Date of this Version
Barnes, P. B. (2021), "Detecting and Evaluating Cracks on Aging Concrete Members using Deep Convolutional Neural Networks," Masters Thesis, Department of Civil and Environmental Engineering, University of Nebraska-Lincoln, Lincoln, NE, 205 pp.
Cracks in concrete structures are evaluated through a timely and subjective manual inspection. The location of cracks is often recorded in an inspection report where some cracks are measured. Although measurements or locations may not be necessary for all cracks observed in concrete members, if quantitative data can be gathered in an autonomous way, allowing measurement data to be used in tracking changes in spatial and temporal scales, this quantitative data can provide useful information not yet captured in the manual inspection process. This thesis aims to construct an image-based crack detection and evaluation pipeline that can assist health monitoring of aging concrete structures, by providing crack locations and measured crack properties for the entire structural member. Over 16,000 images of aging concrete bridge deck were collected from cameras attached on an unmanned aerial vehicle, machine vision cameras attached on a ground vehicle, and other literature. Mask and Region based Convolutional Neural Network (Mask R-CNN) was utilized to train 256 by 256-pixel patches of collected images using three distinct training strategies to detect and segment concrete cracks on bridge decks. Resulting crack masks were translated into binary data (crack or non-crack pixels) and skeletons of the mask were created where the Euclidean distance from the center of the skeleton to the edge of the mask were measured. This allowed to calculate the relative crack width, length, and orientation of each detected crack. Relative crack properties were transformed into real-world unites using the ground sampling distance of the host image. Image patches were then compiled to construct a crack map of the entire structural member.
A case study was conducted on the deck and pier of an aging concrete bridge to test the robustness of the proposed data pipeline. The study yielded that the model was able to successfully detect cracks with an average width of 0.020 inches and were able to make accurate measurements of crack widths that are larger than 0.080 inches. In order to improve the measurements for smaller crack widths, the ground sampling distance needs to be to the scale of the crack width in interest. The image-based data pipeline developed in this study demonstrates potential for the application in autonomous inspections of concrete members. In addition, the data pipeline can be used as a reference framework to provide an example on how computer-vision based data analytics can provide useful information for structural inspections of aging concrete members.
Advisor: Chungwook Sim