Real-Time Long-Term Visual Object Tracking

Visual object tracking with focus on occlusion, background clutter, image noise and unsteady camera movements, those all in a long-term domain, remain unsolved despite the popularity it experiences in recent years. This paper summarizes a related work which has been done in trackers field and proposes an object tracking system focused on solving mentioned problems, especially the occlusion, rough camera movements and the long-term task. Therefore, a system combined from three parts is proposed here; the tracker, which is the core part, the detector, to re-initialize tracker after a failure or an occlusion, and a system of adaptive learning to handle long-term task. The tracker uses newly proposed approach of bidirectional tracking of points, which are generally weaker then commonly used keypoints. Outputs of both the tracker and the detector are fused together and the result is also used for the learning part. The proposed solution can handle mentioned problems well and in some areas is even better then the state-of-the-art solutions.