Geometric Features Recognition Of UAV Swarm Targets

In the UAV swarm(UAS)confrontation,the detection technology is facing many challenges.First of all,conventional radars are difficult to achieve effective detection for low and slow targets such as drones.Also,it is difficult to accurately match the traj ectories when crossing,exiting,and entering scenes between targets.In addition,how to define the features of the swarm from a macro perspective and achieve the organic unification of detection and countermeasures is also a major difficulty.This thesis analyzes the above difficulties and challenges,designs an anti-UAS detec-tion system,and achieves real-time accurate detection of UAS targets under the condition of limited computing power,and further identifies the macro features of the UAS.Based on the outstanding performance of deep learning in the field of computer vision,the topic designs and optimizes multi-target real-time detection and tracking algorithms,and de-fine macro geometric features of UAS targets,moreover,designs algorithms for swarm feature recognition.In summary,in addition to the basic theories of feature engineering,deep learning,and target detection and tracking technologies,the main research content of the subject includes:1.Aiming at the contradiction between the computing power and accuracy when de-tecting small targets,based on the YOLO algorithm,a deep learning technology combining FPN and the residual network is proposed,with the technical route of hierarchically detecting of large,medium and small targets,as well as features fu-sion between high and low levels.The proposed algorithm effectively improves the problem of ineffective detection of small targets and guarantees a detection speed that meets real-time detection simultaneously.2.Aiming at the problem of low accuracy of trajectory matching in MOT tasks,based on the Kalman filter algorithm,a trajectory matching algorithm combining cosine distance is proposed,which combines the target motion features and the apparent features.The proposed algorithm could solve the problem of trajectory matching when the target generates occlusion,exit,and enter the scene.3.Under high and low image brightness dynamic range,the target is easily confused with the background,which makes it difficult to separate the foreground and back-ground of the image,and affects the accuracy of detection and tracking.In order to solve this problem,this thesis has studied image enhancement and tone mapping technology to design an image enhancement algorithm based on biological retinas.The adaptive convolution method is used to effectively improve the brightness level distribution of the image,leading an effectively improving the detection accuracy.4.Based on the above research,the definition of macro swarm features combined with conventional countermeasures at the current stage is proposed,and a three-dimensional reconstruction algorithm of cluster targets based on binocular vision is designed.Based on this,the geometric features of the cluster are extracted,further,the UAS target geometric feature recognition technology that meets the require-ments of anti-UAV cluster tasks is implemented,and the problem of macro feature description of the UAS targets is overcome.This thesis designed a UAS detection system combining traditional algorithms and deep learning,also performed simulation experiments on a large number of data sets to verify the theory.It has the functions of detecting small swarms of targets in real-time and describing swarm features.It can be used in the detection phase of anti-UAV cluster technology to solve the problem of identification based on swarm geometry.