A four state video eye tracking algorithm using maximum likelihood estimation.

Abstract

This thesis presents the design, analysis and implementation of an algorithm which determines the yaw, pitch, roll and pupil diameter state of an eye viewed with a standard video camera. The usual illumination is infrared and no invasive contact lens is required. A maximum likelihood estimation technique tracks the location and size of the pupil in a video image to find horizontal and vertical eye position. Simulations and analyses show that the noiseless measuring resolution of horizontal and vertical movements is less than 0.05 pixel on an image. Based on accurate measurements of pupil position, counterroll movements are calculated using cross correlations between one dimensional templates which consist of equidistant pixels on a partial annulus overlying the iris and concentric with the pupil center. Another advantage of the algorithm is a robustness with respect to intrusions of droopy eyelids and random light reflections. Analysis shows that eyelids which cover pupils by less than a third of pupil radius do not cause a bias in pupil position estimates. Light reflections on the pupil boundary have a minimal effect on estimate bias, while light reflections embedded inside the pupil have a lesser effect. The algorithm's overall robustness is shown through the analysis of millions of eye images from the Microgravity Vestibular Investigation (MVI) experiments. The speed of image analysis (about 10 frames per second on Macintosh IIfx computer), the robustness for eyelid cover and random light reflections, and the ability to track 4 dimensional eye movement (horizontal, vertical, counterroll movement and pupil size) are major characteristics of the algorithm. This algorithm has been used to generate eye movement data for millions of video images from MVI experiments. Extensive use of the algorithm proves that it generates reliable eye movement data with good accuracy.