Calibration-Aware Driver Monitoring System for Real-Time Embedded Applications
Abstract
Driver fatigue and distraction play a crucial role in causing road traffic accidents, which necessitates an effective real-time driver monitoring system that is feasible in a resource-constrained embedded environment. In this paper, a system tailored for the user employing MediaPipe FaceMesh for extracting real-time facial landmarks, used for personalization and ML-based classification of the driver's state is proposed. The behavioral measures include Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), and head position measures. Personalization accounts for the variations between different individuals' faces and behaviors. For increasing the accuracy of the classification, features associated with eye closure, yawning, and distraction are used for training the Random Forest Classifier. Training is performed using both internal samples and modifications of NTHU Driver Drowsiness Detection Dataset, with classification carried out on five different individuals under various conditions in real time. The system runs at 30 frames per second, successfully detecting fatigue, drowsiness, and distraction, eventually classifying the driver's risk state as Normal or High Risk.