NoFatigueDriving: Adaptive Cross-Modal Fusion for IoT-Enabled Wearable Driver Fatigue Detection

Driver fatigue is a major cause of traffic accidents, yet single-modality wearable approaches struggle to capture its multidimensional nature under resource constraints. This paper presents an end-to-end Internet of Things (IoT)-enabled driver fatigue detection system that decouples data acquisition on a lightweight headband from cloud-hosted inference. At its core is the Trimodal Fatigue Detection Network (TFD-Net), which utilizes signal-specific encoders, bidirectional pairwise cross-modal attention, and a two-level adaptive gated fusion module. On a 12-participant dataset, TFD-Net achieves a state-of-the-art accuracy of 94.40 ± 2.93%, outperforming the best baseline by 2.16%. Ablation and t-SNE analyses confirm that our trimodal fusion improves accuracy by 5.44% over Electroencephalography (EEG)-only models while providing high biophysical interpretability. This system demonstrates the potential of cloud-assisted multimodal wearable sensing for robust, real-time driving safety applications. The code is available at GitHub.