International Journal of Science, Engineering and TechnologyAuthors: Aayush Chowdhury, Debdutta Basu, Sagnik Roy, Soumi Mukhopadhyay, Samarjeet Kumar, Anjan Kumar Payra
Abstract: Driver fatigue contributes significantly to traffic accidents worldwide. This paper presents a hybrid driver- monitoring framework combining Eye Aspect Ratio (EAR) analysis with Convolutional Neural Networks (CNN) to enhance drowsiness detection reliability. Unlike single-cue approaches, the proposed system leverages both geometric eye-landmark features and learned visual patterns through MediaPipe Face Mesh for facial landmark localization. Three eye-landmark configurations were systematically evaluated, with Set 3 achieving optimal performance. A lightweight CNN was trained on 64×64 pixel eye images using 5-fold stratified cross-validation with data augmentation. The hybrid system employs an OR- based fusion rule prioritizing safety sensitivity. Results demonstrate that the standalone CNN achieved 79.07% accuracy (AUC = 0.7732), while the optimized EAR model (Set 3) reached 89.53% accuracy (AUC = 0.8929). The hybrid approach reduced false negatives by approximately 92%, achieving 97.3% sensitivity and 81.40% accuracy (AUC = 0.8348). The system operates at 20-25 FPS on standard CPU hardware, confirming real-time viability for Advanced Driver Assistance Systems (ADAS) integration.
