A Comparative Review of Vision-Based Driver Drowsiness Detection Methods

Driver drowsiness is a safety-critical driver state that can impair attention, slow reactions, and precede microsleep episodes. Vision-based monitoring is attractive because it can estimate behavioural signs without requiring the driver to wear physiological sensors. This review compares geometric and rule-based indicators, facial landmarks, eye-closure measures, head-pose analysis, conventional classifiers, deep neural networks, temporal models, and hybrid systems. The literature is orga-nized by sensing pipeline, temporal reasoning, evaluation design, computational demand, and deployment robustness. Founda-tional work supports the use of eye-closure measures such as PERCLOS, while landmark-based Eye Aspect Ratio (EAR) provides a lightweight blink cue. Deep models can learn richer representations, and temporal models can exploit the evolution of eye and mouth behaviour; however, reported results vary sub-stantially with datasets, label definitions, participant separation, and evaluation protocols. For example, one multi-CNN study reported 97.1% evaluation accuracy on NTHU-DDD, whereas a sequence-based UTA-RLDD study reported approximately 60% test accuracy and 93% specificity for a fuzzy-logic alternative, demonstrating that headline values cannot be compared without context. Persistent challenges include night-time illumination, sunglasses, occlusion, subject-independent generalization, class imbalance, and privacy. The review concludes that practical systems should combine temporal evidence with a reliability-aware fallback, report false-negative and false-alarm behaviour, and be tested on unseen drivers in realistic conditions. No new experiments are claimed. Index Terms—driver drowsiness, driver monitoring, Eye As-pect Ratio, PERCLOS, facial landmarks, deep learning, temporal modelling