VISUAL TRACKING with AUTOMATIC CONFIDENT REGION EXTRACTION

In this work, a novel efficient algorithm for visual object tracking in complex conditions is proposed. The main component of this work includes two parts: Bayesian decision based confident region extraction, and mean shift iteration based tracking. A unique characteristic of the proposed algorithm is that instead of tracking the entire object, the method automatically extracts the confident region of the object through fusing multiple cues in the Bayesian framework. Those cues contain object's color feature, motion character, and dynamic surrounding color information. We tested the performance of the algorithm with video sequences under difficult conditions (complex and dynamic background, fast camera motion, object maneuvering, rotations and partial occlusion) and achieved satisfied results.