aerial vehicle group sub-clusters in the image; Step 2: for the unmanned aerial vehicle group sub-clusters, adopting an unmanned aerial vehicle group target tracking algorithm, extracting the motion trajectories of each unmanned aerial vehicle in the cluster, and forming unmanned aerial vehicle group member trajectory clusters; Step 3: for the unmanned aerial vehicle group member trajectory clusters, establishing a trajectory similarity model, identifying the cluster head unmanned aerial vehicle of each sub-cluster by comparing the differences between the motion trajectories in the cluster; Step 4: for the cluster head unmanned aerial vehicle of the sub-cluster, taking targeted countermeasures, and realizing the directional driving away of the unmanned aerial vehicle group. Compared with the traditional one-by-one countermeasure method, the present application has the advantages of strong operability, high economic benefit, small equipment burden, etc., and can effectively cope with the multi-cluster structure unmanned aerial vehicle group saturation countermeasure problem.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of unmanned aerial vehicle defense, and particularly relates to a multi-cluster structure unmanned aerial vehicle group countermeasure method based on motion trajectory mining. BACKGROUND

[0002] The multi-cluster structure unmanned aerial vehicle group refers to a mesh group with multiple cluster structures formed by multiple unmanned aerial vehicles. Each sub-cluster can be regarded as an independent structure, and different sub-clusters allocate tasks through information interaction and negotiation. Each sub-cluster includes one or more cluster leader aircrafts, and the cluster leader aircraft is responsible for communication between the cluster leader aircraft and the follower aircrafts in the cluster and interaction between the follower aircrafts in the cluster and other cluster unmanned aerial vehicles. The follower aircrafts in the cluster adjust their states in real time with the cluster leader aircraft as the reference object to achieve the purpose of target tracking and group control. The one-way information transmission structure from the virtual target to the cluster leader aircraft and from the cluster leader aircraft to the follower aircrafts enables the multi-cluster structure unmanned aerial vehicle group to realize intelligent cooperation more quickly and efficiently, and has strong fault tolerance and combat response capability. In the military field, the multi-cluster structure unmanned aerial vehicle group can carry out high-density reconnaissance and saturation attack on the target area with quantity advantage, which poses a great threat to the air defense system. This combat mode has become a new combat mode and has attracted widespread attention from the world military powers. In order to offset the threat of the multi-cluster structure unmanned aerial vehicle group, it is necessary to actively carry out research on unmanned aerial vehicle group countermeasures.

[0003] The current countermeasure means for the unmanned aerial vehicle group still follows the idea of countering single unmanned aerial vehicle, that is, each member in the group is counteracted one by one. This method is only applicable to small-scale unmanned aerial vehicle groups. When facing the multi-cluster structure unmanned aerial vehicle group, since the number and scale of the unmanned aerial vehicle group are large, the method of countering each unmanned aerial vehicle is difficult to achieve effective countermeasures. At this time, how to use the internal emergence mechanism of the multi-cluster structure unmanned aerial vehicle group for targeted overall countermeasures becomes a key problem to be solved. SUMMARY

[0004] In the multi-cluster structure unmanned aerial vehicle group, the unmanned aerial vehicle group sub-cluster is the smallest combat system for executing tasks, and the cluster members fly around the cluster leader aircraft for escort. The cluster leader aircraft, as the command center, control center and communication center of the sub-cluster, plays a decisive role in maintaining the stability of the sub-cluster. In addition, the cooperation between different sub-clusters is also carried out through communication and negotiation between the cluster leader aircrafts. Therefore, the cluster leader aircraft is the key node for countering the multi-cluster structure unmanned aerial vehicle group. As long as the cluster leader aircraft is accurately identified and countermeasures are carried out, the effect of countermeasures can be achieved.

