Vision and Deep Learning-Based Algorithms to Detect and Quantify Cracks on Concrete Surfaces from UAV Videos

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Manual inspection of civil infrastructure post-disaster is time-consuming and error-prone.
• Accessing all areas of large structures for damage assessment is challenging.
• Unmanned Aerial Vehicles (UAVs) offer a potential solution for timely infrastructure health assessment.

Purpose of the Study:

• To propose an AI-driven framework for automated detection and quantification of cracks in civil infrastructure.
• To enhance the speed and accuracy of structural integrity assessments after natural disasters.
• To provide a robust method for evaluating the global stability of structures.

Main Methods:

• Utilizing computer vision and deep learning algorithms for crack detection from images.
• Implementing image segmentation with a U-Net deep neural network for pixel-level crack classification.
• Applying morphological operations to quantify crack geometry (length, width, area, orientation).

Main Results:

• Successful detection, quantification, and localization of concrete cracks using the proposed framework.
• Validation through a laboratory experiment involving a UAV-mounted camera and a concrete beam.
• Demonstrated efficacy in providing dense measurements of individual crack geometries.

Conclusions:

• The proposed AI and UAV-based framework offers a viable and efficient alternative to manual infrastructure inspection.
• Automated crack analysis improves the accuracy and timeliness of structural health prognosis.
• This technology can significantly aid in assessing a structure's ability to withstand service loads post-disaster.