An algorithm for model‐based stable pupil detection for eye tracking system

Abstract

AbstractThis paper considers eye images observed by a system such as an infrared eye‐movement observation system in which the relative position of the head and camera is kept constant. A highly accurate detection method for the pupil image is proposed. Assuming that the eyeball has a spherical shape and the contour of the pupil is a circle, the pupil is observed as an ellipse accompanying movement of the eyeball. By accurately detecting this pupil image of elliptical shape, it is possible to detect the eye direction stably. Various methods of detecting an ellipse have been presented, and the Hough transform is one of these. The Hough transform has the feature that it is less affected by noise and occlusion. In the detection of ellipses in general, the number of degrees of freedom is represented by five parameters to be determined, which may produce problems of computational complexity and memory requirements. On the other hand, in order that the pupil may rotate around the eyeball center, the pupil image observed depends on a position and a direction from the eyeball center, without becoming general ellipse form. In this study, the above constraint is utilized, and an algorithm is discussed in which the pupil is detected with high accuracy based on the eyeball model. In the algorithm, the processing efficiency of the whole system is considered. In the algorithm, first, initial detection of the pupil is performed by ellipse fitting based on the edges in multiple observed images, and the parameters of the eyeball model are estimated on the basis of the results. Then, the Hough transform is applied on the basis of the estimated eyeball parameters, and the pupil contour is detected with high accuracy. The experimental results show that the accuracy and stability of pupil contour detection is improved, indicating that the proposed method is effective. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(13): 21–31, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10544