Random values of a state-space variable are generated (the so-called particles), used in a description of the system, and checked against the current output measure to generate a weight value, or likeliness of that particle being the one that best describes the current state of the system. Particle filtering is a model estimation technique based on Monte Carlo simulations. The constraints we impose on this tracker are: it has to be completely automatic, robust against rapid movement and partial occlusion, work with just one camera (no stereo-vision) on a conventional PC, and be able to track several faces simultaneously in real-time.

This system could be applied to a wide range of video-based applications such as computer mediated communications. Our motivation for developing a real-time face video tracker is to advance research on a system that can infer conversation structure from video sequences of face-to-face communication, in which it is assumed that the gaze direction of the participants, which provide cues for discerning the conversation structure, can be identified from head directions. Fast and robust object tracking in video sequences is required by many applications in many fields: automated surveillance systems follow their targets using these techniques, robots rely on it to perform navigation tasks or man-machine interaction, augmented reality systems depend on the position data acquired by visual tracking to place their virtual objects in the real world, video-games or assisted devices can be controlled thanks to a camera and some face or hand tracking software, to name just a few.

Realtimes seek tracker