Superpoint Transformer–Based Bridge Component Recognition Using UAV-Mounted LiDAR and Synthetic Point Cloud Generation

Given the rapid rise in aging bridges, accurate and efficient inspection and maintenance have become increasingly critical. Point cloud data, rich in geometric details, plays a vital role in this process. Identifying semantic information for each bridge component is essential for effective analysis. This paper presents a novel Superpoint Transformer–Based method for automatically identifying bridge components from unmanned aerial vehicle (UAV)-mounted light detection and ranging (LiDAR) and simultaneous localization and mapping (SLAM)-acquired point cloud data. The method consists of two stages: acquisition and registration of point cloud data, followed by three-dimensional (3D) semantic segmentation using the Superpoint Transformer. Additionally, a new strategy simulates a virtual UAV equipped with LiDAR to generate synthetic data, reducing the domain gap between real and synthetic data. This simulation-based approach outperformed traditional sampling methods, significantly enhancing model performance. By fine-tuning the Superpoint Transformer for bridge component recognition, the proposed method achieved a mean intersection over union (mIoU) of 86.125%. This approach offers an effective solution for bridge component recognition, facilitating efficient inspection and maintenance of aging bridge infrastructure.