Eye-camera systems record a user's field of view and gaze points simultaneously. We have previously proposed an eye-camera system that has wide field of view and no parallax. To use the wide field of view effectively, this paper introduces an appearance based gaze estimation method. Although appearance based gaze estimation basically needs a long time to obtain lots of data for learning, the proposed method enables us to obtain data for the learning in a short time period by using a moving target. Normally a moving target data set is not suitable for learning because samples are biased towards the trajectory of the target. In our method, bias in the data set is normalized by using a discrete Voronoi tessellation and the gaze point is estimated by a k nearest neighbor method. The results of our evaluation show that the error of the gaze estimation is around 2.5° and the time for capturing a data set is around 60seconds.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/
