Eye Gaze Estimation Based On Accurate Eye Corner Location

With the continuous development of Human-Computer Interaction Techniques, the eye gaze has been widely concerned as a new interactive interface. Most of current gaze estimation methods limit the user’s head movement, besides, they need multiple cameras and IR lights and other auxiliary equipment, system set up is complicated and costly. In order to simplify the system building process and eliminate the restrictions on head movement, a gaze estimation system based on a single camera and IR light is proposed in this paper. The system, combined with the accurate location of eye corners, realizes the eye gaze estimation within a certain range.The eye gaze estimation approach proposed in this paper makes full use of the eye characteristic parameters when looking in different directions. Based on the dark pupil phenomenon on the corneal surface under IR light, the pupil edge is extracted and filtered, and finally the pupil ellipse is obtained by fitting at least squares principle. According to the characteristic that the Purkinje image appears to be highlight, the weighted centroid of the Purkinje image area is located near the pupil. Different from previous gaze estimation methods, the eye corners’ location is used in this paper, and a two-step eye corner locating method combining USM coarse positioning and fine positioning by Gabor eye corner filter is presented. Accurate feature detection provides a strong basis for the gaze estimation.To solve the problem that the accuracy of traditional use of pupil-Purkinje image vector for polynomial fitting decreases when head movement happens, this paper proposes to correct the vector by using the distance of inner eye corners, reducing the effect of human-computer distance’s change on the fitting accuracy. In addition, for the error cased by other forms of head movements, this paper puts forward to compensate by using the Support Vector Regression to fit the relationship between eye characteristic parameters and the error. Experimental results show that the error of polynomial fitting can be effectively reduced and accurate eye gaze estimation under free head movements can be realized by the distance compensation and SVR error compensation. The average estimating error of screen coordinates of the fixation point is about9.54mm.