Research On Target Tracking Technology Based On Correlation Filtering And Siamese Network

Nowadays,with the rapid development of artificial intelligence,computer technology and other technologies.Target tracking,as the middle-level foundation of computer vision applications,has been utilized in various advanced tasks,such as: intelligent transportation,human-computer interaction,intelligent video surveillance,and modern military and so on.The basic principle of target tracking is to predict the position and the shape of the target in the subsequent frame and the shape of the target according to the given target in the initial frame of the video sequence,which will lay the foundation for further analysis of the target's motion behavior and the realization of more advanced visual tasks.The target tracking technology has been paid close attention to because of its basic and challenging visual applications.The basic is for that the target tracking technology is the basis of many high-level applications related to computer vision;and the challenge is for that although domestic and foreign scholars has made a great contribution in target tracking,but during the target tracking process,the variation of target and the complex tracking environment including deformation,scale variation and so on,which leads to the limitation of the existing target tracking algorithms.That means the target tracking algorithm still needs to be further improved.In order to improve the tracking accuracy in the complex environment of target tracking,especially occlusion,fast motion,deformation and rotation in the plane.This paper has studied correlation filter algorithm and correlation filter siamese network,and did some improvement.The main work shows as following.To against the traditional target tracking algorithm cannot deal with the problems of target occlusion,fast motion and in-plane rotation.A failure discrimination mechanism based on optical flow is proposed on this paper.Then,when obtaining the tracking failure result of the discrimination mechanism,this paper adopts the strategy of patch training to modify the filter template update strategy for the improvement on tracking accuracy.Experiments show that the improved algorithm based on correlation filtering can achieve target tracking more robustly than the other five filter algorithms in the three complex tracking cases of target occlusion,fast motion and in-plane rotation;At the same time,the tracking effect can be improved on the entire OTB100 data set.In addition,to against the problem like as occlusion of target,fast motion and the deformation in the tracking process and the siamese network template branch doesn't updateto adapt these conditions.Correlation filtering siamese network based on IOU target tracking algorithm is proposed in this paper.Firstly,an online learning method based on IOU is proposed in this algorithm,which can determine whether the target drifts or is lost during the tracking process by calculating the IOU value between prediction result in the current frame and frame tracking result in the previous during the tracking process;second,on the basis of IOU online learning,a adaptive update strategy is adopted for template branches,and template branches are adaptively updated according to the value of IOU.Experiments show that the proposed algorithm in this paper can achieve the goal more robustly than the other five algorithms when facing complex tracking problems such as target occlusion,fast motion and target deformation.And the overall tracking performance has been improved on the two classic data sets of OTB100 and OTB50.