Learning monocular depth estimation for defect measurement from civil RGB-D dataset

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Civil infrastructure is aging, increasing the need for effective structural inspections.
• Current inspection methods face challenges due to labor intensity and the difficulty of quantifying defects from single images.
• Deep learning for image-based defect detection and segmentation shows promise but lacks practical 3D measurement capabilities.

Purpose of the Study:

• To develop a method for recovering 3D scene geometry from single images for infrastructure defect quantification.
• To address the lack of specialized datasets for training and evaluating spatial computer vision models in civil engineering.
• To create a LiDAR-based RGB-D dataset for the civil engineering domain.

Main Methods:

• Utilized deep learning-based monocular depth estimation to recover 3D geometry from single images.
• Developed and curated a novel, in situ Light Detection and Ranging (LiDAR) RGB-D dataset specifically for civil engineering applications.
• Evaluated the proposed monocular depth estimation approach for quantifying defects in civil infrastructure using the new dataset.

Main Results:

• Successfully demonstrated the recovery of 3D scene geometry from single images.
• Established a valuable, publicly available LiDAR-based RGB-D dataset for civil engineering research.
• Validated the practical application of monocular depth estimation for defect quantification in real-world infrastructure scenarios.

Conclusions:

• Monocular depth estimation offers a viable solution for 3D reconstruction and defect quantification in civil infrastructure from single images.
• The developed dataset is crucial for advancing spatial computer vision techniques in the civil engineering field.
• This research facilitates more efficient, accurate, and potentially cost-effective infrastructure inspection processes.