Abstract: A gaze estimation system based on a single camera and IR light was proposed. The eye corners’ location was used in this paper， and a twostep eye corner locating method combining USM coarse positioning with fine positioning by Gabor eye corner filter was presented. To solve the problem that the accuracy of traditional use of pupilPurkinje image vector for polynomial fitting decreases when head moves， the vector was corrected by using the distance of inner eye corners， and the error was compensated by using the support vector regression to build the relationship between eye characteristic parameters and the polynomial error. Combined with the accurate location of eye corners， the relationship between 2D features around the eye and the screen coordinates was built based on the designed system. Experimental results showed that the accurate eye gaze estimation within a certain range of head movement could be realized with the proposed system.

Key words: eye corner detection, eye gaze estimation, polynomial fitting, support vector regression