IncorporationNet: a novel bimodal EEG-EOG vigilance estimation method via time-frequency-space feature fusion network.

Abstract

The assessment of driver vigilance is critical for promoting road safety, as it evaluates a driver's ability to sustain appropriate levels of attention and reaction capabilities. Electroencephalogram (EEG) and electrooculogram (EOG) signals have proven effective in this context. We propose a bimodal time-frequency-space feature fusion framework aimed at enhancing the integration of EEG and EOG features to improve the predictive accuracy of vigilance estimation. We combine LSTM with a Band-Spatial Attention Module (BSAM) to analyze EEG sub-band dynamics and EOG temporal patterns, then fuse both modalities through regression to enhance vigilance estimation while reducing noise. Validated on the SEED-VIG dataset, our solution achieves near-state-of-the-art performance in both RMSE and COR metrics. This bimodal vigilance monitoring approach introduces novel methodology with promising potential for real-time fatigue detection applications.