A Framework for Fast Online Object Tracking and

In this paper, we address the problem of online informative feature selection for a class of tracking techniques called “tracking by detection” which has been shown to give promising results at real-time speed. In tracking by detection methods, an online discriminative classifier is trained to separate the target object from the background. The classifier is incrementally updated using positive and negative samples from the current frame. How to select the most informative features to update the classifier is very important in order to avoid the drift problem. We propose a feature selection approach by minimizing the information entropy which is able to select more informative features than most state-of-the-art tracking algorithms. Experimental results on challenging sequences demonstrate that the proposed tracking framework is robust, effective and accurate.