[0005] Based on this, the present application aims to provide a multi-cluster structure unmanned aerial vehicle group countermeasure method based on motion trajectory mining. The method can accurately identify the positions of the cluster leaders of the multi-cluster structure unmanned aerial vehicle group through motion trajectory mining, and can achieve the overall countermeasure effect of point-to-surface by carrying out countermeasures on the cluster leaders.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] A multi-cluster structure unmanned aerial vehicle group countermeasure method based on motion trajectory mining, comprising:

[0008] Step 1: Obtain the unmanned aerial vehicle group target image sequence, and extract the unmanned aerial vehicle group sub-cluster in the image frame by frame by using the unmanned aerial vehicle group sub-cluster detection algorithm;

[0009] Step 2: For the unmanned aerial vehicle group sub-cluster of step 1, use the unmanned aerial vehicle group target tracking algorithm to extract the motion trajectory of each unmanned aerial vehicle in the cluster, and form the unmanned aerial vehicle group member trajectory cluster;

[0010] Step 3: For the unmanned aerial vehicle group member trajectory cluster of step 2, establish a trajectory similarity model, and identify the cluster leader unmanned aerial vehicle of each sub-cluster by comparing the differences between the motion trajectories in the cluster;

[0011] Step 4: For the cluster leader unmanned aerial vehicle of the sub-cluster of step 3, take targeted countermeasures to realize the directional displacement of the unmanned aerial vehicle group.

[0012] Preferably, the step 1 comprises:

[0013] Step 1.1: Obtain the aerial unmanned aerial vehicle group image sequence I={I i |i=1,2,...,N} by using the ground optical imaging device, wherein N represents the total number of frames;

[0014] Step 1.2: For the unmanned aerial vehicle group target image sequence I of step 1.1, use the unmanned aerial vehicle group sub-cluster detection algorithm based on attention mechanism to calculate the sub-cluster distribution prediction graph Y={Y i |i=1,2,...,N} frame by frame;

[0015] Step 1.3: For the sub-cluster distribution probability graph Y i generated in step 1.2, use the threshold segmentation algorithm τ to extract the unmanned aerial vehicle group sub-cluster M represents the total number of sub-clusters.

[0016] Preferably, the unmanned aerial vehicle group sub-cluster detection algorithm based on attention mechanism of step 1.2 adopts an improved U-Net structure, which includes an encoder, an enhancer and a decoder, and the input is the i th frame unmanned aerial vehicle group image I i, the i-th frame of the unmanned aerial vehicle group sub-cluster distribution probability graph Y is output by deep coding, feature enhancement and information decoding i , the calculation process of the algorithm is as follows:

[0017]

[0018] Wherein, is an encoder function for encoding the depth features of the unmanned aerial vehicle group image; is a decoder function for decoding the depth features of the unmanned aerial vehicle group image to generate the unmanned aerial vehicle group sub-cluster distribution prediction graph; and φ is an enhancer function for suppressing high-frequency signal interference of the depth features of the unmanned aerial vehicle group image and enhancing the expression ability of task-related features.

[0019] Preferably, the step 2 comprises:

[0020] Step 2.1: for the j-th unmanned aerial vehicle sub-cluster Taking the sub-cluster member unmanned aerial vehicle as a node V={v1, v2,...v n , the action relationship between the members is an edge E(i,j), the group consistency size is the weight w(i,j) of the edge, the unmanned aerial vehicle group structure graph G=(V,E,w) is constructed, and the random matrix is used to model the extended shape of the unmanned aerial vehicle sub-cluster;

[0021] Step 2.2: for the unmanned aerial vehicle group sub-cluster in step 2.1 The Kalman filtering algorithm is used to predict the centroid position and the extended shape of the unmanned aerial vehicle group in the next frame, and the range of N times the extended target size near the predicted unmanned aerial vehicle group centroid is selected as the search area;

[0022] Step 2.3: for the unmanned aerial vehicle group structure graph in the search area The spectral clustering algorithm is used for subgraph division to obtain the unmanned aerial vehicle sub-group with high internal consistency Wherein, P is the total number of sub-groups, and g k is taken as a candidate unmanned aerial vehicle sub-group in the search area;

[0023] Step 2.4: for the candidate unmanned aerial vehicle sub-group g k in step 2.3, a twin graph convolution network model is constructed, the similarity between the members of the unmanned aerial vehicle group and the members of the previous frame of the unmanned aerial vehicle group is calculated, and the unmanned aerial vehicle group member tracking is realized.

