A Robust Appearance Model for Object Tracking

Abstract

Patch strategy is widely adopted in visual tracking to address partial occlusions. However, most patch-based tracking methods either assume all patches sharing the same importance or exploit simple prior for computing the importance of each patch, which may depress the tracking performance when the target object is non-rigid or the background information is included in the initial bounding box. To this end, an importance-aware appearance model with respect to the target patches and background patches is built, which is able to adaptively evaluate the importance of each target/background patch by means of the local self-similarity. In addition, we propose a novel bi-directional multi-voting scheme, which integrates a multi-voting scheme and the two-side agreement scheme, to produce a reliable target-background confidence map. Combining the importance-aware appearance model and the bi-directional multi-voting scheme, a robust patch-based tracking method is proposed. Experimental results demonstrate that the proposed tracking method outperforms other state-of-the-art methods on a set of challenging tracking tasks.