A brief review on visual tracking methods

Long-term robust visual tracking is still a challenge, primarily due to the appearance changes of the scene and target. In this paper, we briefly review the recent progress in image representation, appearance model and motion model for building a general tracking system. The models reviewed here are basic enough to be applicable for tracking either single target or multiple targets. Special attention has been paid to the on-line adaptation of appearance model, a hot topic in the recent. Its key techniques have been discussed, such as classifier issue, on-line manner, sample selection and drifting problem. We notice that the recent state-of-the-art performances are generally given by a class of on-line boosting methods or `tracking-by-detection' methods (e.g. OnlineBoost, SemiBoost, MIL-Track, TLD, etc.). Therefore, we validate them together with typical traditional methods (e.g. template matching, Mean Shift, optical flow, particle filter, FragTrack) on a challenging sequence for single person tracking. Qualitative comparison results are presented.