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

## TL;DR

This paper introduces a method using laser scanning and deep learning to automatically detect geometric deviations in steel structures during construction.

## Contribution

The novel contribution is an automated geometric monitoring pipeline using 2D projections of 3D scans and a YOLOv8 model for steel cross-section detection.

## Key findings

- The methodology achieved 94% coverage of structural elements in real assemblies.
- Segmentation metrics reached 70.20% mAp@50-95 with synthetic data augmentation.
- 97% of segmentations were valid, enabling reliable geometric verification under EN standards.

## Abstract

Ensuring the quality and structural stability of industrial steel buildings requires precise geometric control during the execution stage, in accordance with assembly standards defined by EN 1090-2:2020. In this context, this work proposes a methodology that enables the automatic detection of geometric deviations by comparing the intended design with the actual as-built structure using a Terrestrial Laser Scanner. The integrated pipeline processes the 3D point cloud of the asset by projecting it into 2D images, on which a YOLOv8 segmentation model is trained to detect, classify and segment commercial steel cross-sections. Its application demonstrated improved identification and geometric representation of cross-sections, even in cases of incomplete or partially occluded geometries. To enhance generalisation, synthetic 3D data augmentation was applied, yielding promising results with segmentation metrics measured by mAp@50-95 reaching 70.20%. The methodology includes a systematic segmentation-based filtering step, followed by the computation of Oriented Bounding Boxes to quantify both positional and angular displacements. The effectiveness of the methodology was demonstrated in two field applications during the assembly of industrial steel structures. The results confirm the method’s effectiveness, achieving up to 94% of structural elements assessed in real assemblies, with 97% valid segmentations enabling reliable geometric verification under the standards.