An Improved Pupil Detection Method under Eyeglass Occlusions

There are various challenges of detecting pupil during eye tracking, such as changing illumination conditions, occlusion of eyelashes or eyelids, obstruction of prescription glasses, poorly recorded images, highly off-axial positions, and so forth. Prior state-of-the-art method namely ExCuSe undertakes these problems based on analysis of histogram intensity. However, ExCuSe fails to analyze some pupil images with poor illumination and light reflection occlusion caused by prescription glasses. To overcome this problem, this research proposes an improvement in ExCuSe by incorporating two image filtering techniques in the preprocessing step. The median filter is utilized to diminish noise while the guided filter is implemented to preserve edges in the image. We evaluated the improved and the state-of-the-art algorithm on over 16,000 hand-labeled images in three data sets that contain eyeglass occlusions. The experimental result of data set III shows that the proposed method significantly outperformed the state-of-the-art algorithm with a 22.53% higher detection rate (p<0.05). Although implementation on the other two data sets did not achieve a statistically significant result, the overall performance of the proposed method was still better than the state-of-the-art algorithm. Our study indicates that the proposed method is more sophisticated to handle poor illumination and light reflection occlusion compared with the prior state-of-the-art technique. In future, the proposed pupil detection method can be implemented in an eye tracker for interactive systems as well as for passive monitoring system.