The long-term object tracking with online model learning

In this paper, the problem of long-term tracking an object in video sequences is addressed by means of online model learning. LK (Lucas - Kanade) algorithm is adopted in the tracker, and the object model is updated by online learning. In each frame, the object is described by the location and the scale. When the LK tracker fails to track the object chosen in the first frame, the online model is started to detect the potential object by the stored object models and reinitialize the LK tracker for subsequent tracking. In order to improve accuracy and stability of tracking, a criterion is proposed to estimate whether the LK tracker is failed. A threshold is introduced as well to control the number of online object models and further improve the real-time performance of the algorithm. The experimental results show that the algorithm can realize long-term stable tracking of the interested object in video sequences.