부산대학교 도서관

Bridge inspections are typically expensive and time-consuming, especially in regards of the inspection of difficult-to-reach areas. In recent years, unmanned aerial systems (UASs) have gained attention due to their flexible data acquisition. However, UAS inspections generate large quantities of image and video data, which are currently analysed manually. Additionally, identified damages are currently not assessed accurately in their geometric characteristics and location. In this paper, we propose a time-effective framework for a UAS-based bridge inspection methodology that combines 3D information from photogrammetry and machine learning based object detection to allow direct measurements in the images. Concretely, we propose the use of a two-step flight planning to accurately reconstruct the bridge using limited manual effort. Second, we detect frequently occurring damages such as exposed rebars and concrete spalling on the inspection imagery. Finally, we use the spatial location of the imagery to significantly improve the detection results and geolocate them. We evaluate our proposed framework on a decommissioned concrete bridge. The trained YOLOv8 models prove capable of transfer learning on both our own data and online benchmarks. The photogrammetric reconstruction also proves to be sufficiently reliable. Overall, these are the first steps in automating routine bridge inspections and provide crucial evidence to continue developing the method.