Abstract
The problem of eye tracking has been actively studied since the nineteenth century; its applications include a wide range of fields such as visual perception, ocular diagnosis, and human-machine interaction. The rapid development of compact electronic imaging sensors has made video-based methods among the most successful eye-tracking techniques. In these systems, an observer's point of regard is mapped from the difference vector between the pupil centroid and the first-surface corneal reflection (CR) of a light source. As such, localizing the pupil centroid and CR are crucial in making a robust eye tracking system. However, in existing eye-tracking systems, detection of the two eye features can fail when the eye moves to extreme positions and/or specular reflection artifacts occur due to uncontrolled ambient lighting. These challenges are particularly acute in wearable eye-tracking systems used to monitor natural behavior outside of the laboratory. In this paper, a new structured lighting configuration is proposed, which implements an array of infrared LEDs to illuminate the eye. The design provides robust difference vectors even during extreme eye movements. The challenge of spurious specular reflections from external sources or scleral reflections is solved by a two-stage processing approach. First, global and local statistical information from the eye image are utilized to isolate and determine the pupil centroid and potential CRs. Second, genuine CRs are distinguished by a novel physiological optics based technique - CR position prediction via a well-correlated pupil-CR offset ratio. A prototype implementation in the RIT Wearable Eyetracker is described.