Driver Drowsiness Detection Using Multi-Metric Modeling Based on Facial Landmarks

Abstract

Drowsiness is a major factor contributing to traffic accidents, as it significantly reduces driver alertness, reaction time, and decision-making ability. This study aims to develop a real-time driver drowsiness detection system based on multi-metric modeling using facial landmarks. Three physiological indicators were employed: Eye Aspect Ratio (EAR) to measure eye openness, Mouth Aspect Ratio (MAR) to identify yawning activity, and Percentage of Eye Closure (PERCLOS) to assess prolonged eye closure patterns. These features were extracted using MediaPipe Face Landmarker, a lightweight and efficient facial landmark detection framework. A quantitative approach with a rule-based method was applied without requiring machine learning training, making the system computationally efficient and easily deployable. Sliding window smoothing was incorporated to reduce false detections and improve overall detection stability. The system was implemented as an Android mobile application and evaluated in real-time conditions using the device's front camera. Experimental results demonstrate that PERCLOS serves as the most stable and reliable drowsiness indicator, while the integration of all three metrics yields significantly more accurate detection compared to relying on a single indicator alone. This system offers a promising non-intrusive, accessible, and practical solution for real-time driver monitoring.