A Lightweight MobileNetV3-FPN Framework for Driver Drowsiness Detection

Driver drowsiness remains a major cause of traffic accidents, requiring a drowsiness detection system to achieve high accuracy and real-time inference speed. To overcome these challenges, we propose a lightweight classification model integrating MobileNetV3 with a Feature Pyramid Network (FPN). The proposed model enhances multiscale feature extraction, offering significant advantages in detecting subtle signs of driver fatigue. The model was evaluated on two popular public datasets and achieved superior accuracy of 97.15 % on the Drowsiness-Prediction-Dataset and 99.4% on the NTHU-DDD dataset, outperforming state-of-the-art models. These results demonstrate its potential for practical deployment in real-time driver drowsiness detection scenarios.