Object contour refinement via confidence voting

We propose a voting scheme for object detection and tracking in image sequences. When an object's contour is derived from such as the interframe difference data or from other approaches, a verification method is often desired to properly identify and further refine the contour of the detected object. The voting scheme is thus designed to extract a more accurate object contour by synthesizing those derived from several approaches with different levels of local confidence. The confidence on a contour indicates the reliability of segments of the contour generated through such as edge maps, motion detection or colour segmentation, and reflects how well the conditions that underpin the associated algorithms are met near the corresponding segments. Our experiments show the final synthesized contour will better represent the object to be detected and tracked.