Driver sleepiness detection with deep neural networks using electrophysiological data.

Area of Science:

• Neuroscience
• Computer Science
• Transportation Safety

Background:

• Driver sleepiness is a significant factor in road accidents.
• Existing methods for sleepiness detection often rely on subjective or limited physiological measures.

Purpose of the Study:

• To develop and evaluate a driver sleepiness detection model using electrophysiological data and a deep neural network (DNN).
• To assess the effectiveness of combining convolutional neural networks (CNNs) and long short-term memory (LSTM) architectures for this task.

Main Methods:

• Utilized electrooculographic (EOG) and electroencephalographic (EEG) time series data from 269 drivers across 1187 driving sessions.
• Inputted 16,634 data segments into a DNN, employing both binary classification and regression outputs based on the Karolinska Sleepiness Scale (KSS).
• Evaluated model performance on a large, labeled dataset collected under naturalistic driving conditions.

Main Results:

• Achieved a subject-independent mean absolute error (MAE) of 0.78.
• The regression model demonstrated a binary classification accuracy of 82.6%, slightly outperforming a dedicated binary classification model (82.0%).
• Eye-derived data (EOG) proved more informative than brain-derived data (EEG); combined inputs offered minimal performance gains.

Conclusions:

• A regression-based DNN model effectively captures the graded nature of driver sleepiness, outperforming simple binary classification.
• The model surpasses traditional algorithms based on expert-defined features, indicating its ability to detect subtle signs of drowsiness.
• This approach offers a promising advancement for real-time driver sleepiness monitoring and enhanced road safety.