EEG-Based Driver Drowsiness Estimation Using Self-Paced Learning with Label Diversity

Drowsy driving is one of the major contributors to traffic accidents. Continuously detecting the driver’s drowsiness and taking actions accordingly may be one solution to improving driving safety. Electroencephalogram (EEG) signals contain information of the brain state, and hence can be utilized to estimate the driver’s drowsiness level. A challenge in EEG-based drowsiness estimation is that when directly applied to a new subject without any calibration, the system’s performance usually degrades significantly. Many efforts have been devoted to reducing the calibration data requirement, but there are still very few approaches that can completely eliminate the calibration process. This paper proposes a self-paced learning approach, which also takes the label diversity into consideration. The model learns from the easiest samples when the training first starts, and then more difficult ones are gradually added to the training process. This training strategy improves the generalization performance of the model on a new subject. Experiments on a simulated driving dataset with 15 subjects demonstrated that the proposed approach can better reduce the estimation error than several other approaches.