Research On Vision-based Detection And Tracking Technology For Small UAVs

The advancement of science and technology has driven the development of UAV industry.Nowadays,UAVs are widely used in all walks of life,but the "dark flying" and "spread flying" phenomenon of drones have greatly increased the difficulty of the drone aerial supervision.Moreover,It has seriously endangered personal safety,social security and national security.Therefore,the research on anti-UAV technology has become extremely important.UAV detection and tracking technology is an important part of the anti-UAV system.In view of the characteristics of UAV,such as small size,variable flight height and complex flight environment,this paper focuses on the study of vision-based UAV detection and tracking technology.Firstly,since there are no open UAV data sets at present,we have taken many kinds of UAV videos with different flying altitudes and attitudes,and made 4000 UAV sample sets.Then,this paper introduces the traditional detection algorithm,the candidate region-based detection algorithm and the end-to-end detection algorithm represented by YOLO,and improves and optimizes YOLOv3 network model,including data enhancement,bounding box improvement,the selection and fusion of multi-scale features and loss function parameter design.Finally,we analyze and compare the algorithm before and after the improvement,so as to verify that the improved network framework can detect the small UAV more accurately and in real-time.Secondly,in this paper,the YOLOv3 detection result is used as the initialization input of the tracking algorithm to realize automatic tracking.After comparing and analyzing the experimental results of particle filter and TLD tracking algorithm,this paper selects KCF algorithm which has strong real-time performance for UAV tracking.To solve the problem of size changes,multi-scale calculation is added.And for the problem of fast and large-scale shaking of the target,Kalman filter is added to predict and correct the position of the drone.At the same time,for the target missing situation in the tracking process,a confidence test scheme is proposed,and the tracking model is initialized by the detection algorithm again.The experimental results show that the optimized tracking algorithm has a better tracking effect.