Abstract:In order to improve the accuracy of drowsiness detection, in this study multi source data for young and middle aged experienced drivers including vehicle lateral position, steering wheel controlling, and eye movement are collected in a driving simulator experiment. Meanwhile, the subjective drowsiness level of the drivers were also recorded and validated by replaying the videos. Based on those data, the decision tree model was established. The results indicate that the most significant variables to estimate drowsiness level are percentage of eye closure(PERCLOS), the standard deviation of lateral position, the time space area of lane crossing, the reverse rate of steering wheel and those variables are positively correlated to drowsiness level. Among these variables, PERCLOS is the best variable to divide drowsiness level. When PERCLOS is lower than 2.8%, there are no drivers in seriously drowsy state and when PERCLOS is higher than 21.9%, there are no drivers in non drowsy state; the total correct predicting rate is 64.31%. To verification the model, 4 drivers were selected from the 15 drivers randomly. The results of model validation showed the correct predicting rate of the decision tree model is 63.22%. In both experiments, the decision tree model doesn’t mistake seriously drowsy state for non drowsy state.