Detecting the gaze direction for a man machine interface

In this paper we describe a technique to determine the gaze direction of a human operator. Beginning with a head extraction, done by a color segmentation, the eye position gets estimated by a corner detector, taking into account, that the eyes must lie on approximately the same eye level. The iris position is found in two steps. First, a region growing algorithm marks the darkest area for each eye, and a template matching determines the precise position in this pre-marked region. Second, a hough transformation builds up a second hypothesis on the iris position, based on the results of a canny edge detector. By fusing these two results, wrong determinations are excluded. Furthermore, the eye corners had to be found to evaluate their relative position to the iris. This is done by a parametric eye model, finding the best correspondence of the model and the image features. Now, having the iris and the eye corner positions, a gaze direction can be determined. The heading direction is considered to be known.