Robust Tracking based on Failure Recovery

Object tracking is a issue in the domain of computer visual, most of current state-of-art approaches for visual tracking adapt tracking-by-detection, using detection to address tracking problem. While suitable for cases when the object is always in the sense and these algorithms always results in failures and can’t track back after failure. This paper we propose a tracking method based on failure recovery. After we choose an object to tracking in the first frame, the object is tracked by improved optical flow method forward and backward in time then compute the distance between these two trajectories. While the distance larger then threshold tracking likely to fail, but the latest object model return by detector will reinitialize the tracker. Tracking an object on camera video approve our approach can work at 20fps with long-time robustness.