Abstract: To more accurately extract the features of train operation track and generate precise boundaries for foreign object intrusion detection, this paper proposed a model-based and rule-based method for extracting features of train operation track. The paper improved the YOLO (You Only Look Once) v8n algorithm by incorporating a variable kernel convolution module into the backbone network to increase the receptive field, establish long-range dependencies, better obtain the features of long strip-shaped large targets in the orbit, and designed a multi feature fusion attention module that integrated features from shallow layer, deep layer, and variable kernel convolution modules, performed attention weighting to highlight effective track information, and obtained accurate track masks and position information. The paper also conducted a preliminary screening of the track segmentation results and used threshold and region of interest screening to avoid track interference in non train operational areas, extracted features of train operation track based on the preliminary screening results and prior information. The experimental results show that compared to the original YOLOv8n algorithm, the Box mAP@0.5 Value and MaskmAP@0.5 increase by 0.9% and 1.5% respectively by using the rail YOLOv8n improved algorithm. The algorithm achieved ideal results in scenarios such as straight track, bends, and turnouts.