Visual Object Tracking Based On Correlation Filters And Online Learning

In recent years,with the development of computer technology and the increasing demand of video analysis,visual object tracking has become one of hot topics of compute vision.In video or image sequences,the aim of object tracking is to estimate the trajectory of target in each image frame with the initialized state(e.g.,position and scale)and using a graph to label the states of target in the subsequent frames.Object tracking has wide applications,such as motion-based recognition,unmanned aerial vehicle(UAV)surveillance,human-computer interaction,autonomous driving,and other photogrammetry and computer vision fields.However,due to some reasons,accurate and robust object tracking is still a huge challenge.Firstly,complex scenarios increase the challenges of object tracking.The target will suffer from many challenging factors in the process of object tracking,such as occlusion,deformation,scale or illumination variation,background clutters,motion blur or fast,low resolution of ground-truth,etc.It is difficult to develop an object tracking algorithm which can overcome all of these challenging factors.Secondly,the prior knowledge for object tracking is limited.The only prior knowledge for object tracking is the position and scale of target in the initial image frame.During object tracking,tracking drifts may occur due to the effect of challenging factors and the accumulation of tracking errors.Therefore,it is an urgent problem that how to utilize limited prior knowledge;overcome the challenging factors to develop an accurate and robust tracking algorithm.Aiming at these aforementioned problems,this paper propose an object tracking framework by combining correlation filters and online learning.Two contributions of this paper are as following.Firstly,object tracking methods based on multi-features.In this paper,a kind of local area dense sampling method is used to extract the multi-feature of the search area to training the discriminative correlation filters for target tracking.Secondly,object tracking based on correlation filters and online learning.Tracking drifts and failure may occur under some challenging factors,when the targets are tracked by methods based on discriminative correlation filters.In this paper,a threshold is used to judge the failure of object tracking.An online learning method is proposed to detect the location of target in the case of tracking failure.In this paper,we evaluate the performance of the proposed algorithm in comparison with state-of-the art trackers using some videos from Online Tracking Benchmark(OTB)dataset.The experimental results show that the proposed method is superior to the state-of-the-art approaches in some challenging factors,such as occlusion.