Semantic segmentation of bridge components and road infrastructure from mobile LiDAR data

Abstract

Emerging mobile LiDAR mapping systems exhibit great potential as an alternative for mapping urban environments. Such systems can acquire high-quality, dense point clouds that capture detailed information over an area of interest through efficient field surveys. However, automatically recognizing and semantically segmenting different components from the point clouds with efficiency and high accuracy remains a challenge. Towards this end, this study proposes a semantic segmentation framework to simultaneously classify bridge components and road infrastructure using mobile LiDAR point clouds while providing the following contributions: 1) a deep learning approach exploiting graph convolutions is adopted for point cloud semantic segmentation; 2) cross-labeling and transfer learning techniques are developed to reduce the need for manual annotation; and 3) geometric quality control strategies are proposed to refine the semantic segmentation results. The proposed framework is evaluated using data from two mobile mapping systems along an interstate highway with 27 highway bridges. With the help of the proposed cross-labeling and transfer learning strategies, the deep learning model achieves an overall accuracy of 84% using limited training data. Moreover, the effectiveness of the proposed framework is verified through test covering approximately 42 miles along the interstate highway, where substantial improvement after quality control can be observed.