[0024] Preferably, the step 2.4 is specifically:

[0025] The twin graph convolution network model includes a template branch and a candidate branch, which respectively input the template unmanned aerial vehicle sub-group structure graph g t and the candidate unmanned aerial vehicle sub-group structure graph g kWherein, the template UAV subgroup structure diagram is the UAV subgroup structure diagram of the previous frame, the deep topological features of g t and g k are extracted respectively through the graph convolution network with weight sharing, to generate UAV subgroup member embedding vectors F t and F k The constructed graph convolution network includes three graph convolution layers, and the information aggregation mode of the graph convolution layer is as follows:

[0026]

[0027] Wherein, F l represents the member feature vector of the lth layer UAV subgroup; represents the weighted self-loop adjacency matrix of the UAV subgroup structure diagram; is the degree matrix of the UAV subgroup structure diagram; W l is the feature transformation matrix;

[0028] Based on the UAV subgroup member embedding vector, the group member similarity matrix between the template UAV subgroup and the candidate UAV subgroup is calculated Finally, the nearest neighbor matching algorithm is adopted, the most similar group members in the two UAV subgroups are associated according to the group member similarity matrix, to form the group member matching pair, and the matched node pairs are connected through iterative processing of the time sequence image, to form the UAV group sub-cluster member motion trajectory T rack .

[0029] Preferably, the step 3 includes:

[0030] Step 3.1: For the UAV group sub-cluster member motion trajectory T rack , the member flight trajectory segment is intercepted through the sliding time window, the structural feature information of the UAV group sub-cluster member motion trajectory is extracted, and the feature weight vector W = {W D ,W S ,W A ,W L ,W T} is constructed, the elements in the vector correspond to the direction D, speed S, angle A, position L and time T of the trajectory respectively, each weight value is greater than or equal to zero, and W D +W S +W A +W L +W T =1, the sensitivity of the trajectory structure feature attribute is adjusted through the feature weight, and the trajectory structure similarity is the difference degree similarity between the trajectory segments, for the trajectory segment sample L, the trajectory structure similarity is calculated by the following formula:

[0031] DIST(L) = D dist ×WD +S dist ×W S +A dist ×W A +L dist ×W L +T dist ×W T ;

[0032] Step 3.2: structure comparison of all segments of different trajectories by using structural similarity, DIST(L) represents the structural difference of trajectory segments, the smaller the value, the more similar the trajectory segments, and vice versa, the less similar the trajectory segments, the trajectory segments with similarity less than threshold T h are selected as the motion trajectory mining segment, the motion trajectory of which contains obvious differences between the leader and follower, and the unmanned aerial vehicle group member with obvious motion trajectory change is taken as the cluster head unmanned aerial vehicle C drones .

[0033] Preferably, the step 4 is specifically:

[0034] For each cluster head unmanned aerial vehicle C drones , a false navigation deception signal is sent to make the cluster head unmanned aerial vehicle deviate from the expected task target position, and the unmanned aerial vehicles in the cluster take the actual position of the cluster head unmanned aerial vehicle as a reference and do not directly receive the virtual target information, so the unmanned aerial vehicles in the cluster will follow the cluster head unmanned aerial vehicle to deviate from the expected task trajectory, thereby realizing the directional expulsion of the unmanned aerial vehicle group sub-cluster.

