3D Dense Reconstruction for Structural Defect Quantification

Abstract

Recent advancements in vision-based visual inspection enable the identification, localization, and quantification of damage on structures. However, existing damage quantification methods are limited to measuring one- or two-dimensional attributes such as length or area, which is insufficient for certain damage types such as spalling that require depth in addition to in-plane measurements, as outlined in inspection manuals. To address this limitation, we propose utilizing image-based dense 3D reconstruction to perform full 3D quantifications to assess damages for concrete structure inspections. The proposed method is applied to quantify spalling damage in 3D to compute volumetric loss and maximum depth of the damage in line with bridge inspection manuals. Our approach involves using a convolutional neural network-based interactive segmentation algorithm to accurately segment spalling boundaries from images. Structure-from-motion and multiview stereo algorithms are then applied to generate a detailed 3D point cloud reconstruction of the spalling using multiple images. From this point cloud, a 3D mesh representation of the spalling is created for precise quantification. To validate our proposed technique, we conducted laboratory and field experiments to capture images and interactively segment the damage. The results demonstrate the effectiveness and reliability of our approach for 3D damage quantification in structure inspections.