Doppler Sensor-Based Blink Duration Estimation by Analysis of Eyelids Closing and Opening Behavior on Spectrogram

Blink duration is one of the useful indicators to estimate drowsiness and fatigue. A Doppler sensor could be a key device to realize the non-contact blink duration estimation, which is very useful for the drowsiness and fatigue monitoring in real life. However, none of the blink duration estimation methods has been proposed so far. In this paper, we propose a novel Doppler sensor-based blink duration estimation method based on the analysis of eyelids closing and opening behavior on a spectrogram. When one blinks, the energies caused by eyelids closing and opening behavior appears on a spectrogram. Hence, the blink duration can be estimated by integrating such energies and then detecting the timings when the integrated energy caused by the eyelids closing behavior appears and when the energy caused by the eyelids opening behavior disappears. To precisely detect such timings, a spectrogram with a short time window is better. However, thorough the preliminary experiments, we confirmed that when a shorter time window is used, more noise appears around such timings, which makes it difficult to detect such timings. To deal with this issue, the proposed method multiplies several integrated energies for spectrograms with different time windows. By multiplying the integrated energies with such noise and the ones without such noise, the energy level of such noise gets significantly declined over the product. Furthermore, the levels of the redundant blink energies also decline to the noise level, where the redundant blink energies denote the ones that appear before and after the eyelids closing and opening timings because of a long time window. The results of the experiments on 8 subjects watching a movie showed that our method achieved an average root mean squared error (RMSE) of 49.3 ms.