Detection and prediction of driver drowsiness using artificial neural network models

Area of Science:

• Road Safety
• Driver Behavior Analysis
• Artificial Intelligence in Transportation

Background:

• Detecting driver drowsiness is essential for preventing accidents.
• Predicting the onset of drowsiness is more challenging than detecting its current state.
• Existing methods often rely on limited data sources.

Purpose of the Study:

• To assess if standard drowsiness detection data can predict future drowsiness levels.
• To investigate the impact of additional data (driving time, participant info) on prediction accuracy.
• To develop and validate artificial neural network models for drowsiness detection and prediction.

Main Methods:

• Utilized a car simulator to induce drowsiness in 21 participants over 110 minutes.
• Collected physiological (heart rate, respiration) and behavioral (head/eyelid movements, lane position) data.
• Developed two artificial neural network models for real-time drowsiness detection and prediction of time to reach a target drowsiness level.

Main Results:

• Behavioral indicators combined with additional data yielded the best detection and prediction performance.
• The detection model achieved a mean square error of 0.22 for drowsiness level.
• The prediction model accurately estimated the time to reach a specific drowsiness level with a mean square error of 4.18 minutes.

Conclusions:

• Driver impairment dynamics, specifically drowsiness, can be effectively predicted in controlled, monotonous driving conditions.
• Behavioral indicators are key predictors of future drowsiness states.
• Integrating diverse data sources enhances the accuracy of driver state monitoring systems.