Online Feature Selection for Visual Tracking

Object tracking is one of the most important tasks in many applications of computer vision. Many tracking methods use a fixed set of features ignoring that appearance of a target object may change drastically due to intrinsic and extrinsic factors. The ability to dynamically identify discriminative features would help in handling the appearance variability by improving tracking performance. The contribution of this work is threefold. Firstly, this paper presents a collection of several modern feature selection approaches selected among filter, embedded, and wrapper methods. Secondly, we provide extensive tests regarding the classification task intended to explore the strengths and weaknesses of the proposed methods with the goal to identify the right candidates for online tracking. Finally, we show how feature selection mechanisms can be successfully employed for ranking the features used by a tracking system, maintaining high frame rates. In particular, feature selection mounted on the Adaptive Color Tracking (ACT) system operates at over 110 FPS. This work demonstrates the importance of feature selection in online and realtime applications, resulted in what is clearly a very impressive performance, our solutions improve by 3% up to 7% the baseline ACT while providing superior results compared to 29 state-of-the-art tracking methods.