A Hybrid System for On-line Blink Detection

Automatic non-obtrusive deception detection is highly desirable because of its objectivity, accuracy and reliability. Eye blinking has been shown to be one of the informative nonverbal behavioural cues for solving this problem. Traditional blink recognition methods tend to use a tracker to extract static eye region images and classify those images as open and closed eyes to detect blinks. However, those recognition systems are frame based and do not incorporate temporal information. For this reason, they perform poorly as the tracker fails to detect eyes due to rapid head movement or occlusion. In this paper, we present an approach which combines Hidden Markov Models and Support Vector Machines to model the temporal dynamics of eye blinks and improve the blink detection accuracy.