Bridge Point Cloud Completion Using Deep Learning Obtained in Actual Bridge Structures

Abstract

Point cloud, which can be obtained by optical measurement, is recently recognized to be useful in SHM for the maintenance and management of existing civil structures. However, there are some issues in measuring point cloud of large structures such as bridges. First, multiple measurements from different locations are required to reconstruct the point cloud of a whole structure. Second, the parts, where are interrupted by trees or other objects cannot be measured well even in uses of cameras or 3D scanners. Therefore, the point clouds acquired in actual structures cannot prevent missing parts or lack of details of structural configurations. This study aims to show applicability of deep learning for reproducing the partial point cloud obtained from measurements in actual structures into a completed point cloud. The experiment results show that even with limited data, transfer of training weight and component-wise completion can yield greater accuracy compared to completion of the entire bridge.