Visible and Invisible Dense 3D Reconstruction-based Robotic Nondestructive Inspection

ABSTRACT
Nondestructive structural health monitoring is critical for safety condition assessments of infrastructure. While conventional mono-view-based inspection has been proved to be successful, none has yet been able to answer what the metric information of the defect is? Can we detect and reconstruct subsurface objects? Also, vertical structures are challenged with accessibility. In this talk, I will present my research endeavor on automated surface and subsurface (seeing invisible) 3D inspection through robotics, computer vision, and machine learning. Firstly, for surface reconstruction, I will focus on the RGB-D camera depth calibration, completion, and semantic dense fusion towards generating defect-highlighted 3D models through our pioneering work on concrete structural flaw dataset creation and semantic recognition model. I will also discuss the design of our wall-climbing robot for automated inspection capable of maneuvering over the vertical surfaces. Then, I will present my recent work on data-driven-based subsurface objects detection and reconstruction using radio-wave signals collected from a ground-penetrating radar, producing the state-of-the-art straightforward 3D visualization of subsurface objects within a structure model. Finally, I will envision my future researches on critical problems in robotics and machine vision.