A Fast, Robust, Automatic Blink Detector

Abstract

Introduction “Blink” is defined as closing and opening of the eyes in a small duration of time. In this study, we aimed to introduce a fast, robust, vision-based approach for blink detection. Materials and Methods This approach consists of two steps. In the first step, the subject’s face is localized every second and with the first blink, the system detects the eye’s location and creates an open-eye template image. In the second step, the eye is tracked, using sum of squared differences (SSD). This system can classify the state of the eyes as open, closed, or lost, using the SSD-based classifier. If the eyes are closed as in usual blinking, the blink will be detected. To classify eyes as closed or open, two adaptive thresholds were proposed; therefore, factors such as the subject’s distance from the camera or environment illumination did not affect the system performance. In addition, in order to improve system performance, a new feature, called "peak-to-neighbors ratio", was proposed. Results The accuracy of this system was 96.03%, based on the evaluation on Zhejiang University (ZJU) dataset, and 98.59% in our own dataset. Conclusion The present system was faster than other systems, which use normalized correlation coefficient (NCC) for eye tracking, since time complexity of SSD is lower than that of NCC. The achieved processing rate for ZJU dataset was 35 fps.

Keywords

• Eye Localization
• Blink Detection
• Sum of Squared Differences