Driver Drowsiness Detection Using Facial Analysis and Machine Learning Techniques

Abstract

Driver drowsiness detection is an important safety task because fatigue-related impairment can reduce attention and reaction time before a driver becomes fully asleep. Vision-based systems are suitable for practical deployment because facial cues can be extracted using a camera without requiring wearable sensors. However, systems mainly based on eye closure may miss drowsiness patterns expressed through yawning or head movement. This paper presents a driver drowsiness detection pipeline using a balanced Random Forest classifier trained on seven visual features: Eye Aspect Ratio, Mouth Aspect Ratio, Nose_X, Nose_Y, Pitch, Yaw, and Roll. The features were extracted from the 22-subject training and evaluation portion of the National Tsing Hua University (NTHU) Driver Drowsiness Detection Dataset, containing 794,254 frames. The pipeline applied standard scaling, random over-sampling, random under-sampling, and a fixed-hyperparameter Random Forest model. On a 70/30 stratified split, the proposed model achieved 98.30% accuracy, 94.93% sensitivity, 99.55% specificity, 96.80% F1 score, and 99.84% area under the curve. Five-fold cross-validation produced stable performance of 98.39% ± 0.03%. The proposed model also generalized well to 14 unseen subjects from the NTHU-DDD testing set, reaching an accuracy of 95.23%, which is only 3.07 percentage points lower than the 70/30 split result, confirming that the added head-pose features did not lead to overfitting. The results show that incorporating 3D head-pose features and class balancing improves sensitivity.