A diffusion-based feature enhancement approach for driving behavior classification with EEG data

Abstract

The recognition and prediction of driving behaviors play a significant role in addressing the substantial human factors involved in traffic safety. Electroencephalogram (EEG), as a sensitive physiological indicator, has unique advantages in detecting driving behavior compared to vehicle data. However, most existing studies only focus on a few specific driving behaviors, such as only considering braking, with a small amount of data. In this paper, we utilized an event-related simulated driving experiment to test five types of driving behaviors, and collected EEG signals from 35 subjects during the experiment. We proposed an encoder–decoder model structure containing a DDPM module for EEG signal classification. DDPM is able to enhance EEG features and solve the problem of insufficient sample size by generating new samples and learning reconstruction errors. We also analyzed the EEG response to event-induced behavior from the perspective of power spectrum. The topographical map of the power spectrum indicates a significant response to event-induced driving behavior within specific brain regions. In the classification experiment, our model achieved a classification accuracy of 82.12% on the partial dataset, and an accuracy of 83.65% across all participants, representing an improvement of 10.01%, 7.11% over comparison model EEG-Inception and EEG-Conformer. The results indicate that EEG physiological signals can be utilized for decoding driving behavior, thereby laying the groundwork for further in-depth investigations into real-world road traffic safety.