Inspecting infrastructure with drones and deep learning
- Article
- Computer Vision
Roads, bridges, and buildings wear out under weather, load, and age, so they need regular inspection. The usual practice is non-destructive testing with laser scanners and cameras, and much of it still comes down to a person looking. Visual inspection is flexible but slow, subjective, and sometimes dangerous, since an inspector has to get close to the defect. Unmanned aerial and ground vehicles were introduced to take the person out of the risky part, and computer vision is being added so the images they bring back can be checked automatically.
This post, published by UP Data Science Society on May 3, 2024, summarizes a 2024 paper in Automation in Construction by Benyun Zhao, Xunkuai Zhou, Guidong Yang, Junjie Wen, Jihan Zhang, Jia Dou, Guang Li, Xi Chen, and Ben M. Chen.
What the researchers built
The team released CUBIT-Det, an open dataset of more than 5,500 high-resolution images taken with professional equipment across three scenarios: buildings, pavements, and bridges. It covers three defect types chosen from inspection standards and guidelines: cracks, spalling, and moisture. A self-built unmanned platform with custom collection routines gathered the images, which cut down manual work. After collection the team removed invalid photos, blurred video frames, and unfocused shots, then annotated each image in LabelImg with the file path, dimensions, defect class, and bounding-box coordinates.
To check that the dataset is useful, they benchmarked more than 25 detection models on it, measured how each defect class affected accuracy, and ran a model trained on CUBIT-Det against a real industrial building in Hong Kong.
What they found
Buildings make up 65 percent of the images, pavements 29 percent, and bridges 6 percent. Cracks are the most common defect, followed by spalling. Most annotated objects are large, more than 10 percent of the image, which means the dataset is best at defects that are visible at a glance rather than hairline ones.
The best real-time detector reached 82.9 percent mAP at an IoU threshold of 0.5 and 55.9 percent on the stricter 0.5 to 0.95 range. The fastest model processed a 1024 by 1024 image in 15.9 milliseconds. The Hong Kong field test produced good inspection results, which the authors take as evidence that models trained on the dataset transfer to real structures.
Why it matters
Compared with existing defect datasets, CUBIT-Det spans more scenarios and uses higher-resolution images. The authors say the dataset still needs to grow and cover more defect types, but the combination of unmanned platforms and automated detection already shortens inspections and keeps people off the scaffolding.
Source: Zhao, B., Zhou, X., Yang, G., Wen, J., Zhang, J., Dou, J., Li, G., Chen, X., and Chen, B. M. (2024). High-resolution infrastructure defect detection dataset sourced by unmanned systems and validated with deep learning. Automation in Construction, 163, 105405. doi:10.1016/j.autcon.2024.105405