Driver Drowsiness Classification Using Data Fusion of Vehicle-based Measures and ECG Signals

Reduced alertness due to the drowsy state that impairs driving performance has been reported to be one of the significant causes of road accidents. This paper aims to present a data fusion of vehicle-based and ECG signals for classifying three levels of driver drowsiness, including alert, moderately drowsy, and extremely drowsy. Lateral deviation from the road centerline, steering wheel angle, and lateral acceleration are employed as vehicle-based signals. Two ECG leads are also exploited to collect heart rate variability of drivers. Thirty-nine features from vehicle-based data and ten features from heart rate variability signals are extracted. Finally, k-nearest neighbors and random forest are used as classifiers to classify the level of drowsiness using selected features by the sequential feature selector. Age and gender, as the two most effective human factors, are considered to assess the performance of the method in different age/gender groups. The proposed method is evaluated on experimental data that were collected from 93 manual driving tests using 47 different human volunteers in a driving simulator. Results show that hyperparameter-optimized random forests obtain an accuracy of 82.8% for the detection of drowsiness levels based on vehicle signals only, and an accuracy of 88.5% based on ECG derived data only. Data fusion of ECG signals and vehicle data improves the accuracy of classification to 91.2%. The model performs slightly better on older than on younger drivers, but no gender difference was found.