Research On Long-Term Object Tracking Based On Complementary Trackers

Object tracking based on correlation filter is one of the most popular problems in computer vision,has a broad developmental prospect and important applications in video surveillance,robot,autonomous vehicle navigation and intelligent transportation.Limited by the inherent boundary effects and high risk update strategy of correlation filters and stability-plasticity dilemma,unitary correlation filter could not satisfy realworld applications.Therefore,this paper focus on the research that merge different complementary trackers to overcome the drawbacks of correlation filters and realize accuracy tracking at complicated environments.Standard correlation filter employ circulant samples to train ridge regression classifier,which lead to inherent boundary effects.Focusing on the boundary effects,this paper utilize the merits of color histogram to construct a Bayes classifier to track targets that encounter low resolution and fast deformation.Meanwhile,Bayes classifier is employed to alleviate the adverse impact of boundary effects by ensemble learning.Later,Bayes classifier based on local sensitive histogram is proposed to strengthen the discriminative ability on the gray video sequences that lacking of color information.Finally,high-confidence update strategy is proposed to prevent Bayes classifier and correlation filter form updating with corrupted samples.The qualitative and quantitative experimental results show that the proposed methods are obtain different progress.Existing correlation filter and Bayes classifier adapt the change of target quickly by online learning,the high risk update strategy and the stability-plasticity dilemma of online learning make tracking model drift once target encounter heavy occlusion or outof-view.We propose an adaptive learning rate to deal with the high risk update strategy of classifier,the discriminative ability of classifiers are strengthened by increasing the weights of high quality samples and decreasing the weights of corrupted samples.Aiming at the stability-plasticity dilemma,we propose a long-term tracker by combining complimentary trackers with soft margin support vector machine to re-detect target after tracking failures.The proposed long-term tracker is proved robust when target encounter fast motion,out-of-view and heavy occlusion.