Blur-Resilient Tracking Using Group Sparsity

Pengpeng Liang, Yi Wu, Xue Mei, Jingyi Yu, Erik Blasch, Danil V. Prokhorov, Chunyuan Liao, Haitao Lang, Haibin Ling

Abstract: In this paper, a Blur Resilient target Tracking algorithm (BReT) is developed by modeling target appearance with a groupwise sparse approximation over a template set. Since blur templates of different directions are added to the template set to accommodate motion blur, there is a natural group structure among the templates. In order to enforce the solution of the sparse approximation problem to have group structure, we employ the mixed $$\ell _1+\ell _1/\ell _2$$ norm to regularize the model coefficients. Having observed the similarity of gradient distributions in the blur templates of the same direction, we further boost the tracking robustness by including gradient histograms in the appearance model. Then, we use an accelerated proximal gradient scheme to develop an efficient algorithm for the non-smooth optimization resulted from the representation. After that, blur estimation is performed by investigating the energy of the coefficients, and when the estimated target can be well approximated by the normal templates, we dynamically update the template set to reduce the drifting problem. Experimental results show that the proposed BReT algorithm outperforms state-of-the-art trackers on blurred sequences.