AI-Based Driver Drowsiness Detection Using Eye Tracking: A Real-Time EAR–CNN Fusion Framework

Abstract

Driver drowsiness is a gradual and frequently underestimated safety risk that can impair attention before the driver recognises fatigue. This paper presents a realtime, non-intrusive drowsiness-detection framework using a dashboard-facing camera, facial landmarks, Eye Aspect Ratio (EAR), and a lightweight convolutional neural network (CNN). Each video frame is processed to locate the face and eyes, compute geometric eye openness, classify the eye state as open or closed, and apply temporal smoothing that separates normal blinks from sustained closure. The system is trained on approximately 84,000 labelled eye images from the MRL Eye Dataset and validated end-to-end on the NTHU Drowsy Driver Dataset. The fused EAR + CNN pipeline achieves 96.2% classification accuracy, 94.8% precision, 95.6% recall, and a 95.2% F1-score for the drowsy class, with an average alert latency of 1.3 s and a 3.1% falsealarm rate. A graded visual and audible alert is produced after sustained closure. The results show that a camera-only design can provide practical early warning on commodity and embedded hardware without wearable sensors.