Blink is a physiological signal that reflects drowsiness and concentration. It is important to detect driver's blinks without any wearable devices. For this purpose, a Doppler sensor has been used and several blink detection methods where a subject sits in front of such sensor have been proposed. However, it is challenging to detect driver's blinks because of face and body movement. In this paper, we propose a Doppler sensor-based driver's blink detection method in existence of face and body movement in a car. In the proposed method, blinks are detected through two steps: pre-detection and classification. In the first step which we call pre-detection, the time candidates of subject's blinks are detected based on spectrograms calculated from a received signal. Then, in the second one which we call classification, a set of features are calculated from a spectrogram and are fed into a supervised machine learning classifier to identify which time candidates are truly blinks. We leverage the fact that the distribution of the energy on a spectrogram differs between a blink and non-blink. Specifically, features are extracted based on the distribution of energy on a spectrogram. We conducted a series of experiments for the evaluation in the situation where a subject drives a real car in public road. As a result, we confirmed our method outperforms the conventional one in terms of F-measure calculated from recall rate and precision rate.
展开▼
