Salient feature selection for visual tracking

Abstract

Proposed is a novel method that can adaptively extract discriminative features and learn the target region accurately for object tracking. Only the region selected as salient pixels by the proposed weighted log likelihood ratio is employed, instead of using all data in the tracker window, for learning the object appearance accurately. The selected pixels are used to train a new weighted likelihood ratio which is employed to select new salient pixels. The proposed method has a recursive structure between selecting salient pixels and learning the weighted likelihood ratio. Experimental results show that the approach by the proposed adaptive feature selection is effective to adapt to object appearance change and alleviate tracking drift or the occlusion problem.