Personalized Classification of Scenario-Derived Operational Driver-State Classes from Non-Intrusive Wearable Signals...

Area of Science:

• Human-Computer Interaction
• Automotive Safety
• Wearable Technology

Background:

• SAE Level 2 automation requires continuous driver supervision.
• Robust driver monitoring systems are essential for safety.
• Wearable-based personalized monitoring is underexplored for real-world driving.

Purpose of the Study:

• Evaluate a personalized, wrist-worn driver monitoring system for SAE Level 2 automation.
• Assess the feasibility of intra-subject classification using non-intrusive signals.
• Investigate classification accuracy across different training scenarios.

Main Methods:

• Collected physiological and motion data from a wristband during real-world driving.
• Processed data using ResNet-50, PCA, and a supervised classifier.
• Employed a Leave-One-Experience-Out protocol for personalization assessment.

Main Results:

• Target-driver-only training achieved the highest classification accuracy (60%).
• External-user-only training yielded 50% accuracy, mixed training 54%.
• The low-demand baseline class achieved 88.4% accuracy and 70% F1-score.

Conclusions:

• Personalized wrist-worn classification of driver states is feasible under real-world SAE Level 2 conditions.
• Intra-subject classification demonstrates potential for unobtrusive driver supervision.
• Further research can refine wearable-based systems for enhanced automotive safety.