Robust visual tracking via principal component pursuit

The traditional subspaces based visual trackers are prone to model drifting. To deal with this problem, we propose a robust visual tracking method based on principal component pursuit. The proposed method represents objects with subspaces spanned by multiple templates, and finds error components of target candidates via principal component pursuit. The optimal state parameters are estimated by the error components of object candidates in particle filter framework. To adapt to changes of object appearance and avoid model drifting, a template update method is proposed. The proposed method updates the template set using tracking result when the tracking result is very similar to the templates; otherwise, it updates the template library with low-rank component corresponding to the tracking result. The experimental results on several challenging sequences show that the proposed method has better performance than that of the state-of-the-art tracker. ©, 2015, Chinese Institute of Electronics. All right reserved.