Robust eye localization in video by combining eye detector and eye tracker

Eye detector and eye tracker have been individually used to solve the task of eye localization in video. Although the eye detection based approach seems to be robust especially in frontal view faces and opened eyes, its performance drops dramatically in the presence of large head pose change and closed eyes. Meanwhile, eye tracking based approaches can estimate closed eyes and eyes in extreme head poses using information from previous frames. Therefore, in this paper, we proposed to combine both tracker and detector for robust eye localization in video. Rather than sequential integration of these systems, our main idea is to use the eye locations suggested by an eye detector for initialization and measurement updating steps of particle based tracker. Experiments were conducted on two benchmark databases: TRECVID and Boston University Head Pose databases. The results show that our proposed method achieves a remarkable improvement compared to the state-of-the-art approach.