Contents I. Introduction
Many powerful algorithms have been proposed for object tracking. They can be classified into four classes: Region-based methods[1] [2] [3], Feature-based methods[4] [5] [6] [7], Deform-able-template-based methods[8] [9][10] [11] and Model-based methods[12] [13] [14]. However, most of these traditional tracking approaches depend on some expensive assumptions. For example, they assume that motion and appearance are continuous and that the representative fixed features can always distinguish the interested objects from background well. The representative fixed features are selected before the tracking task starts. Unfortunately, motion and appearance continuity is often not satisfied because of an abrupt motion of the object or the camera. An object detector trained offline can be integrated into the tracker to solve this problem. Moreover, the fixed features cannot always distinguish the objects from the background well. There are two reasons for this: the object appearance will change when the illumination changes, occlusion happens or viewpoint varies; and the background will change as the target object moves from place to place. The remedy for the drawback of fixed features is using online selection of discriminative features for object tracking. For example, Collins et al[15] proposed a method in which a feature evaluation mechanism is embedded in a mean-shift tracking system that adaptively selects the top-ranked discriminative features for tracking. Jianyu Wang et al[16] online selected discriminative features from a set of Haar features into the appearance model for tracking. However, the problem of an online learning system is obvious: setting the online features updating ratio may be very difficult because the over updating may even ruin the original model. It can be solved by combining offline learning and online learning. For example, Yuan Li et al[17] proposed a cascade particle filter with discriminative observers of different Lifespans. The features selected online can represent object appearance more specifically, while the features selected offline can produce more accurate result.
