REAL-TIME DRIVER DROWSINESS DETECTION AND ALERT SYTEM USING MACHINE LEARNING

Abstract

Driver drowsiness is implicated in approximately 20% of motorway traffic fatalities, yet most production vehicles lack real-time physiological monitoring. This paper presents a comprehensive, open-source driver drowsiness detection system that monitors Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR) from a consumer-grade webcam using MediaPipe Face Mesh at 30 FPS. A three-state finite state machine (NORMAL, WARNING, CRITICAL) with configurable temporal thresholds classifies drowsiness severity, and a three-tier alert escalation pipeline delivers proportional responses: on-screen visual overlay, audio alarm, and WhatsApp notification. A personalised calibration module adapts EAR and MAR thresholds to individual facial geometry. Evaluation on a 30- minute annotated driving sequence achieved 91.2% truepositive rate at WARNING level (8.1% false-positive rate) and 95.0% correct CRITICAL escalations with zero false WhatsApp alerts. The modular Python architecture (MediaPipe, OpenCV, Streamlit) enables deployment on commodity hardware without GPU acceleration.