Adaptive Multi-Cue Kernel Tracking

This paper is a new attempt to introduce multiple cues to the kernel tracking by adaptive manner to improve the reliability and robustness of target tracking in the time-variant scenario. Based on Fisher rule, we construct the measure of discriminability to represent the ability of each cue in distinguishing the target from the background. According to the discriminability the weight of each cue is adjusted in time to accommodate the scene change, and then the cues are adaptively fused with kernel tracking method by these weights. In addition, we present a selective submodel update strategy via the discriminability for alleviating the model drift. In experiments, our scheme based on color cue and LBP texture cue is shown better effectiveness, compared with the well known mean shift tracker.