[0035] Compared with the prior art, the beneficial effects of the present application are:

[0036] The present application proposes a multi-cluster structure unmanned aerial vehicle group countermeasure method based on motion trajectory mining, first, an saliency target detection algorithm is used to segment the sub-cluster distribution area of the unmanned aerial vehicle group from the background, then the multi-target tracking algorithm is used to obtain the motion trajectory of the cluster member unmanned aerial vehicle, then a trajectory similarity measurement model is constructed, the cluster head unmanned aerial vehicle of each sub-cluster is found out by comparing the trajectory similarity of each unmanned aerial vehicle in the cluster, and finally the countermeasure is carried out for each sub-cluster cluster head unmanned aerial vehicle, so as to achieve the overall countermeasure effect of point to surface. Compared with the traditional one-by-one countermeasure method, the present application has the advantages of strong operability, high economic benefit, small equipment burden, etc., and can effectively cope with the multi-cluster structure unmanned aerial vehicle group saturation countermeasure problem, and has important significance for unmanned aerial vehicle group countermeasure. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application.

[0038] In the drawings:

[0039] Figure 1 The method flow is implemented as an embodiment of the application.

[0040] Figure 2 The method flow is implemented as an embodiment of the application.

[0041] Figure 3 The multi-cluster structure UAV group countermeasure effect.

[0042] Figure 4 The sub-cluster area detection result.

[0043] Figure 5 The sub-cluster member motion trajectory extraction result. DETAILED DESCRIPTION

[0044] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0045] Embodiment:

[0046] Referring to the accompanying drawings, Figures 1-5 A multi-cluster structure UAV group countermeasure method based on motion trajectory mining, comprising:

[0047] Step 1: Obtain the UAV group target image sequence, and extract the UAV group sub-cluster in the image frame by frame by using the UAV group sub-cluster detection algorithm. Specifically:

[0048] Step 1.1: Use a ground optical imaging device (visible light camera or infrared camera) to obtain an aerial UAV group image sequence I={I i |i=1,2,...,N}, N represents the total number of frames;

[0049] Step 1.2: For the UAV group target image sequence I of step 1.1, use the UAV group sub-cluster detection algorithm based on attention mechanism to calculate the UAV group sub-cluster distribution prediction map Y={Y i |i=1,2,...,N} frame by frame;

[0050] Further, the UAV group sub-cluster detection algorithm based on attention mechanism of step 1.2 uses an improved U-Net structure, including an encoder, an enhancer and a decoder, inputs the i-th frame of UAV group image I i , and outputs the i-th frame of UAV group sub-cluster distribution probability map Y i through deep encoding, feature enhancement and information decoding, and the calculation process of the algorithm is:

[0051]

[0052] wherein, is an encoder function for encoding the UAV swarm image depth feature; is a decoder function for decoding the UAV swarm image depth feature to generate the UAV swarm sub-cluster distribution prediction map; φ is an enhancer function for suppressing high-frequency signal interference of the UAV swarm image depth feature and enhancing the expression ability of task-related features.

[0053] The step 1.2 encoder function The input is the UAV swarm image I i , and the output is the UAV swarm image depth feature F i d The encoder structure is composed of four convolutional blocks in series, each convolutional block has two convolutional layers, one batch normalization layer and one ReLU activation layer, and the last layer adopts maximum pooling for feature down-sampling, and the calculation process of the encoder function is:

[0054]

[0055] The input of the enhancer function φ is the UAV swarm image depth feature F i d , and the output is the UAV swarm image depth enhanced feature To alleviate the interference of high-frequency signals at the bottom, the UAV swarm image depth feature enhancer is designed. Before the UAV swarm image depth feature is input into the decoder, the feature enhancer is used to strengthen the representation weight of the task-related features and weaken the background noise interference. The feature enhancer includes a spatial enhancer φ S and a channel enhancer φ C , and the calculation process is:

[0056]

[0057] The spatial enhancer φ S is realized by inter-pixel attention, and the input UAV swarm image depth feature F R represents the real number field, and first, the feature map is expanded by row to arrange a two-dimensional feature map Then the correlation between each pixel point in the feature map is calculated to obtain the inter-pixel attention Finally, the is multiplied by F i v pixel by pixel, and is rearranged back to the input feature map size to obtain the spatial enhanced feature Wherein, π represents the rearrangement operation, represents the Hadamard product.

