Learning to track curves in motion

Recent developments in video-tracking allow the outlines of moving, natural objects in a video-camera input stream to be tracked live, at full video-rate. The system used here is based on Kalman filtering with a B-spline representation of curves to track the silhouettes of moving non-polyhedral objects. For example hands, lips, legs, vehicles, fruit can be tracked at video-rate without any special hardware beyond a desktop workstation and a video-camera and framestore. The novel contribution of this paper is a tracking algorithm that uses a bootstrapping technique to learn a stochastic, dynamic model for given motions from example video-streams. Incorporating such a model into the tracking algorithm greatly enhances maximum tracking speed and robustness to distraction from background objects. Experiments with learning both rigid and non-rigid motions, using moving hands and lips, clearly show the increased tracking power resulting from the learned dynamics