Bogie-YOLO: A key component detection model for high-speed train bogies

Abstract The performance of high-speed trains heavily relies on the bogie, a critical component supporting the body and steering. Traditional methods for inspecting bogie components, which rely on manual identification, are not just time-consuming but also expensive. Furthermore, deep learning models often face difficulties in handling dimly lit, intricate backgrounds and detecting numerous small objects. To solve these challenges, we propose Bogie-YOLO, a targeted detection model tailored for bogie photography scenes. Bogie-YOLO utilizes the YOLOv5 framework and introduces the Coordinate and Squeeze-and-Excitation Attention Module (CSAM), improving positional information and channel relationships. Additionally, to expedite convergence, we adopt EIoU as the bounding box loss. Experimental results demonstrate an average Precision of 87.9% and an average mAP0.5 of 83.2% on the dataset of key components of train bogies.