A flexible eyetracker for psychological applications

Abstract

We describe a practical method for measuring eye movements during psychological tests. This is an important class of applications including clinical evaluations and marketing studies. Existing methods in common use for psychological measurement, for example infrared reflection methods, are invasive involving head stabilization and special purpose lighting. In our experiments, we need to observe subjects for long periods, on the order of one hour. In addition, subjects must verbalize, which makes it difficult to stabilize their heads relative to the camera. We track the head using a lightweight spectacle framework worn by the subject. It has a set of easily visible colored balls. We segment each image into four characteristic colors, corresponding to iris, yellow ball, red ball, and background, which are obtained by sampling the images for each subject. The classification into colors is done by training a simple neural network for each characteristic color. We match a template to color-reduced image regions to find the balls and the two irises. We use a model-based object pose method, which uses a prior measurement of the relative positions of the balls, to calculate the spectacle framework pose (the head pose). A linear method is used for calibrating gaze position against head pose and iris positions. The subject's gaze position can be traded reliably for periods of more than an hour. The locations of image features are found with an accuracy of approximately one pixel of the image. In a 640/spl times/480 image of the whole face, the eyes are each about 80 pixels across. This gives a corresponding accuracy of calculated eye gaze position on a 17 inch monitor of about 1 cm horizontally and 2 cm vertically. This method has shown itself in practice to be very flexible for psychological measurement, giving sufficient accuracy and being noninvasive.