Institute of Transportation, MOTC

      Fatigue or drowsy driving is one of the major concerns for road transport safety. Statistics show that more than 70% of vehicle accidents come from risky driving behaviors, including fatigue driving. Drowsy driving is hard to be detected by inspecting the vehicle dynamic data because it accounts for very small proportion, hence it is highly data imbalanced. Synthetic minority oversampling technique (SMOTE) is applied to preprocessing the vehicle dynamics data, which is labeled by the fleet manager, for the data imbalance issue. Four machine learning models are applied for predicting drowsy driving
including support vector machine (SVM), random forest, InceptionTime, and Stacked-LSTM. Results show that although the average accuracy is 0.96 of these four models by using SMOTE, however, it cannot identify drowsy driving correctly. To more accurate predict the drowsy driving events by using the vehicle dynamic data, a time window slicing combining with target sample augmentation method is proposed for the data imbalance preprocessing issue.
     InceptionTime and Stacked-LSTM models are applied for training and learning the correlations within the vehicle dynamic data and drowsy driving style. Experiment results show the accuracy of proposed sliding window method with InceptionTime model is 0.7763, the F1-score is 0.791 which is better than the models using SMOTE method with the average F1-score 0.5. With the proposed sliding window method, it helps the deep learning models to better predict the drowsy driving.