Non-intrusive eye gaze estimation using a projective invariant under large head movement

Abstract

This dissertation deals with development of an eye gaze estimation system. We can know the thought or the object that human is interested in by understanding human eye gaze. Eye gaze estimation system can be used as human-computer or human-machine interface, and especially, it is very useful for the disabled who cannot use their arms. Many researches to estimate eye gaze direction or eye gaze point have been done. The researches focus on the improvement of accuracy or head movement. The conventional methods utilize eye gaze direction and eye position in 3D for al-lowing head movement. Thus, eye gaze direction estimation system and eye pose estimation system are integrated to find eye gaze point. There are two sources of error: the error of eye gaze direction and the error of eye pose. Especially, small error of eye pose results in large error of gaze point. In this dissertation, eye gaze point estimation method without eye pose is suggested in order to overcome the problem of the conventional methods. For allowing large head movement, five light sources and two cameras are used. The cameras are mounted on a pan-tilt unit: face tracking camera and eye tracking camera. The face tracking camera is a wide view camera to track a face. Face detection algorithm finds the position of the face in the image from the face tracking camera, and the pan-tilt unit is controlled according to the position of the face. Then, the eye tracking camera captures the eye in large scale because it has high zoom lens. In the image from the eye tracking camera, the positions of the five glints and the pupil center are determined by feature tracking methods. After obtaining the feature positions, the eye gaze point can be computed by the projective invariant. First, fast face tracking method is proposed for eye gaze estimation system. In order to make fast face tracking system, search space reduction by using color filtering and AdaBoost pattern classification method are used. The face de...