MF-AHFNet: A Fatigue Driving Detection Method Based on Multi-Feature Asymptotic Hierarchical Fusion Network

Abstract

This study investigates driver fatigue detection using frontal electroencephalogram (EEG) signals and proposes a Multi-Feature Asymptotic Hierarchical Fusion Network (MF-AHFNet). The proposed model adopts a dual-branch architecture, in which Power Spectral Density (PSD) and Differential Entropy (DE) features are extracted via a multi-scale convolution module. Cross-attention and self-attention mechanisms are progressively integrated through an asymptotic hierarchical fusion strategy, enabling hierarchical semantic interactions between feature domains and enhancing discriminative representation learning. The proposed method was evaluated on the SEED-VIG dataset, which contains EEG recordings from 21 subjects across 23 sessions. Each 118-minute session was segmented into non-overlapping 8-second samples, resulting in 20,355 labeled samples in total. A sample-wise five-fold cross-validation scheme (80% training and 20% testing per fold) was employed, together with regularization techniques such as dropout, batch normalization, and early stopping to mitigate overfitting. Comparative experiments across multiple brain regions demonstrate the effectiveness of the proposed approach for fatigue detection under the present evaluation setting and validate the feasibility of using frontal EEG signals in this benchmark scenario. Specifically, MF-AHFNet achieved an accuracy of 95.77% using frontal channels, demonstrating competitive performance within the current protocol while retaining potential value for reduced-channel fatigue monitoring. These results indicate that the proposed method achieves a favorable balance between detection accuracy and model compactness under the present benchmark setting. Subject-independent generalization and broader real-world applicability remain to be further investigated in future work.