[0058] The channel enhancer φ S is realized by inter-channel attention. Given the input spatial enhanced feature First, the global average pooling is used to extract the global representation of each channel of the spatial enhanced feature; then, the attention vector between channels is obtained through the nonlinear transformation of the fully connected layer where G AP represents the global average pooling operation, M LP represents the fully connected layer. Finally, V i is used as the weight to determine the importance of each channel to obtain the deep enhanced feature of the UAV group image where denotes the dot product.

[0059] The input of the decoder function is the deep enhanced feature of the UAV group image and the output is the sub-cluster distribution probability map Y i of the UAV group. The decoder structure has the same number of deconvolution blocks as the encoder, and each deconvolution block also contains the same number of deconvolution layers, batch normalization layers and activation layers. The last layer performs upsampling and fuses with the corresponding decoder feature. The calculation process of the decoder function is as follows:

[0060]

[0061] Step 1.3: Using threshold segmentation algorithm τ to extract the sub-cluster distribution probability map Y i generated in step 1.2 to extract the sub-cluster distribution of the UAV group M represents the total number of sub-clusters.

[0062] Using Otsu threshold segmentation algorithm to extract the binary graph of the sub-cluster distribution area from the prediction map Y i generated in step 1.2. Let the total number of pixels in the prediction map be N, the gray scale range be [0, 255], and the frequency of gray scale i be n i ,i = 0, 1, 2,..., 255. The pixel points in the graph can be divided into two classes by threshold T, i.e. C1 ∈ [0, T] and C2 ∈ [T+1, 255], and the number of pixels in C1 is N1 and the number of pixels in C2 is N2. The mean values of C1 and C2 are respectively:

[0063]

[0064] Let

[0065]

[0066] Then the mean value of the whole graph is:

[0067] μ = ω1μ1 + ω2μ2

[0068] Define the inter-class variance as:

[0069] σ2 = ω1(μ1 - μ) 2 + ω2(μ2 - μ) 2 = ω1ω2(μ1 - μ) 2

[0070] Iterate all gray levels in turn, set the threshold T1 with the maximum inter-class variance as the Otsu method optimal threshold. Use the threshold to binarize the distribution probability graph, extract the unmanned aerial vehicle group sub-cluster As Figure 2 shown.

[0071] Step 2: For the unmanned aerial vehicle group sub-cluster of step 1, an unmanned aerial vehicle group target tracking algorithm is used to extract the motion trajectory of each unmanned aerial vehicle in the cluster, forming an unmanned aerial vehicle group member trajectory cluster. Specifically:

[0072] Step 2.1: For the jth unmanned aerial vehicle sub-cluster Take the sub-cluster member unmanned aerial vehicles as nodes V = {v1, v2,...v n}, the interaction relationship between members as edge E(i,j), and the group consistency size as the weight w(i,j) of the edge, construct the unmanned aerial vehicle group structure graph G = (V,E,w), and use a random matrix to model the extended shape of the unmanned aerial vehicle sub-cluster;

[0073] Step 2.2: For the unmanned aerial vehicle group sub-cluster of step 2.1 Use the Kalman filter algorithm to predict the centroid position and extended shape of the unmanned aerial vehicle group in the next frame, and select the range of N times the size of the extended target near the predicted unmanned aerial vehicle group centroid as the search area;

[0074] Step 2.3: For the unmanned aerial vehicle group structure graph in the search area Use the spectral clustering algorithm to divide the subgraph and obtain the unmanned aerial vehicle sub-group with high internal consistency where P is the total number of sub-groups, and g k is taken as a candidate unmanned aerial vehicle sub-group in the search area;

[0075] Step 2.4: For the candidate unmanned aerial vehicle sub-group g k described in step 2.3, a twin graph convolution network model is constructed to realize unmanned aerial vehicle group member tracking by calculating the similarity of its members with the previous frame unmanned aerial vehicle sub-group. Specifically:

