earticle

영어

This paper proposes a real-time object detection system for railway safety using the Fast YOLO deep learning framework. Using a dataset of over 10,000 annotated images captured from onboard cameras, the system detects people, animals, and obstacles on railway tracks under various environmental conditions. Preprocessing methods including background subtraction and Gaussian modeling enhance detection robustness, achieving 15% relative improvement over baseline in low-light conditions. Experimental results demonstrate high precision (0.925 for people, 0.893 for animals, 0.878 for obstacles) with real-time processing at 38 FPS on NVIDIA GTX 1080 GPU. Our Fast YOLO implementation outperforms Faster R-CNN by 3.2- fold in speed while maintaining comparable accuracy (mAP of 0.84 vs 0.86) and surpasses SSD by 8.7% in detection accuracy. The system achieves 95.2% detection rate for stationary hazards and 91.6% for moving objects, with false positive rates below 2.3%. Field tests over 6 months demonstrated 99.7% uptime reliability and successful prevention of 12 potential incidents. The findings confirm Fast YOLO's effectiveness for automated railway safety monitoring, providing a practical solution for real-world deployment.