A Subject-Independent Temporal Framework for Behavioural Eye-Based Driver Drowsiness Detection

Driver drowsiness detection systems are often evaluated using data partitions that may retain observations from the same individuals across training and testing, making their ability to generalise to unseen drivers difficult to assess. This study investigates behavioural eye-based drowsiness detection under strict subject-independent evaluation using 65,929 facial image frames from four drivers in the NTHU Driver Drowsiness Detection dataset. Eye Aspect Ratio (EAR), Rolling Mean EAR, EAR Delta and an Eye Closure Indicator were represented over temporal windows of 10, 20 and 30 frames and evaluated using Logistic Regression, Random Forest, Long Short-Term Memory and Bidirectional Long Short-Term Memory classifiers under Leave-One-Driver-Out validation. Logistic Regression achieved the highest mean accuracy, whereas Random Forest at 30 frames produced the highest drowsy-class F1-score and Area under the Precision-Recall Curve. Controlled baseline analysis showed that temporal context contributed more substantially to performance than the additional engineered EAR features, while recurrent architectures did not consistently outperform the classical classifiers. Performance also varied across unseen drivers, and the modest absolute accuracy cautions against interpreting the results as evidence of deployment-ready detection. The study demonstrates the importance of separating driver identities during evaluation and provides a reproducible framework for examining temporal behavioural representations under unseen-driver conditions.