A Robust Framework for Visual Object Tracking

Visual object tracking is an important problem in computer vision. Recent proposed tracking methods based on online learning a discriminative classifier have drawn considerable interest. However, most of existing approaches make a simple assumption about initializing the object to be tracked; on-line adaptation of binary classifier only has to discriminate the current object from its surrounding background can lead to tracking failure (drifting) without a recovering. This paper presents a novel framework for robust object tracking. The system comprises of a strong learned object detector incorporation with an online adaptation tracking mechanism. The main contributions are: (1) an efficient visual object learning algorithm based on online boosting, which provides a reliable object detector for the tracking process; (2) a robust strategy to deal with tracking failures and recovery of such failures. Our idea is to incorporate decision of given by the prior learned strong detector and an on-line boosting tracker. This allows almost completely avoiding the drifting problem in tracking. Complex object can be learned and the object is initiated automatically at its first appearance. Moreover, the distinct advantage is we can almost completely make sure that the object is always detected and tracked when it appears; the abruption is also detected and failure will be recovered by re-detecting the object. The online adaptation tracker monitors the whole process and gives output of the system. In the intensive set of experiments on challenging data set for several applications, we demonstrate the out performance of our framework over very recent proposed approaches.