3D Scene Reconstruction for Robotic Bridge Inspection

Abstract

This paper provides a comparative study of two methods for reconstructing three-dimensional (3D) scenes from monocular two-dimensional (2D) images with respect to their applicability to robotic civil infrastructure inspection. The two methods considered are dense structure from motion (DSfM) and image mosaicing (IM). The primary metrics used in the comparison include: (1) the accuracy of the reconstructions for inspection measurement purposes, (2) the nature and prevalence of reconstruction artifacts, and (3) resolution demands for adequate reconstruction accuracy. A series of small-scale model reconstructions using both approaches were developed and compared. Findings from the small-scale tests were then applied to a full-scale field test of the robotic reconstruction system. Both methods generally produced geometrically accurate reconstructions, but the DSfM method was found to be more adaptable to complex field scenarios. There were also key differences in the frequency and nature of artifacts introduced by each method.