Discriminative nonorthogonal binary subspace tracking

摘要: Visual tracking is one of the central problems in computer vision. A crucial problem how to represent object. Traditional appearance-based trackers are using increasingly more complex features order be robust. However, representations typically will not only require computation for feature extraction, but also make state inference complicated. In this paper, we show that with a careful selection scheme, extremely simple yet discriminative can used robust object tracking. The component proposed method succinct and representation image template non-orthogonal binary subspace spanned by Haar-like features. These bases selected from over-complete dictionary variation OOMP (optimized orthogonal matching pursuit). Such inherits merits original NBS it efficiently describe It incorporates information distinguish foreground background. We apply through SSD-based matching. An update scheme devised accommodate appearance changes. validate effectiveness our extensive experiments on challenging videos demonstrate its capability track objects clutter moving