Blink detection robust to various facial poses.

Applications based on eye-blink detection have increased, as a result of which it is essential for eye-blink detection to be robust and non-intrusive irrespective of the changes in the user's facial pose. However, most previous studies on camera-based blink detection have the disadvantage that their performances were affected by the facial pose. They also focused on blink detection using only frontal facial images. To overcome these disadvantages, we developed a new method for blink detection, which maintains its accuracy despite changes in the facial pose of the subject. This research is novel in the following four ways. First, the face and eye regions are detected by using both the AdaBoost face detector and a Lucas-Kanade-Tomasi (LKT)-based method, in order to achieve robustness to facial pose. Secondly, the determination of the state of the eye (being open or closed), needed for blink detection, is based on two features: the ratio of height to width of the eye region in a still image, and the cumulative difference of the number of black pixels of the eye region using an adaptive threshold in successive images. These two features are robustly extracted irrespective of the lighting variations by using illumination normalization. Thirdly, the accuracy of determining the eye state - open or closed - is increased by combining the above two features on the basis of the support vector machine (SVM). Finally, the SVM classifier for determining the eye state is adaptively selected according to the facial rotation. Experimental results using various databases showed that the blink detection by the proposed method is robust to various facial poses.