Structure-from-motion algorithm fused with real-time kinematic location information and improved YOLOv11 network for bridge tower damage recognition method

The rapid advancement of Unmanned Aerial Vehicles (UAVs) and Real-Time Kinematic (RTK) technologies has created new opportunities and challenges for bridge tower inspection. This study proposes a damage identification method that integrates RTK-enhanced Structure-from-Motion (SfM) with an optimised neural network to improve structural assessment of bridge towers. First, a two-dimensional virtual flight path is generated using RTK positioning and virtual projection to guide UAV-based image acquisition. The RTK positional data are then incorporated into the SfM pipeline via the Umeyama and Bundle Adjustment (BA) algorithms to enhance 3D reconstruction accuracy. The four flight parameters are systematically optimised through 16 cross-experiments and eight factor quality analyses. The test bridge model, which uses reconfigured, optimised flight parameters, achieves millimeter-scale accuracy and produces millimeter-resolution images, improving YOLOv11 neural network recognition. Experimental results show that the median position error of the projection centre point of the experimental bridge image is 11.07 mm, and the average Ground Sampling Distance (GSD) resolution is 1.02 mm/pixel. Compared with the original YOLOv11 model, the improved model achieves increases of 3.9%, 1.3%, 1.0%, and 2.5% in accuracy, recall, mAP@50, and mAP50-95, respectively. Compared with the traditional bridge-tower detection method, the proposed UAV-based non-contact approach offers higher accuracy and efficiency.