An eye-tracking based approach to gaze prediction using low-level features

Abstract: In this master's thesis, an attempt is made to automatically predict where people will look when watching video sequences. An application in the form of foveation for video compression is discussed. A relatively simple prediction model is built based on eyetracking data from several subjects and low-level features generated from the video frames, using simple image processing algorithms. The prediction model uses a new method to extract differences in feature distributions between frame regions that are watched and those that are not. It is first shown that these differences are significant. The differences are then used to predict which regions will be looked at in a new video sequence. The prediction is evaluated against eye-tracking data for the new video sequence and it is shown that the prediction is significantly better than random. Moreover, the accuracy of the prediction is compared to that of a group of humans predicting another group of humans. This comparison indicates that the proposed model needs improvement. Finally, a discussion follows about the possibilities and problems of the selected approach to gaze prediction.