Multistage Adaptive CNN–3D-CNN Framework for Robust Driver Drowsiness Detection Under Diverse Driving Conditions

Abstract

Driver drowsiness remains a leading yet often underestimated cause of road accidents, driven largely by impaired self-awareness and delayed reaction time among fatigued drivers. This study proposes a multistage adaptive CNN–3D-CNN framework for real-time driver drowsiness detection, designed to overcome the temporal and contextual limitations of conventional single-stage classifiers. The framework decomposes detection into specialized sub-modules covering head orientation, eye closure, mouth activity, and contextual conditions, before fusing these outputs through a 3D-CNN layer for spatiotemporal reasoning. Two benchmark datasets, UTA-RLDD (University of Texas at Arlington Real-Life Drowsiness Dataset) and NTHU-DDD (National Tsing Hua University Driver Drowsiness Detection Dataset), were harmonized and pre-processed to support robust model training, while data augmentation and class-balancing techniques addressed demographic and environmental variability. Comparative evaluation against existing face-detection models, including YOLOv7-Face, MTCNN (Multi-task Cascaded Convolutional Networks), and Hyperface-ResNet, demonstrated the proposed framework's superior accuracy, recall, and robustness under occlusion and low-light conditions. Cross-dataset validation across KEC-DDD (Kathmandu Engineering College - Driver Drowsiness Detection Dataset), NTHU-DDD, and UTA-RLDD further confirmed the model's generalizability across diverse driving populations. These findings support the framework's suitability for integration into Advanced Driver Assistance Systems, offering a deployable, explainable, and accurate solution for early fatigue detection.