Image based eye-gaze estimation irrespective of head direction

One can roughly estimate eye-gaze direction of a person or a portrait. In particular, it seems likely that one can accurately recognize whether or not one meets a person's eyes directly. If a machine realizes such visual information processing as a man does, one can tell one's intention to the machine with one's eyes. Most previous methods of eye-gaze estimation use a special light source or a hardware device. These methods however are not user friendly because they require the user to restrict his motion. We propose a new method for eye-gaze estimation. This method, based on image processing, uses only a single facial image obtained by a CCD camera under ordinary lighting condition without using any instrument except a pair of small marks. Experimental results show that the proposed method produces good accuracy irrespective of the head direction.