Research On Vision-based Target Detection And Tracking Algorithm

In recent years,computer vision technology is entering a period of rapid development,in artificial intelligence,security monitoring,intelligent city,unmanned vehicles and many other fields have wide application and unlimited development prospects.As one of the important contents of computer vision,the realization of online detection and tracking of unknown targets has gradually become a hot topic in the research of experts,scholars and enterprises.Based on visual target detection and tracking,it mainly uses computer vision technology to detect,analyze and process the object to be tracked in the input video image,obtain the position information of the moving target,predict and track the trajectory of the target,so as to realize the goal of real-time on-line tracking of the moving object.At present,the main factors that affect the target detection and tracking are the illumination changes,target occlusion,target deformation and size change,and the disappearance and recurrence of targets.In view of this kind of problem,based on the analysis and summary of the development course and present situation of target detection and tracking technology,this paper puts forward an algorithm combining detection,tracking and online learning to complete the long time online tracking of unknown target.The main research contents are as follows:(1)In this paper,several common target detection and tracking algorithms are analyzed in detail.In the research of target detection algorithm,the advantages and disadvantages of the background difference method,the frame difference method and the optical flow method are compared and analyzed,and the calculation method and the realization process of the pyramid Kanade optical flow tracking method are emphatically introduced.In the research of target tracking algorithm,the Meanshift tracking algorithm is studied in detail,the modeling process of Meanshift is confirmed,and the practical effect of meanshift algorithm in realizing target tracking is verified by experiment,which shows that the algorithm has good real-time performance,occlusion of moving target,In the case of attitude change,it has strong stability and robustness.(2)The traditional method of tracking the moving target only depends on the detector or tracker,which is often not able to be detected and tracked after the target is temporarily lost.In order to solve this problem,this paper introduces the n online learning device,and continuously trains new samples through on-line learning,so as to constantly update the detector and tracker,and finally realizes the real-time on-line tracking of unknown moving target.Experiments show that thealgorithm can solve the problem of tracking target after disappearing,but the tracking effect needs to be improved in the complex situation of target occlusion and target deformation.(3)Because the original TLD algorithm has a large computational load,especially in the case of the moving target occlusion and attitude change,considering the advantages of Meanshift algorithm,this paper proposes the idea of Meanshift to improve the original TLD algorithm.Finally,through the comparison of the two groups of experiments,it is shown that the improved algorithm can not only realize the on-line tracking of unknown target,but also has a good tracking effect in other disturbances.