A Hough Based Eye Direction Detection Algorithm without On-site Calibration

This paper proposes an algorithm to estimate gaze direction without on-site calibration even when the position of the head relative to the camera moves. We define eye direction as geometrical direction of the eyeball and gaze direction as long fixed looking detection in this paper. In the algorithm to detect the eye direction, a cornea curvature center is detected based on a relation model between a light source and the cornea curvature center. A shape of the pupil is approximated using the detected center by ellipse Hough transform. The eye direction is estimated as a rotation angle of the approximated pupil using a cornea curvature radius and distance between the cornea curvature center and a pupil rim. This personal information on the eyeball is essential to the eye direction detection and the third personal information, a gap of the fovea, to the gaze direction detection. They are required measuring one time beforehand. However, on-site calibration is never to do in the gaze direction detection, which corrects the eye direction by the gap of the fovea. The effectiveness of our algorithm was confirmed about three point of views. The first is that the cornea curvature center was detected stably, the second that the eye direction for 4 examinees was measured with dispersion of almost same level as the involuntary eye movement, and the third that the gaze direction was detected stably moving the camera instead of the head.