Comparative Study: Physiological-based Driver Drowsiness Detection Utilizing Traditional and Hybrid Methods

Driver drowsiness is identified as a major factor leading to traffic accidents and fatalities worldwide. To address this critical public concern, researchers are developing various driver-centered drowsiness detection systems. However, most of the studies in this field rely on either visual-based monitoring techniques or physiological-based indicators. This paper focuses on the latter and presents a comparative study aimed at identifying the most reliable and promising method for detecting the level of drowsiness using physiological signals. This study explores various techniques employing traditional machine learning methods or pre-trained deep neural networks. The former focus on generating features leveraging the temporal aspect of the signals using time series and frequency representation through Fast Fourier Transform, Discrete Cosine Transform, and Discrete Wavelet Transform. The latter type utilizes the spectrogram, which encapsulates both temporal and frequency information through visual representation, subsequently employed within a pre-trained Convolutional Neural Network model for feature extraction. The different approaches explored in this study, using the publicly available ULg DROZY dataset, have produced a multitude of results. However, the most promising approach achieved a top accuracy of 88%.