A new approach for long-term person tracking

This paper investigates long-term visual person tracking using particle filter as the underlying framework and online boosting as the detection strategy. In the case of the being tracked person with abrupt motion, under occlusion or in low sample rate of video source, two main issues rise inevitably. One is the poor constraint of person motion model, and the other is the drastic variation of pose or incomplete appearance when the person reappears. We address the problems with an integrated framework of multiple observers, and online boosting algorithm with independent features and its static and dynamic combination aiming to balance the tradeoff of adaption and drift.