Aerial crack detection in building structures using convolutional neural networks

Evaluating cracks is a crucial step in preserving concrete structures. Manual visual observation of the surface is the common method for inspecting concrete cracks, but it is inherently subjective and depends on the inspectors’ expertise. Moreover, it is inefficient, costly, and often hazardous when building structures are hard to reach. In structural health monitoring, unmanned aerial vehicles (UVAs) outfitted with advanced sensing and imaging technologies offer the capacity to detect and assess defects, such as corrosion and cracking, in buildings and infrastructure, thereby facilitating proactive maintenance and early hazard identification. This study presents the design and implementation of a UAV system integrating both sonar and visual sensors, aimed at detecting and classifying building cracks through deep learning techniques. The combined use of heterogeneous sensors represents a novel approach that capitalizes on their respective strengths while mitigating individual drawbacks. The detection of crack presence and depth is accomplished using sonar sensors. Visual capture is then initiated triggering the cameras only after a crack is detected by the sonar to minimize processing overhead during image analysis. The images of the cracks acquired by the vision sensors are processed by a customized convolutional neural network model which shows 99% accuracy in detecting and identifying the cracks.