Evaluating Performance of Non-intrusive Indicators on Drowsy Driving Detection

To detect the drowsiness state of the driver more accurately, a high-fidelity driving simulator was adopted to collect the driving behavior and eye movement data of drivers, and questionnaires were used to collect their subjective sleepiness scales. Based on these data, 23 non-intrusive indicators were developed. The indicators were evaluated from the aspects of their correlation with drowsiness, binary classification performance, sensitivity to road alignment, and individual homogeneity; and then ranked in terms of comprehensive performance. The results show that the percentage of eyelid closure time (PERCLOS) has the highest correlation (0.443) with the Karolinska sleepiness scale (KSS). The best binary classification performance of eye movement and driving indicators lies on PERCLOS and the lateral position standard deviation (LPSD). For eye movement indicators, the difference between straight line and curve is lower than 20%. When KSS is 7, all the indicators show significant individual difference except the percentage of lane crossing time of vehicle center. The top three non-intrusive indicators with the best comprehensive performance are PERCLOS, LPSD, and the average lateral speed during lane departure.