Semantic segmentation of bridge components based on hierarchical point cloud model

Abstract

Geometric information such as the volumetric dimensions and type of a bridge can be retrieved by means of a terrestrial laser scanner, and the point cloud (PC) data thus obtained can be used to build 3-dimensional (3D) objects such as bridge components in the Building Information Modeling (BIM) framework. For modeling of PC data, PointNet (Qi et al., 2018) and a graph-based convolutional neural network (GCNN) model known as dynamic GCNN (DGCNN; Wang et al., 2019) have been widely employed. In this study, a graph-based hierarchical DGCNN (HGCNN) model is proposed for more accurate and realistic representation of railway bridges having electric poles. The model obtains detailed local features by incrementally considering neighboring points while the total number of neighbors remains the same. Field application reveals that the proposed HGCNN model can represent tall components such as electric poles more precisely, while the overall accuracy of semantic segmentation is dominated by bulky components such as decks, so that the differences among the models (PointNet, DGCNN and HGCNN) are not significant. Specifically, the recall and intersection over union (IoU) rate of the electric pole were improved by about 3% when using the proposed model. A few parametric studies were also performed, and it was demonstrated that the proposed model with its expanded local features provides more precise information near the object boundaries.