Multi-sensor driver monitoring for drowsiness prediction

Area of Science:

• Automotive Safety
• Human Factors Engineering
• Biomedical Engineering

Background:

• Driver monitoring systems (DMS) are crucial for enhancing road safety.
• Integrating diverse data sources can improve the accuracy and timeliness of drowsiness detection.
• Wearable technology offers a promising avenue for collecting physiological data.

Purpose of the Study:

• To develop and compare drowsiness detection models using vehicular, behavioral, and physiological data.
• To augment camera-based DMS with vehicle data and heart rate variability (HRV) measures.
• To analyze the timeliness of drowsiness prediction models.

Main Methods:

• Utilized data from a driving simulator, a production-type DMS, and an Empatica E4 wristband for physiological data.
• Collected drowsiness data via external rater observation and the Karolinska Sleepiness Scale (KSS).
• Developed nine binary random forest models using various data combinations and ground truths.

Main Results:

• Model classification accuracy ranged from 0.77 to 0.92.
• The top-performing model integrated physiological data and excluded missing segments post-HRV computation.
• The most timely model predicted drowsiness 6.7 minutes prior to a drowsy lane departure.

Conclusions:

• Physiological measures slightly improved model accuracy.
• Models using observational ratings were more timely in detecting drowsiness onset than those using KSS.
• Combining multiple data streams enhances driver drowsiness detection capabilities.