Eye localization in video by combining eye detector and eye tracker

Abstract

In this paper, we propose a method  for combining eye tracker and eye detector for robust eye localization in video. Instead of sequential intergration of the two systems, we use eye locations suggested by an eye detector for initialization and measurement of updating steps of particles used in an eye tracker. This combination allows  to improve the localization performance, since the detector provides good estimation of eye location candidates, meanwhile the tracker allows to find the best eye location by using temporal information. Experiments  are conductedon two benchmark video databases (TRECVID and Boston University Headposedatasets) and videos from Vietnamese Television. The obtained results show that our method achieves  better effectiveness  compared to the state-of-the-art eye detector and eye tracker.