An eigenspace approach to eye-gaze estimation

In this paper, we present a new approach to eye-gaze estimation using eigenspaces. The context of our application is monitoring car drivers. Within this context, eye-gaze information will be an important component in the development of systems that monitor the driver and alert him/her of any insecure driving conditions. Based on the requirements of our application, accurate gaze estimates are not really necessary. On the other hand, speed is very important. The approach we have taken involves classifying gaze into five directions ("straight", "up", "down", "left", and "right"). Each gaze direction is modeled by a distinct eigenspace. Novel gaze directions are classified as one of the above five gaze directions by computing the distance of the novel direction from each eigenspace. The eigenspace associated with the smallest distance determines the classification result. The proposed approach is very simple and fast. Also, it is not intrusive and does not required extracting any special features. Our preliminary results illustrate that it is a promising approach.