@

초록

Semantic segmentation of bridge components from UAV-mounted LiDAR point clouds is essential for automated structural inspection and digital twin generation, yet remains challenging due to non-uniform point density and occlusions. This paper introduces a large-scale real-world UAV LiDAR bridge dataset with fine-grained component-level annotations. To establish a strong benchmark, we propose a lightweight UAV-oriented segmentation framework based on RandLA-Net, incorporating a hybrid neighborhood feature aggregation strategy and a class-weighted focal loss to address sparse regions and class imbalance. Experiments on three unseen bridge scenes show that the proposed framework significantly improves segmentation performance, achieving an average overall accuracy of 96.93% and a mean Intersection over Union of 93.74%, outperforming the baseline RandLA-Net under realistic background-inclusive settings. In addition, notable accuracy gains are observed across individual structural components, including relatively underrepresented elements such as pier caps, demonstrating the effectiveness of the proposed dataset and framework for UAV-based bridge segmentation.

키워드