Guided Dual-Attention Networks for Compact Driver Drowsiness Detection: Performance, Calibration, and Edge Deployment

Abstract

We present a comprehensive empirical study of attention mechanisms for eye-based driver drowsiness detection, evaluating 13 model variants across accuracy, calibration, cross-dataset generalization, and FPGA edge deployment. We introduce the Guided Dual-Attention Unit (GDAU), which combines position-aware spatial attention with SE channel attention. The data pipeline first splits the full imbalanced dataset (103,803 images, 5.9:1 ratio) into stratified train/validation/test subsets; then, it applies undersampling only to the training set. On the naturally imbalanced test set, no attention mechanism significantly outperforms the others on an identical backbone: Channel attention achieves the highest raw accuracy (83.86%), while CBAM-L achieves the highest balanced accuracy (86.77%) and ROC AUC (0.927). GDAN achieves 82.83% accuracy (85.99% balanced) with only 2.19 M parameters—half of the baseline’s 4.29 M. Component ablation confirms spatial–channel complementarity (+2.37% balanced accuracy over baseline). No single model dominates all calibration metrics. Cross-dataset transfer fails for all architectures (48–57%, near random chance). On the physical Xilinx Kria KV260, DPU-accelerated inference achieves 0.481–1.116 ms (896–2077 FPS), confirming real-time edge deployment feasibility.