Robust object tracking based on sparse representation

Abstract

ABSTRACT In this paper, we propose a novel and robust object tracking algorithm based on sparse representation. Objecttracking is formulated as a object recognition problem rather than a traditional search problem. All targetcandidates are considered as training samples and the target template is represented as a linear combination ofall training samples. The combination coecients are obtained by solving for the minimum l 1 -norm solution.The “nal tracking result is the target candidate associat ed with the non-zero coecie nt. Experimental resultson two challenging test sequences show that the proposed method is more eective than the widely used meanshift tracker.Keywords: Object tracking, sparse representation, l 1 Š norm minimization, 1. INTRODUCTION Object tracking in the computer vision community usually refers to the eorts of consistently “nding the motionstate of the tracked target in consecutive video frames. In other words, given a target template in advance, thegoal of object tracking is to obtain the position and size of the target in current frame by checking all targetcandidates and “nd the one which is the most similar with the target template. It is used in many practicalapplications, such as automated surveillance, video analysis, human computer interfaces and vehicle navigationand so on.In the previous literature, numerous tracking algorithms have been proposed and they usually fall into twomajor groups: deterministic methods and stochastic methods. In deterministic methods, the target object islocated in current frame by maximizing the similarity between the target model and a target candidate.