DEVELOPMENT OF DRIVER DROWSINESS CLASSIFICATION BASED ON EEG SIGNALS USING LONG SHORT-TERM MEMORY METHOD

Abstract

Driving accidents have a high prevalence cause of death with 1.3 million fatalities each year. Drowsy driving accounts for 26% of these accidents. A classification system for driver drowsiness was developed as a preventive countermeasure for the issue. The classification system was developed using Electroencephalograph (EEG) data recorded from four test subjects aged 19-24 during Virtual Reality (VR) driving simulations. The system categorizes drowsiness into four classes: awake, and three levels of drowsiness: mild, moderate, and severe, based on the reaction times measured during the Psychomotor Vigilance Test (PVT). EEG signals were processed using Discrete Wavelet Transform (DWT) to extract alpha, beta, and theta wave features. These features were then analyzed statistically using means and energy. This study compared the performance of four variations of Long Short-Term Memory (LSTM) models as a classification system. The most effective model was an LSTM with six hidden layers and 50 units per layer, achieving a test accuracy of 91% with the highest precision, sensitivity and F1-score, and accuracy average values of 0.93, 0.91, 0.92 across all classes.