Research And System Implementation Of UAV Tracking Method Based On Deep Learning

In recent years,the UAV industry has developed rapidly and has been widely used in social production.However,due to the characteristics of low altitude,slow speed and small targets,UAVs are difficult to control,and illegal " black flight" incidents that endanger the safety of airspace occur from time to time.How to effectively solve the problem of "black flight" of UAVs has become the fundamental motivation and purpose of this subject.The solution of the "black flight" problem is mainly composed of detection and control.After the UAV target is detected first,it can be decoyed and controlled.This paper mainly improves the target tracking algorithm based on correlation filter,studies and implements the UAV visual tracking system based on deep learning.After the UAV target is detected in the image,it will be tracked.From two perspectives of theory and engineering,the main research contents are as follows:1.This paper studies the basic theory of target detection and tracking,the target tracking algorithm based on correlation filter and the target detection algorithm based on deep learning are introduced respectively,and the classic algorithms in the field of target detection and tracking are analyzed and compared.2.In terms of theoretical research,this paper uses a variety of complementary features to describe the appearance of the target in view of the single feature and poor robustness of traditional correlation filter algorithms.Considering the influence of the boundary effects,this paper introduces the context-aware framework to extract one layer of convolution features of four image blocks around the target as background information.In the stage of feature response fusion,a new multi-feature fusion strategy is adopted to effectively combine the deep and shallow complementary features.Finally,the average peak correlation energy(APCE)index is used to evaluate the confidence of the response map and the reliability of the tracking results,and decide whether to update the model.The algorithm proposed in this paper is tested on the OTB-2013 benchmark,and the results show that compared with the HCF algorithm,the tracking accuracy of our algorithm is increased by 0.3%,and the tracking success rate is increased by 1.9%.3.This paper transplants and optimizes the YOLO v4 target detection algorithm in Qt IDE,uses Hikvision camera to capture video stream in real-time and decodes it.We use the producer/consumer multi-threaded design mode to improve the speed of the program,and the link list queue is used as the buffer,the GPU accelerates the detection algorithm.Finally,we use PID algorithm to control the steering gear of the camera.The system has been tested in outdoor environment,and the results show that our system can detect and track UAV targets.