Accurate Eye Tracking For Autostereoscopic Displays

Abstract

Autostereoscopic displays have the potential to play a signi?cant role in the multimedia industry in the coming years. Their big advantage is that the viewer does not need glasses to experience 3D. However, static view cones in which the 3D looks best, restrict the viewer to very speci?c positions in front of the display. For single viewer usage, Dimenco developed dynamic clear view, which allows the viewer to move freely in front of the display and still experience good quality 3D on each position. This is obtained by tracking the eyes of the viewer and adjusting the rendering of the 3D content accordingly. However, tracking the viewers eyes for this purpose means the tracker has to be fast enough to keep up with the content and accurate enough to retrieve the best possible viewers location. The speed required is 30 frames per second and this is particularly di?fficult because state of the art methods that obtain good tracking result, often lack speed and vice versa. Obtaining good accuracy from the eye tracker is a challenge because this tracker has to work under many di?fferent circumstances. This thesis presents an eye tracker that meets the requirements set by Dimenco in both speed and accuracy. These results are obtained by combining the state of the art face tracker introduced by Viola and Jones with a particle fi?lter for fast face tracking, and by using the local minimum, adaptive gradient masks and isocenters to determine the pupil location. This combination of methods contributes to this fi?eld by running at atleast 30 frames per second and by tracking as accurate as needed. The accuracy that is needed was determined by an experiment with an autostereoscopic display from Dimenco and for that accuracy this method outperforms other eye tracking methods to which it is compared to. Problems with eye tracking still remain when the viewer wears glasses, mostly when light reflection is visible on these glasses.