Dictionary Learning for Visual Tracking with Dimensionality Reduction

Abstract

Recently, visual tracking has seen much progress in either accuracy or speed. However, due to drastic illumination variation, partial occlusion, scale variation and out-of-plane rotation, visual tracking remains a challenging task. Dealing with complicated appearance variations is an open issue in visual tracking. Existing trackers represent target candidates by a combination of target templates or previous tracking results under some constraints. When a drastic appearance variation occurs or some appearance variations occur simultaneously, such target representations are not robust. In this paper, we present a discriminative dictionary learning based target representation. A target candidate is represented via a linear combination of atoms in a learnt dictionary. The online dictionary learning can learn the appearance variations in tracking processing. So, the learnt dictionary can cover all of kinds of appearance variations. Based on this kind of target representation, a novel tracking algorithm is proposed. Extensive experiments on challenging sequences in popular tracking benchmark demonstrate competing tracking performances against some state-of-the-art trackers.