ORB tracking via random model and sample consensus

The ORB feature is a fast binary keypoint descriptor that is efficient enough to be applied to many real-time computer vision problems. In this paper, we propose a novel visual object tracking approach by representing the tracked object with an ORB feature set. We find the keypoint correspondences between two consecutive frames using the nearest neighbor search method in the two corresponding ORB feature sets. A random model and sample consensus method is then applied to the keypoint pairs to find the most appropriate motion model among multiple object transformation models. While tracking, we will update the ORB feature set by adding new features and pruning outliers. The proposed method is tested on several challenging video clips to show impressive tracking performance.