MSFD-Net: A Multiscale Spatial Feature Descriptor Network for Semantic Segmentation of Large-Scale Bridge Point Clouds

Digital Twins (DTs) provide a promising solution for bridge operation, thanks to their ability to mirror the physical conditions into a digital representation. At the core of the DTs is a geometric-semantic model. The modeling process for existing bridges, however, requires extensive manual effort. Given the high number of bridges in operation worldwide, there is an urgent need for automating this process. Available low-effort capturing methods, including laser-scanning and photogrammetry, generate raw point cloud data (PCD) that requires further processing to achieve a high-quality model. This paper focuses on the semantic segmentation of the PCD, which is the essential first step in an automated processing pipeline. A novel deep learning model, called multi-scale spatial feature descriptor network (MSFD-Net), is proposed for the semantic segmentation of PCD. The model is tested using the PCD of six bridges in Bavaria, Germany. The results show that MSFD-Net can automate semantic segmentation of bridges with mean accuracy (mAcc) of 98.29 % and mean intersection over union (mIoU) of 93.57 %.