Real-Time Driver Drowsiness Detection Using Facial Biometrics

Abstract

Driver drowsiness is a progressive loss of vigilance whose visual manifestations may appear before a critical driving error. This paper presents an interpretable real-time framework that combines Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), Percentage of Eye Closure (PERCLOS), head-pose estimation, and temporal fusion to distinguish Normal, Yawning, Fatigue, and Microsleep states. The NTHU Driver Drowsiness Detection (NTHU-DDD) benchmark is used as the principal experimental reference, following subject-separated training, evaluation, and test partitions to reduce identity leakage. The proposed approach integrates frame-level facial measurements with temporal persistence and multi-cue decision rules, allowing short normal behaviors to be distinguished from sustained drowsiness-related events. Evaluation is performed at the event level over 2,000 driver-state events using accuracy, precision, recall, F1-score, false alarms per hour, and detection latency. The multi-cue framework achieves an overall accuracy of 91.65% and a macro F1-score of 91.41%, with 1.8 false alarms per hour and a mean detection latency of 0.84 s. Ablation analysis further indicates that temporal eye-closure information plays an important role in classification performance, while the complete multi-cue configuration reduces false alarms. The results support the use of interpretable facial and temporal information for real-time driver monitoring, while further validation under natural driving conditions remains necessary.