Feasibility Analysis of RT-DETR in the Context of Foreign Object Intrusion at Railway Boundaries

Detection of railway perimeter intrusion is of great significance for railway safety operation, but existing object detection algorithms cannot fully meet the needs of actual railway perimeter foreign object intrusion detection. This article is based on a self-built dataset for railway perimeter intrusion problem. Under the same computer environment, three target recognition algorithms, YOLOv5, YOLOv8, and RT-DETR, are trained and evaluated for their performance indicators. The experimental results showed that in terms of accuracy, RT-DETR achieved better performance than YOLOv5 and YOLOv8 with fewer training rounds and fewer data augmentation strategies. In terms of FPS, RT-DETR’s FPS value after removing NMS was much higher than that of the YOLO series. This comparison verified the feasibility of RT-DETR in the field of railway perimeter intrusion and can be applied to the actual use needs of railway perimeter intrusion.