[0076] The twin graph convolution network model includes a template branch and a candidate branch, which respectively input the template unmanned aerial vehicle sub-group structure graph g t and the candidate unmanned aerial vehicle sub-group structure graph g k , wherein the template unmanned aerial vehicle sub-group structure graph is the unmanned aerial vehicle sub-group structure graph of the previous frame, and the graph convolution network with weight sharing is used to extract g t and g kdeep topological features of the UAV sub-group members, to generate an UAV sub-group member embedding vector F t and F k The constructed graph convolution network includes three graph convolution layers, and the information aggregation mode of the graph convolution layer is as follows:

[0077]

[0078] wherein F l represents the member feature vector of the lth layer UAV sub-group; represents the weighted self-loop adjacency matrix of the UAV sub-group structure graph; is the degree matrix of the UAV sub-group structure graph; W l is a feature transformation matrix;

[0079] Based on the UAV sub-group member embedding vector, the group member similarity matrix between the template UAV sub-group and the candidate UAV sub-group is calculated Finally, the nearest neighbor matching algorithm is adopted, and the most similar group members in the two UAV sub-groups are associated according to the group member similarity matrix to form a group member matching pair. Through iterative processing of the time sequence image, the matched node pairs (UAV sub-group members) are connected, and the UAV group sub-cluster member motion trajectory T rack is formed.

[0080] Step 3: For the UAV group member trajectory cluster in step 2, a trajectory similarity model is established to identify the cluster head UAV of each sub-cluster by comparing the differences between the motion trajectories in the cluster. It includes:

[0081] Step 3.1: For the UAV group sub-cluster member motion trajectory T rack , the member flight trajectory segment is intercepted by a sliding time window, the structural feature information of the UAV group sub-cluster member motion trajectory is extracted, and a feature weight vector W = {W D ,W S ,W A ,W L ,W T} is constructed. The elements in the vector correspond to the direction D, speed S, angle A, position L and time T of the trajectory respectively. Each weight value is greater than or equal to zero, and W D +W S +W A +W L +W T = 1. The sensitivity of the trajectory structure feature attribute is adjusted by the feature weight, and the trajectory structure similarity is the difference degree similarity between the trajectory segments. For a trajectory segment sample L, the trajectory structure similarity is calculated by the following formula:

[0082] DIST(L) = D dist × W D +Sdist xW S +A dist xW A +L dist xW L +T dist xW T ;

[0083] Step 3.2: Structural comparison of all segments of different trajectories using structural similarity, DIST(L) represents the structural difference of trajectory segments, the smaller the value, the more similar the trajectory segments, otherwise the less similar, select the trajectory segment with similarity less than threshold T h as the motion trajectory mining segment, which contains obvious differences between the leader and follower trajectories, and the UAV group member with obvious motion trajectory change is selected as the cluster head UAV C drones .

[0084] Step 4: For the cluster head UAV of step 3 sub-clustering, take targeted countermeasures to realize the directional displacement of the UAV group. Specifically:

[0085] For each sub-cluster head UAV C drones , send a false navigation deception signal to make the cluster head UAV deviate from the expected task target position, and the UAVs in the cluster take the actual position of the cluster head UAV as the reference, and do not directly receive the virtual target information, so the UAVs in the cluster will follow the cluster head UAV to deviate from the expected task trajectory, thereby realizing the directional displacement of the UAV group sub-clustering. The sub-clustering navigation deception process is shown in Figure 3 .

[0086] Further illustrated by the following simulation:

[0087] 1. Simulation conditions

[0088] In order to test the effectiveness of the present application, the present application uses a real multi-cluster UAV group image sequence to simulate and verify the UAV group sub-clustering detection algorithm and the UAV group target tracking algorithm. The software framework is Pytorch1.10, and the hardware running platform is a workstation computer containing Xeon E5 processor (3.6GHz), NVIDIA 2080Ti graphics card (11GB) and 32GB running memory.

