| With the development of new military combat technologies such as intelligence and unmanned operations,drone swarm combat technology has become a hot issue in various countries.Due to the low cost of drone swarms,we usually spend hundreds of times the cost to attack and defend them.If the formation formed by a swarm of drones is known,the key nodes can be found through the prior knowledge of the formation type,which greatly improves the efficiency of strikes.Therefore,identifying drone swarm targets and giving their formation category is of great significance in the field of drone defense.In order to detect the weak and small targets of the drone swarm movement in the infrared sequence diagram,a target detection algorithm for the drone swarm based on foreground segmentation is proposed.This algorithm aims at some background pixels being mistaken when the background changes greatly.Judged as the problem of the former scenic spot,an adaptive threshold generation method based on background dynamic factor is proposed,which adaptively generates a dynamic segmentation threshold according to the degree of background change to ensure the accuracy of foreground segmentation.Aiming at the problem that the foreground target is repeatedly detected as multiple targets caused by the ghost phenomenon,a ghost elimination method based on the judgment of the ghost image category is proposed to generate a real background model and avoid the ghost phenomenon.After obtaining the binarized image of the foreground and background,morphological processing and connected domain search methods are used to generate the smallest circumscribed frame of each UAV target to complete the target detection task.In order to identify the drone swarm formation,this paper introduces the graph neural network into the field of drone swarm formation recognition.First,the detected target coordinate points are converted into the input data of the graph neural network,and then a method based on intra-layer summation is designed.The graph convolutional network architecture for feature extraction between layers can output formation types composed of different drone swarms after training.The graph convolutional network architecture can perform feature extraction and feature information aggregation at the node level,and learn at the hierarchical level,so that the final output result can better reflect the feature information of the graph data itself.Finally,a comparative experiment was conducted on the UAV swarm infrared moving target data set and the image classification data set.The final results show that the foreground segmentation algorithm based on the adaptive background model proposed in this paper has better performance and has certain advantages compared with other methods.With good robustness,the graph convolutional network based on intra-layer and inter-layer feature extraction proposed in this paper is superior to other methods in recognition accuracy,and has a strong ability to extract deep features. |
