Eye corner detector robust to shape and illumination changes

Abstract

We introduce a robust and accurate eye corner detector. The purpose of this technology is to improve the robustness of natural user interfaces' inputs, such as eye movements, in real-life environments for various users. Our technology relies on the pupil centers and particularly the inter-pupil direction, to define simple features capturing the structural characteristics of the eye corners independently of the shape and appearance of their neighborhoods. Our method proved to be robust to different lighting environments, eye shapes and different patterns of reflection on glasses. We report an error below 2% of the inter-pupil distance for 98% of the images, and maintain processing time below 3ms for computation on both eyes.