This study proposes a fatigue driving detection model based on image enhancement and facial state recognition to address the issues of low accuracy, large model size, and reduced performance in low-light environments found in existing models. The YOLOv5s model was enhanced for face detection and key point localization under low illumination, incorporating a self-calibrated illumination module to enhance low-illumination images. The downsampling layer was replaced with the StemBlock module to improve feature expression capability, and the backbone network was replaced with ShuffleNetv2 to reduce the model's parameters and computational complexity. Replacing the C3 module with the cbam inverted bottleneck C3 (CIBC3) reduced noise interference in detection and enhanced the model's global perception ability. Furthermore, the wing loss function was added to the total loss function for facial keypoint regression. A fatigue state recognition network was employed to determine the opening and closing statuses of the eyes and mouth located by the face detection model, and evaluation indicators were used to determine the fatigue state. The experimental results obtained on the DARK FACE dataset demonstrate that, compared to the benchmark YOLOv5s model, the improved model reduced parameters and computational complexity by 62.12% and 63.41%, respectively, and improved accuracy by 2.38 percentage points. The proposed fatigue driving detection model achieved accuracies of 96.07% and 94.50% on the YawDD normal lighting and self-built low-lighting datasets, respectively, outperforming other models. The processing time per image was only 27 ms, demonstrating that the proposed model not only ensures detection accuracy in normal and low-light environments but also meets real-time requirements, making it suitable for deployment on edge computing equipment with limited computing power.