Domain-robust vision transformer with hierarchical swin encoding for explainable low-latency driver drowsiness detection

Driver drowsiness is a significant cause of road accidents worldwide, often leading to fatalities due to delayed reaction times and loss of vehicle control. Traditional fatigue detection systems face several limitations, including poor generalization across various conditions, a lack of model transparency, and an inability to operate on low-power platforms. To address these issues, we propose Enhanced Multi-path Token-Fusion Vision Transformer (ViT), Light-VTD, a novel lightweight architecture designed for accurate and explainable fatigue detection from facial images. Unlike existing transformer-based methods, Light-VTD features a fused-MBConv block for efficient local feature extraction, a position-aware token mixer for capturing global context, and a multi-path token fusion mechanism that enhances spatial consistency across different resolutions. Furthermore, we adapt Gradient-weighted Class Activation Mapping (Grad-CAM) to generate interpretable attention heatmaps that align with clinically relevant fatigue indicators. Our experimental models are trained on a large-scale dataset comprising over 167,000 labeled images, compiled from four diverse public datasets: MRL Eye, nthuDDD2, and UTA-RLDD. The Light-VTD model achieves an intra-dataset accuracy of 99.2% and at least 93% in cross-dataset transfer, outperforming current ViT baselines by 2-4%. It supports inference with