Robust visual tracking via binocular multi-task multi-view joint sparse representation

摘要

Visual object tracking has been a major and fundamental topic in computer vision field for decades. Despite of great progress that has been made, robust tracking under drastic illumination changes, continuous occlusions and scale changes remains a very challenging work. In this paper, an efficient binocular object tracker via joint sparse representation is proposed as Binocular Multi-task Multi-view Tracker (BMTMVT). By introducing the unit-norm normalized 3D depth feature into previous multiple 2D views (such as intensity, color, texture and edge) based sparse representation framework, tracking performance is significantly improved. Meanwhile, an approach for occlusion detection utilizing depth based histogram analysis is further proposed to efficiently decide the accurate time to update target template set. Besides, a strategy of particles pretreatment and a screening process are introduced to enhance the particles efficiency and to optimize tracking performance with employing range data respectively. Extensive experiments on various types of challenging sequences from KITTI and Princeton data sets demonstrate that the proposed BMTMVT algorithm outperforms the state-of-the-art trackers, especially when handling frequently changing illuminations, successive obstructions and variations in scale.