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

Area of Science:

• Physiological computing
• Machine learning
• Road safety engineering

Background:

• Driver drowsiness is a major cause of road accidents.
• Traditional drowsiness detection methods lack robustness and flexibility.
• Deep learning models struggle with domain shifts in real-world conditions.

Purpose of the Study:

• To propose DrowsyDG-Phys, a novel domain generalization framework for driver drowsiness detection.
• To enhance the generalization and robustness of drowsiness detection using physiological signals.
• To establish a multi-source domain generalization benchmark for driver drowsiness.

Main Methods:

• Utilized electrocardiogram, electrodermal activity, and respiration signals.
• Developed a backbone network for time and frequency domain feature learning.
• Integrated three novel loss functions: contrastive regularization, feature centralization, and drowsiness assessment alignment.

Main Results:

• Achieved 78.5% accuracy on the domain generalization protocol.
• Achieved 88.4% accuracy on the cross-subject protocol.
• DrowsyDG-Phys outperformed baseline methods in generalization and robustness.

Conclusions:

• The proposed DrowsyDG-Phys framework significantly improves driver drowsiness detection.
• The model demonstrates enhanced generalization across diverse datasets and conditions.
• Physiological signal-based drowsiness monitoring is made more reliable with this approach.