UAV-Based Inspection of Building Envelopes Through 3D Reconstruction and Autonomous Defect Detection
Loading...
Authors
Chodura, Nicholas
Date
2025-11-03
Type
thesis
Language
eng
Keyword
UAV , Building Inspection , 3D Reconstruction , Photogrammetry , Condition Assessment
Alternative Title
Abstract
Visual inspection remains a costly, hazardous, and subjective process for assessing building envelope and rooftop condition. This thesis investigates a framework for unmanned aerial vehicle (UAV)–based rooftop inspection, combining optimized photogrammetric reconstruction with deep learning–based defect segmentation. The first phase investigates how flight parameters such as image overlap, ground sampling distance, and supplementary flight paths influence the accuracy and visual quality of three-dimensional rooftop models. Controlled UAV field tests were evaluated against LiDAR ground truth to determine optimal flight parameters and path combinations. The second phase of the research applied UAV-based 3D reconstruction to cold-climate region applications of roof inspection. Photogrammetry and LiDAR were used in conjunction with point cloud volumetric differencing to estimate rooftop snow depth. This method was evaluated using manual snow probe measurements. Additionally, a method of combined thermal-RGB 3D reconstruction was developed that produces high-quality thermal models without the need for expensive software or complex image pair alignments. The third phase introduces the Residential Rooftop Defect Segmentation (RRD-Seg) dataset, a pixel-wise annotated benchmark of residential roof imagery. A DeepLabV3+ segmentation model with a ResNet-50 backbone achieved an F-score of 69.9\% and a mean Intersection-over-Union of 55.6\%. These results demonstrate the feasibility of semantic segmentation for UAV-based roof defect detection. The goal of these experiments is to apply modern tools and techniques to the traditional building inspection process. This was achieved through the optimization of flight parameters for UAV-based 3D reconstruction, snow depth estimation, thermal 3D reconstruction, and autonomous pixel-wise defect detection using a machine learning model. These tools can reduce the time, cost, and risk of rooftop inspections, while improving inspection accuracy and consistency.