Computer vision systems that monitor human activity can be utilized for many diverse applications. Some general applications stemming from such activity monitoring are surveillance, human-computer interfaces, aids for the handicapped, and virtual reality environments. For most of these applications, a non-intrusive system is desirable, either for reasons of covertness or comfort. Also desirable is generality across users, especially for humancomputer interfaces and surveillance. This thesis presents a method of gaze estimation that, without calibration, determines a relatively unconstrained user’s overall horizontal eye gaze. Utilizing anthropometric data and physiological models, a simple, yet general eye model is presented. The equations that describe the gaze angle of the eye in this model are presented. The procedure for choosing the proper features for gaze estimation is detailed and the algorithms utilized to find these points are described. Results from manual and automatic feature extraction are presented and analyzed. The error observed from this model is around 3± and the error observed from the implementation is around 6±. This amount of error is comparable to previous eye gaze estimation algorithms and it validates this model. The results presented across a set of subjects display consistency, which proves the generality of this model. A real-time implementation that operates around 17 frames per second displays the efficiency of the algorithms implemented. While there are many interesting directions for future work, the goals of this thesis were achieved.
Library of Congress Subject Headings
Eye--Movements--Computer simulation; Vision--Computer simulation; Computer vision
Department, Program, or Center
Computer Engineering (KGCOE)
Hnatow, Justin, "A theoretical eye model for uncalibrated real-time eye gaze estimation" (2006). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus