Building damage inspection method using UAV-based data acquisition and deep learning-based crack detection

Abstract

Detecting cracks early benefits building maintenance by assessing structural safety, which in turn helps prevent potential severe damage and collapse, given that cracks in concrete surfaces often reflect underlying structural damage. However, the conventional method by human hands is time-consuming, inconvenient, and high risk for inspectors. In this present study, an improved framework for inspecting building surface cracks, which integrates digital innovations of Unmanned Aerial Vehicle (UAV) and deep learning technologies with wide-area coverage, high efficiency, and less intervention, is established. The feasibility of the proposed approach is demonstrated by conducting an experimental test on an in-service office building. The results show that not only can we achieve a prediction accuracy of over 97% on the validation dataset, but also that increasing the number and variety of images in the training dataset positively impacts the ability to detect concrete cracks. However, this improvement might not be as notable once the model has already learned sufficient features of concrete cracks. Additionally, a 3D model was created to virtually showcase the detection results. This opens up new possibilities for conducting building damage inspections by integrating these results into a virtual 3D space, which enhances overall structural health management and offers new insights for improving detection performance. Challenges and future directions to improve the effectiveness and address potential improvement approaches of the proposed framework in practice are also suggested.