[0089] 2. Simulation experiment

[0090] The processing object of the UAV group sub-clustering region detection algorithm is a multi-cluster structure UAV group, and the purpose is to extract the distribution region of each sub-cluster of the multi-cluster structure UAV group. The present application uses an attention mechanism based UAV group sub-clustering detection algorithm to realize UAV group sub-clustering region detection, and the detection result is as shown in Figure 4As shown in the figure. It can be seen that the algorithm proposed by the present application can accurately segment each sub-cluster region, the region boundary is clear, and is similar to the true result. Through sub-cluster extraction, it is further used to focus on identifying the cluster head unmanned aerial vehicle of each sub-cluster.

[0091] For each sub-cluster unmanned aerial vehicle group, an unmanned aerial vehicle group target tracking algorithm is used to extract the motion trajectory of the sub-cluster member, and the result is as shown in the figure. Figure 5 Different colored curves represent the historical motion trajectories of different unmanned aerial vehicles. From Figure 5 It can be seen that the motion trajectories of each member in the sub-cluster have high similarity, but the motion trajectory of the cluster head unmanned aerial vehicle is smoother, and the motion of the unmanned aerial vehicle in the cluster has a certain lag. Figure 5 The middle red box represents the identified unmanned aerial vehicle group sub-cluster cluster head unmanned aerial vehicle.

[0092] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for counteracting a multi-clustering structure UAV swarm based on trajectory mining, characterized in that: The method comprises the following steps: Step 1: Obtain a target image sequence of the UAV group, and extract UAV group sub-clusters from the image frame by frame by using a UAV group sub-cluster detection algorithm; Step 2: For the UAV group sub-clusters obtained in step 1, extract the motion trajectories of each UAV in the cluster by using a UAV group target tracking algorithm, and form a UAV group member trajectory cluster; Step 3: For the UAV group member trajectory cluster obtained in step 2, establish a trajectory similarity model, and identify the cluster head UAV of each sub-cluster by comparing the differences between the motion trajectories in the cluster; The step 3 comprises: Step 3.1: UAV swarm sub-cluster member motion trajectory , the structural feature information of the UAV swarm sub-cluster member motion trajectory is extracted by intercepting the member flight trajectory segment through a sliding time window, and a feature weight vector is constructed , the elements in the vector correspond to the direction D, speed S, turning angle A, position L and time T of the trajectory respectively, each weight value is greater than or equal to zero, and , the sensitivity of the trajectory structure feature attribute is adjusted through the feature weight, the trajectory structure similarity is the difference degree similarity between the trajectory segments in structure, and for a trajectory segment sample L, the trajectory structure similarity is calculated by the following formula: ； Step 3.2: Structural comparison of all segments of different trajectories by structural similarity, reflects the structural difference of trajectory segments, the smaller the value, the more similar the trajectory segments, and vice versa, select the trajectory segment with similarity less than threshold T h as the motion trajectory mining segment, the motion trajectory contains obvious differences between the leader and the follower, and the UAV group member with obvious motion trajectory change is selected as the cluster head UAV of sub-clustering ; Step 4: For the cluster head UAV of the sub-cluster obtained in step 3, take targeted countermeasures to realize the directional repulsion of the UAV group. 2.The multi-clustering structure UAV swarm countermeasure method based on trajectory mining of claim 1, wherein: The step 1 comprises: Step 1.1: Acquire aerial drone swarm image sequence using ground optical imaging device , denotes total number of frames; Step 1.2: The UAV group target image sequence of step 1.1 , using the UAV group sub-cluster detection algorithm based on attention mechanism, frame by frame calculating the UAV group sub-cluster distribution prediction map ; Step 1.3: Generating sub-cluster distribution probability map for the sub-clusters generated in step 1.2 using threshold segmentation algorithm extracting sub-clusters of UAVs , denotes the total number of sub-clusters.

