Automated Defect Quantification in Concrete Bridges Using Robotics and Deep Learning

摘要

Abstract This work presents a process for automated end-to-end inspection of area defects—specifically spalls and delaminations—in RC bridges. The process uses a mobile robotic platform to collect three-dimensional (3D) spatial data via lidar, and visual defect data via visible and infrared spectrum cameras. A convolutional neural network (CNN) is implemented to automatically make pixelwise predictions about the presence of defects in the images. Simultaneous localization and mapping is employed to fuse 3D lidar data with labeled images to generate a colorized and semantically labeled 3D map of a structure. Using this 3D map, a procedure was developed to automatically quantify the delamination and spall areas. This procedure was validated on a concrete bridge, and results showed that the automated defect quantification inspection process is feasible to detect and quantify both spalls and delaminations. Error rates in the physical scale of defect areas when using ground truth–labeled versus CNN-labeled images were similar to the corresponding pixel error rates between ground truth and CNN labels in the image domain. The central contribution of this paper is an objective, repeatable, and reference-free approach to area defect quantification from images collected in unstructured environments using a mobile platform.