Efficient Modality Fusion Framework for Driver Cognitive Load Classification

초록

Accurate estimation of drivers' cognitive load is essential for ensuring driving safety and performance. Although unimodal physiological signals, such as EEG or ECG, are widely used, they often provide limited information, leading to suboptimal classification performance. To address this limitation, we proposed an efficient modality fusion framework that leverages multimodal physiological signals for cognitive load classification. The framework can extract features from EEG and ECG signals using parallel 1D convolutional layers and ResNet-Style blocks, followed by self-attention to refine intramodality dependencies and cross-attention to capture complementary inter-modality interactions. Experiments on the CL-Drive public benchmark dataset evaluated the framework's performance under both binary and ternary classification settings, using 10-fold cross-validation and leave-one-subject-out (LOSO) protocols. The proposed framework consistently outperformed conventional machine learning models and state-of-the-art deep learning approaches, achieving accuracies of 85.69% (10-fold CV) and 76.26% (LOSO) for binary classification, and 78.79% (10-fold CV) and 63.68% (LOSO) for ternary classification. These results highlight the importance of attention-based multimodal fusion for robust cognitive load estimation, suggesting its strong potential for applications in intelligent transportation systems and brain-computer interface development.

키워드