Buildings, Vol. 16, Pages 140: Steel and Concrete Segmentation in Construction Sites Using Data Fusion: A Literature Review

Buildings doi: 10.3390/buildings16010140

Authors:
Enrique Martín Luna Gutiérrez
Osslan Osiris Vergara Villegas
Vianey Guadalupe Cruz Sánchez
Humberto de Jesús Ochoa Domínguez
Juan Humberto Sossa Azuela

Construction progress monitoring remains predominantly manual, labor-intensive, and reliant on subjective human interpretation. Human dependence often leads to redundant or unreliable information, resulting in scheduling delays and increased costs. Advances in drones, point cloud generation, and multisensor data acquisition have expanded access to high-resolution as-built data. However, transforming data into reliable automated indicators of progress poses a challenge. A limitation is the lack of robust material-level segmentation, particularly for structural materials such as concrete and steel. Concrete and steel are crucial for verifying progress, ensuring quality, and facilitating construction management. Most studies in point cloud segmentation focus on object- or scene-level classification and primarily use geometric features, which limit their ability to distinguish materials with similar geometries but differing physical properties. A consolidated and systematic understanding of the performance of multispectral and multimodal segmentation methods for material-specific classification in construction environments remains unavailable. The systematic review addresses the existing gap by synthesizing and analyzing literature published from 2020 to 2025. The review focuses on segmentation methodologies, multispectral and multimodal data sources, performance metrics, dataset limitations, and documented challenges. Additionally, the review identifies research directions to facilitate automated progress monitoring of construction and to enhance digital twin frameworks. The review indicates strong quantitative performance, with multispectral and multimodal segmentation approaches achieving accuracies of 93–97% when integrating spectral information into point cloud or image-based pipelines. Large-scale environments benefit from combined LiDAR and high-resolution imagery approaches, which achieve classification quality metrics of 85–90%, thereby demonstrating robustness under complex acquisition conditions. Automated inspection workflows reduce inspection time from 24 h to less than 2 h and yield cost reductions of more than 50% compared to conventional methods. Additionally, deep-learning-based defect detection achieves inference times of 5–6 s per structural element, with reported accuracies of around 97%. The findings confirm productivity gains for construction monitoring.

Source link

Enrique Martín Luna Gutiérrez www.mdpi.com