Railway Foreign Object Intrusion Detection Inspired by Star-operation and Coordinate Attention

Railway construction has advanced, reducing casualties from accidents. However, the number remains high. Identifying and avoiding foreign objects that invade railway safety boundaries is an urgent issue. Omissions and misdetections are common problems. Deep learning-based railway foreign object intrusion detection algorithms can replace manual inspections and traditional methods. These algorithms enable accurate division of intrusion limits and precise detection of foreign objects. However, existing methods face challenges. These include insufficient railway datasets, inaccurate safety limit demarcation, excessive model parameters, long detection times, and issues with missing and false detections. To address these problems, this paper proposes an adaptive improvement to YOLOv8n. The goal is to enhance the accuracy and efficiency of railway foreign object detection. Experimental results show the improved model achieves $55.1 \%$ mAP50 on the OSDaR23 dataset. The inference speed reaches 46.2 FPS. The recall rate increases by $3.1 \%$.