A Kernel Sparse Representation Based Visual Tracking

Abstract

Recently, the sparse representation based visual tracking is very popular, which is robust to occlusion and noise, but not satisfactory in the scenarios of fast motion and blur.In order to solve this problem, a new tracking method based on kernel sparse representation is proposed.In this method, each candidate is represented by kernel sparse representation, then the computed reconstruction error is used to obtain the observation probability, and at last the candidate with the maximal observation probability is determined as the target.Aspect to the solving of kernel sparse representation, the accelerated proximal gradient (APG) is adopted.Experiments on several representative image sequences shows that the proposed tracking method performs better than the sparse representation based visual tracking method in the fast motion and blur scenarios.