Nighttime Driving Fatigue Detection Based on Improved Zero-DCE

Fatigue is a primary factor in traffic accidents, and vision-based driver fatigue monitoring has become a crucial approach for traffic safety surveillance in intelligent driving. To address the decline in accuracy of traditional vision-based monitoring under nighttime conditions, this study proposes a driver fatigue detection method based on image recognition, specifically designed for low-light environments. Firstly, an improved Zero-DCE algorithm enhances facial image recognition accuracy at night. A composite attention network strengthens feature focus on key facial regions. Additionally, a gamma correction-based adaptive exposure loss function suppresses local overexposure and underexposure. A lightweight convolutional structure replaces conventional modules, reducing computational costs while maintaining detection accuracy. Secondly, a fatigue detection model based on facial features is established. The Dlib-HOG algorithm extracts facial features, and a Naïve Bayes classifier identifies fatigue states. Experimental results demonstrate that the proposed algorithm improves image processing quality on public datasets and increases detection accuracy on a self-constructed nighttime driving fatigue dataset. The proposed method supports nighttime human-machine collaborative safety in intelligent driving systems, promoting advancements in low-light humanmachine interaction safety.