Evaluation of three template matching algorithms for registering images of the eye.

Abstract

The purpose of this paper is to evaluate the ability of three similarity measures to register a template in a sequence of near-infrared eye images. Three measures are evaluated: the normalized correlation coefficient (rho), the sum of absolute valued differences (SAVD), and a relatively new technique based on a sign change criterion, called the stochastic sign change criterion (SSC). Performance is measured in terms of the method's ability to track the pupil center derived from fitting a general 2-D ellipse to the pupil contour. Experiments using static eye images indicate that rho is not a reliable similarity measure. Conversely, both SAVD and SSC show enough potential to merit an investigation into their performance in tracking dynamic eye movements. It is anticipated that SAVD and SSC can accommodate a range of eye movements spanning at least 60 degrees horizontally and 40 degrees vertically.