3. The multi-cluster structure UAV group countermeasure method based on motion trajectory mining according to claim 2, characterized in that: The attention mechanism-based UAV swarm sub-cluster detection algorithm of step 1.2 adopts an improved U-Net structure, including an encoder, an enhancer and a decoder, and inputs the i-th frame of UAV swarm image , and outputs the i-th frame of UAV swarm sub-cluster distribution probability map through depth coding, feature enhancement and information decoding , and the calculation process of the algorithm is as follows: ； wherein, is an encoder function for encoding the UAV swarm image depth features; is a decoder function for decoding the UAV swarm image depth features to generate a UAV swarm sub-cluster distribution prediction map; is an enhancer function for suppressing high-frequency signal interference of the UAV swarm image depth features and enhancing the expression ability of task-related features.

4. The multi-clustering structure unmanned aerial vehicle swarm countermeasure method based on trajectory mining of claim 3, wherein: The step 2 comprises: Step 2.1: Sub-cluster for the jth UAV , with sub-cluster member UAVs as nodes , the interaction between members as edges , the group consensus size as the weight of the edge , construct the UAV group structure diagram , and use a random matrix to model the extended shape of the UAV sub-cluster; Step 2.2: Sub-clustering of the UAV group for step 2.1 The Kalman filtering algorithm is used to predict the centroid position and the extended shape of the UAV group in the next frame, and the search area is selected as the range of N times the size of the extended target around the predicted centroid of the UAV group. Step 2.3: Structure diagram of UAV group in search area , using spectral clustering algorithm to divide subgraph, get high consistency of UAV subgroup , wherein, is the total number of subgroups, and is taken as a candidate UAV subgroup in the search area; Step 2.4: For the candidate sub-swarm of unmanned aerial vehicles described in step 2.3 , a twin graph convolution network model is constructed to realize unmanned aerial vehicle swarm member tracking by calculating the similarity of its members with the previous frame of unmanned aerial vehicle sub-swarm.

5. The multi-clustering structure unmanned aerial vehicle swarm countermeasure method based on trajectory mining of claim 4, wherein: The step 2.4 is specifically: The twin graph convolution network model comprises a template branch and a candidate branch, and respectively inputs a template UAV subgroup structure graph and a candidate UAV subgroup structure graph , wherein the template UAV subgroup structure graph is a UAV subgroup structure graph of a previous frame, and deep topological features of and are respectively extracted through a weight-shared graph convolution network to generate UAV subgroup member embedding vectors and , and the constructed graph convolution network comprises three graph convolution layers, wherein the information aggregation mode of the graph convolution layers is: ； wherein, represents the member feature vector of the lth layer UAV subgroup; represents the weighted self-loop adjacency matrix of the UAV subgroup structure graph; is the degree matrix of the UAV subgroup structure graph; is the feature transformation matrix; Based on the UAV subgroup member embedding vectors, a group member similarity matrix between the template UAV subgroup and the candidate UAV subgroup is calculated Finally, a nearest neighbor matching algorithm is used to associate the most similar group members in the two UAV subgroups according to the group member similarity matrix to form a group member matching pair, and through iterative processing of the time sequence images, the matched node pairs are continuously connected to form the UAV subgroup cluster member motion trajectory .

6. The multi-clustering structure UAV swarm countermeasure method based on trajectory mining of claim 5, wherein: The step 4 is specifically: Cluster head unmanned aerial vehicle for each sub-cluster The false navigation decoy signal is sent, so that the cluster head unmanned aerial vehicle deviates from the expected task target position. The unmanned aerial vehicles in the cluster take the actual position of the cluster head unmanned aerial vehicle as a reference and do not directly receive virtual target information. Therefore, the unmanned aerial vehicles in the cluster will deviate from the expected task trajectory following the cluster head unmanned aerial vehicle, so as to realize the sub-cluster directional driving of the unmanned aerial vehicle group.

Images