Component‐level point cloud completion of bridge structures using deep learning
Gen Matono, Mayuko Nishio
Abstract
Point cloud of existing bridges provides important applications in their maintenance and management, such as to the three-dimensional (3D) model creation. However, point cloud data acquired in actual bridges are caused missing parts due to occlusions and limitations in sensor placements. This study proposes a learning method to realize the point cloud completion of such structure: the component-wise learning combining the initial weight transfer, to overcome the difficulty particularly found in the bridge structures, where a whole structure consists of multiple and various components. The learning method was developed and verified using point cloud data acquired in an actual concrete bridge based on the point cloud completion performance of three significant deep learning models. The effectiveness and applicability of the proposed method were shown in that it improved performances in component level in applying it to the bridge point cloud completion by the multiple deep learning models, respectively.