Active Appearance Models for Gaze Estimation

This thesis describes the implementation of and the experimentation with a gaze estimation system based on an approach called the Active Appearance Model (AAM). Active Appearance Models are generative models of a certain visual phenomenon. They can be used to model specific patterns of variability in shape and grey-level appearance, which in turn can be used directly in image interpretation. Our goal for developing a gaze estimation system is a low-cost user friendly system. Therefore, we wish to design a system that can be used with normal off-the-shelf equipment, like a normal resolution webcam without infrared capabilities and infrared lighting. Our method uses separate modules for subtracting both the head pose and the eye angles from an image depicting a face. The head pose is necessary information, since we want to allow the user to move its head. We basically tested two different methods for calculating/estimating the gaze angle. The first one being a direct calculation method based on important positions of parts of the human eye, and a geometric gaze estimation model. The second method uses a Neural Network classifier to estimate the gaze direction based on a representation of the eyes of the user (we tested several representations). This second method proved to be most promising, where the best performing representation was a combination of the landmark vector and the appearance vector of our AAM (of an eye). We showed that is possible to obtain quite accurate results from our gaze estimation system, although these results are not as accurate as the results of other systems using a infrared camera and/or higher resolution images. Depending on the field of application (where only moderately accurate results are required), our method can be a cheap alternative to other more expensive systems.