Adaptive structured sub-blocks tracking

Abstract

Local features have been widely used in visual object tracking for their robustness in illumination, deformation, rotation and partial occlusion. Traditional feature selection algorithms based on accumulated knowledge of previous frames usually adopt the perspective of continuity of changes, which could lead to degradation. Exploiting discrimination and uniqueness of local sub-blocks, we build an automatic preselection mechanism for local features and propose the structured sub-blocks tracking algorithm under particle filter framework. Optimal sub-blocks are chosen automatically according to their discriminant function distribution in current frame. Furthermore, we reduce blocks search costs with help of historical prediction accuracy. Experiments validate the robustness of our algorithm in tackling with small deformation and partial occlusion.