Design Of Autonomous Tracking System For UAV Targets Based On Machine Vision

Along with the development of unmanned aerial vehicle technology ushered in a vigorous period,there are more and more applications of UAV.Using UAV to track the target has become one of the research focuses in the field of machine vision.Whether in the military or civilian fields,the tracking function of UAV can be used to conduct reconnaissance or aerial photography of targets.After years of research and development,the target tracking algorithm has made major breakthroughs in the speed and accuracy of the algorithm.However,in practical applications,the target tracking is often inaccurate or even lost due to factors such as target deformation,illumination changes,motion blur,occlusion,and scale changes.Aiming at the problem of UAV tracking moving targets,an autonomous UAV target tracking system based on machine vision is designed in this paper.The main contents and solutions are as follows:(1)According to the requirements of target tracking using UAV,the flight principle and tracking principle of UAV are mainly introduced in detail.Introduce the flying principle of the UAV by analyzing the five flight motion states of hovering,lifting,pitching,rolling and yaw.Introducing the target tracking principle of the UAV by analyzing the conversion process from the two-dimensional image coordinate system to the three-dimensional UAV coordinate system.(2)Aiming at the change of target scale,this paper introduces the scale filter of DSST(Discriminative Scale Space Tracking)algorithm into KCF(Kernelized Correlation Filters)algorithm,which makes the target frame change with the change of target scale in the tracking process.Therefore,KCF-SA(Scale adaptation)algorithm with target scale adaptation function is formed.After quantitative experiments,compared with KCF algorithm,KCF-SA algorithm reduces the average center error by 2 pixels and increases the average overlap rate by 19.8%.(3)Aiming at the occlusion and fast motion of the target,this paper introduces the confidence estimation of the target result into the KCF-SA algorithm to judge the target missing state.The first frame redetection mechanism to reduce the offset of the target frame,and the extended region redetection mechanism to improve the tracking robustness when the target moves fast,which forms the KCF-RD(Re-detection)algorithm.After quantitative experiments,compared with KCF-SA algorithm,KCF-RD algorithm reduces the average center error by 39 pixels and increases the average overlap rate by 32.3%.(4)Aiming at the change of the rotation angle of the target,Fourier-Mellin Transform is introduced into KCF-RD algorithm to realize the target angle adaptation when the target rotates.In the algorithm,two filters with learning rate are used to calculate the confidence level respectively,and the angle with high confidence level is selected for updating to improve the accuracy of the algorithm which forms the improved algorithm.After quantitative experiments,the improved algorithm reduces the average center error by 6 pixels and increases the average overlap rate by 19.4% compared with KCF-SA algorithm.