Investigating Driver-Specific Variability in Vision-Based Driver Fatigue Monitoring Using a Computer Vision System

This preprint presents a research framework for investigating how driver-specific variability (facial appearance, eyewear, head pose, and illumination) affects the performance and generalizability of computer vision-based driver fatigue monitoring systems. Most existing systems are evaluated on subject-dependent datasets, which can overstate real-world performance. The paper proposes a comparative experimental design with subject-independent evaluation across driver subgroups, a statistical hypothesis-testing plan, and a survey-based sampling framework. It also presents DriverVision, a real-time prototype that computes Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR) from facial landmarks and classifies driver state as alert, drowsy, or microsleep. Prototype results are preliminary and come from a single subject; the full cross-subgroup experiment is planned as future work.