DrowsyDG-Phys: Generalizable driver drowsiness estimation in conditional automated vehicles using physiological...

Driver drowsiness is one of the leading causes of crashes, injuries, and fatalities on the road. Traditional drowsiness detection models relied on manually extracted physiological features processed through machine learning algorithms. However, these methods lacked flexibility and robustness across diverse real-world conditions. Although recent advances in deep learning have improved detection accuracy through automated feature extraction based on larger learnable parameter space, the generalization of existing models is still limited due to domain shifts. In this study, we proposed DrowsyDG-Phys, a novel domain generalization (DG) framework for driver drowsiness detection using three physiological signals (i.e., electrocardiogram, electrodermal activity, and respiration signals) that can be measured by in-vehicle or wearable sensors. Our approach introduced a backbone network for explicit time and frequency domain feature learning. In addition, our approach integrated three novel loss functions: a prior knowledge-based contrastive regularization for robustness, a feature centralization loss to promote generalization in heterogeneities, and a novel loss function to align drowsiness assessment criteria. Finally, we established a multi-source DG benchmark and evaluated our model on three existing datasets and a self-collected dataset involving 60 participants in a simulated SAE Level-3 driving scenario. Our proposed DrowsyDG-Phys achieves 78.5% accuracy on the DG protocol, as well as 88.4% accuracy on the cross-subject protocol. Experimental results demonstrated that DrowsyDG-Phys outperformed baseline methods, and improved generalization and robustness of physiological signal-based drowsiness monitoring.