YOLO-FDCL: Improved YOLOv8 for Driver Fatigue Detection in Complex Lighting Conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety

Background:

• Driver fatigue is a major cause of road accidents.
• Complex lighting conditions in vehicle cabins reduce the accuracy of existing fatigue detection systems.
• Accurate fatigue identification is critical for enhancing road traffic safety.

Purpose of the Study:

• To develop a novel algorithm, YOLO-FDCL, for robust driver fatigue detection under complex lighting conditions.
• To improve the accuracy and performance of fatigue detection systems in challenging driving environments.
• To address the limitations of current methods in varying light scenarios within vehicle cabins.

Main Methods:

• Introduced MobileNetV4 into the backbone network for enhanced feature extraction and reduced model size.
• Incorporated RepFPN with structural re-parameterization in the neck for improved multi-scale feature fusion.
• Evaluated the algorithm on the YAWDD dataset and a self-developed Complex Lighting Driving Fatigue Dataset (CLDFD).

Main Results:

• On the YAWDD dataset, YOLO-FDCL improved precision to 98.8% and recall to 97.5% compared to YOLOv8-S.
• Achieved higher mAP@0.5 (98.8%) and mAP@0.5:0.95 (94.2%) on the YAWDD dataset.
• Demonstrated significant performance gains on the CLDFD, with precision and recall increases of 2.8% and 2.2%, respectively.

Conclusions:

• YOLO-FDCL offers significant advancements in driver fatigue detection, particularly under complex lighting conditions.
• The algorithm enhances feature extraction and multi-scale fusion, leading to superior detection performance.
• This work contributes to improved road safety by providing a more accurate and reliable method for identifying driver fatigue.