Comparison of Features Based on Spectral Estimation for the Analysis of EEG Signals in Driver Behavior

Anticipating such catastrophic road accident is one of the main interest of Malaysian government policies, vehicle manufactures strategies, and research in automotive safety. The vital aspect that causes road accident is a human error due to the behavior of the drivers. This research presents approaches for alertness detection based on the electroencephalography and power spectrum to evaluate driver's vigilance level in a static driving simulator. Twenty datasets were acquired using the eight-channel wireless headset for EEG signals recording. Two driving tasks, that are Alert Driving (AD) and Fatigue Driving (FD), were instigated for data acquisition, to discriminate the driver's alertness level. The AD task is designed to demonstrate the alertness condition and FD task is for the fatigued state. The acquired EEG signals are preprocessed to eliminate the noise interference and segmented into equal-length frame samples. The frame samples (signals) are segregated into four frequency range signals (i.e. Delta, Theta, Alpha, and Beta) and used to select the distinctive attribute. To estimate the alertness level, four different frequency-domain power spectral density (PSD) feature extraction techniques were evaluated (specifically periodogram, Lomb-Scargle, multitaper, and Welch). Multilayer Neural Network (MLNN) is used to evaluate the performance of all the extracted features. Ten-fold cross-validation method was used for validating and testing the reliability of the classifier model. The highest average accuracy is obtained from PSD using Welch with 96.7% and 85.0% for training accuracy and testing accuracy respectively.