Research on Facial Micro-Expressions Based on Deep Learning for Driver Fatigue Detection

Abstract

Detecting driver sleepiness is crucial for preventing accidents. The necessity for more sophisticated solutions is underscored by the fact that conventional detection methods frequently depend on physical indicators or self-reporting, which may be intrusive or inaccurate. This research employs deep learning architectures such as EfficientNetB0, VGG16 and ResNet50 to identify driver weariness using the Driver Drowsiness Dataset (DDD). The DDD is a comprehensive dataset of facial photographs taken under diverse real-world driving scenarios, offering a solid basis for model training and assessment. The study workflow encompasses data augmentation, image preprocessing, facial micro-expression extraction and model training. According to the comparative analyses of models, EfficientNetB0 achieves high accuracy, ensuring its reliability in detecting driver fatigue states. Additionally, the model's lightweight architecture guarantees deployment in embedded systems and in-vehicle platforms. EfficientNetB0 strikes an optimal balance among accuracy, inference duration and model size, rendering it suitable for real-time applications. This study not only demonstrates the effectiveness of EfficientNetB0 in driver fatigue detection but also underscores the importance of model efficiency in safety-critical applications. These findings highlight its potential use for detecting driver weariness in real-world scenarios, providing essential technological assistance to improve road safety.