In recent years in Japan, infrastructures constructed during the period of rapid economic growth have been rapidly aging and their management has become an urgent issue. Because construction drawings and photographs are used for bridge management in Japan, inspectors cannot record 3D-information of damages precisely and accumulate damage records.
To overcome these problems, utilization of 3D models in bridge management has been catching attention. In order to utilize 3D models for bridge management, it is important to establish the method to construct them. It takes a lot of work to manually create an accurate 3D model from construction drawings, and the required accuracy varies depending on the application. In addition, it is also not easy to create 3D models of a huge number of existing structures, and there are existing structures for which no drawings remain.
Under such circumstances, there has been lots of research in which they construct 3D models using various data, including construction drawings and point cloud data. However, there are some problems like the difficulty of applying their method to the vast number of existing bridges. In this research, the authors propose a 3D model construction method using neural network which learned generic shapes of bridge parts. In particular, the authors propose a method to extract latent information from point cloud data by introducing structured knowledge into network calculations and use it to construct 3D models. Furthermore, the authors validate the effectiveness of the method using point cloud data of a simple bridge.