摘要

A tracking method based on adaptive relay detection using primary and auxiliary detectors is proposed. In this framework, the tracking problem is formulated as the continuous relay detection, where primary detector and auxiliary detectors collaborate to locate the target and are updated online. Each of the detectors corresponds to one of the appearances of target that have appeared. Primary detector that always corresponds to the current appearance of target is set to conduct object detection, while auxiliary detectors that correspond to the previous appearances of target are used to re-detect the target when it shows a previous appearance. To achieve better classification with less features and ferns, the detectors are constructed based on the feature selection by the mutual information. As the previous appearances of target are recorded by the detectors correspondingly and only primary detector needs to update, our tracker can achieve long-term real-time object tracking in unconstrained environments. Experimental results on challenging real-world video sequences demonstrate that our tracker outperforms most of the state-of-the-art methods.