Real-Time Pilot Fatigue Monitoring: A YOLO-Based Deep Learning Approach

Abstract

Pilot fatigue constitutes one of the most critical factors compromising flight safety. Currently, the majority of deep learning algorithms focus primarily on improving detection accuracy, while neglecting the critical requirements of real-time performance and lightweight design essential for fatigue driving detection systems. Therefore, this paper proposes an enhanced approach based on the YOLOv8 object detection network. The improved network architecture, integrated with a fatigue driving assessment mechanism, significantly accelerates detection speed while maintaining robust performance. First, a lightweight VanillaNet architecture is adopted to replace the traditional backbone network of YOLOv8, effectively reducing both the computational complexity and parameter count of the model. Subsequently, the original bounding box regression loss function is replaced with a Wise-IoU loss function incorporating a dynamic non-monotonic focusing mechanism, which enhanced the model’s precision. Experimental results demonstrate that the VanillaNet architecture reduces the model’s parameter count by 2 million, while the enhanced Wise-IoU loss function improves detection accuracy by 6.1%, collectively achieving real-time performance and lightweight requirements for fatigue driving detection.