Online failure detection and correction for CAMShift tracking algorithm

Tracking failure is an inevitable problem in any object tracking algorithm. Online evaluation of a tracking algorithm to detect and correct failures is therefore an important task in any object tracking system. In this paper we propose an early tracking failure detection procedure for the Continuously Adaptive Mean-Shift(CAMShift) tracking algorithm. We also propose an algorithm to modify the tracker in order to correct the detected failures. CAMShift is a light-weight tracking algorithm first developed based on mean-shift to track human face as a component in a perceptual user interface, but it easily fails in tracking targets in more complex situations like surveillance applications. With our proposed failure detection and correction algorithm, CAMShift shows promising results in the test video sequences.