Multi-Timescale Drowsiness Characterization Based on a Video of a Driver's Face

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Transportation Safety

Background:

• Drowsiness significantly contributes to fatal transportation accidents.
• Real-time drowsiness detection systems are crucial for driver safety.
• Current camera-based methods face a trade-off between detection accuracy and responsiveness.

Purpose of the Study:

• To develop a multi-timescale system for accurate and responsive drowsiness characterization.
• To address the accuracy-responsiveness trade-off in drowsiness detection.
• To enhance driver warning systems for preventing accidents.

Main Methods:

• Developed a multi-timescale drowsiness characterization system with four binary classifiers.
• Classifiers operated at distinct timescales: 5 s, 15 s, 30 s, and 60 s.
• Introduced a novel multi-timescale ground truth based on Psychomotor Vigilance Tasks (PVTs) reaction times.

Main Results:

• Achieved global accuracies of 70%, 85%, 89%, and 94% for the four classifiers.
• Demonstrated strong performance across varying timescales, balancing accuracy and responsiveness.
• Validated the system on 29 subjects using leave-one-subject-out cross-validation.

Conclusions:

• The multi-timescale system effectively characterizes drowsiness with adaptable accuracy-responsiveness trade-offs.
• This approach enhances the potential for timely and accurate driver drowsiness warnings.
• The system shows promise for improving safety in transportation and other critical domains.