A bridge point cloud databank for digital bridge understanding

Abstract

Despite progress in automated bridge point cloud segmentation based on deep learning, challenges persist. For instance, the absence of a public point cloud dataset specifically designed for bridge instances, and the existing bridge point cloud datasets display a lack of diversity in bridge types and inconsistency in component labeling. These factors may hinder the further improvement of accuracy in bridge point cloud segmentation. In this paper, a universal multi-type bridge point cloud databank, named BrPCD, consisting of a total of 98 point cloud data (PCD; 10 of them are obtained from scanning, and the rest is obtained by data augmentation) from small to long-span bridges, is established. Additionally, a method for augmenting bridge PCD is proposed, significantly enriching the spatial feature information of bridges within the dataset. Furthermore, based on the introduced data annotation rules, a uniform categorization of semantic labels for bridge components is implemented, enhancing the applicability of our dataset across various semantic segmentation tasks for different types of bridges. A benchmark testing was conducted on the BrPCD using the PointNet model. The segmentation results indicate that the parameters learned through the BrPCD enable accurate segmentation at the level of various types of bridge components. In other words, the BrPCD can function as a universal dataset, applicable for testing various networks aimed at bridge point cloud segmentation.