Analysis Of Two Video Eye Tracking Algorithms

We are using a video tracking system for recording eye movements during vestibular experiments. The data are taken by a helmet mounted camera using infrared illumination. The measurement system need not run in real time but it must be tolerant to partial eye closings and other obstructions such as Purkinje spots. Two algorithms, each with its own advantages and disadvantages, have been adapted to our program shell which runs on the Macintosh IIx series computer. The algorithm requiring the least computation is a least squares (LS) fit of partial pupil edge data to a circular arch. It is vulnerable to out liers and biases due to finite edge width and asymmetries. A maximum likelihood method is computationally intense but is robust under many conditions of image obstruction which sends the LS algorithm astray.