In-Depth Inception–Attention Time Model: An Application for Driver Drowsiness Detection

Drowsiness while driving is a common problem for many drivers and a significant problem in contemporary society. This study presents a method for detecting drowsiness while driving. The key finding is that six channels of EEG data are closely associated with drowsiness detection; this finding will contribute significantly to the development of new drowsiness detection systems. To process EEG data with high frequencies and large datasets, an in-depth Inception model suitable for time-series data was employed, incorporating a self-attention mechanism. This model effectively extracts the time–frequency representation of EEG data using a short-time Fourier transform and selectively learns important features by applying the self-attention mechanism within the Inception block structure. Additionally, channel-wise convolution is utilized to reduce the dimensionality of input data, and modified Inception blocks are stacked to enable more profound data representation. The model manages its complexity by adding partial sequential convolution filters and self-attention to the Inception blocks while performing complementary roles. Our method achieved high-performance drowsiness detection with an accuracy of 79.02% using only six EEG channels. The method contributes to ensuring accurate detection by minimizing information loss through the introduction of a self-attention mechanism.