Geometric Monitoring of Steel Structures Using Terrestrial Laser Scanning and Deep Learning

Area of Science:

• Structural Engineering
• Computer Vision
• Metrology

Background:

• Geometric precision is critical for industrial steel building quality and stability, adhering to EN 1090-2:2020 standards.
• Manual geometric verification is labor-intensive and prone to errors.

Purpose of the Study:

• To develop an automated methodology for detecting geometric deviations in steel structures.
• To compare as-built structures with design models using Terrestrial Laser Scanning (TLS) and AI.

Main Methods:

• A pipeline processing 3D point clouds from TLS by projecting them into 2D images.
• Training a YOLOv8 segmentation model for steel cross-section detection, classification, and segmentation.
• Applying synthetic data augmentation for improved model generalization.
• Quantifying positional and angular displacements using Oriented Bounding Boxes.

Main Results:

• Achieved 70.20% mAp@50-95 in segmentation metrics with data augmentation.
• Successfully assessed up to 94% of structural elements in field applications.
• Obtained 97% valid segmentations for reliable geometric verification.

Conclusions:

• The proposed methodology effectively detects geometric deviations in industrial steel buildings.
• Automated detection ensures compliance with assembly standards and enhances structural integrity.
• The approach is robust, handling incomplete or occluded geometries effectively.