Robust visual tracking via spatio-temporal cue integration

Appearance modeling is an important and yet challenging issue for online visual tracking due to the accumulation of errors which is prone to potential drifting during the self-updating with newly obtained results. In this paper, we propose a novel online tracking algorithm using spatio-temporal cue integration. Specifically, the object is represented as a set of local patches with respect to the spatial cue. In terms of the temporal cue, we keep the appearance models at different time and do appearance updating alternately. Taking full advantage of both historical and current information of the tracked object, the drift problem is alleviated. We also develop an effective cue quality measurement that combines similarity and motion information. Both qualitative and quantitative evaluations on challenging video sequences demonstrate that the proposed algorithm performs comparable against the state-of-the-art methods.