Railroad Track Defect Detection Using YOLO Models: A Comparative Study

Detecting substructure defects is critical to ensuring track stability, preventing derailments, and avoiding costly slow orders. This study explores the use of surficial thermal and optical imagery captured by uncrewed aerial vehicle (UAV)-mounted sensors to detect subsurface issues such as ballast contamination, mud pumping, and differential settlement. We evaluate real-time object detection models You Only Look Once (YOLO)v5 through YOLOv10 across optical, thermal, and combined datasets. Results show that YOLOv10 achieves state-of-the-art performance, reaching 0.915 in precision and 0.918 in recall on optical imagery, and leading across metrics on thermal and combined modalities as well. Zero-shot cross-modal inference experiments, however, reveal limited transferability: models trained on thermal imagery generalize poorly to optical data, and vice versa, with mAP@50 below 0.25 in both directions. Overall, our results establish both a strong baseline for UAV-based defect detection and a roadmap for future research into more generalizable multimodal railway inspection models.