Automating Dynamic Scaffolding Construction Progress Tracking Using Deep Learning-based Semantic Image Segmentation

Scaffolding plays a critical role in construction by providing safe access to elevated work areas, yet it is often overlooked in progress monitoring and omitted from reports and building information models (BIM) due to its temporary and dynamic nature. This study presents a deep learning–based semantic segmentation system for tracking scaffolding installation progress, distinguishing between scaffolding with and without safety sheets. A custom dataset of annotated site images was used to train the model, and performance was evaluated on both validation and test sets. The system achieved real-time processing speeds of 33.81 Frame Per Second (FPS) (validation) and 31.28 FPS (test), with mean Average Precision scores of 0.496 and 0.480, respectively. Class-specific results showed consistently higher accuracy for scaffolding with safety sheets, with peak Intersection over Union (IoU) values exceeding 93% in a case study time point. Two multi-time point construction case studies demonstrated the system’s robustness across varying site conditions. An automatic mask modification algorithm was applied to address missed detections, improving IoU by up to 3.10% in challenging scenarios. The calculated progress, based on segmented masks and predefined building boundaries, was compared with groundtruth measurements, confirming the system’s capability for quantitative progress tracking. Results indicate that scaffolding with safety sheets is more reliably detected, while detection of scaffolding without sheets remains more challenging. The proposed method offers a practical tool for reducing inspection workload, improving safety compliance, and enabling more comprehensive progress tracking in construction projects, particularly in scenarios where scaffolding installation is a major operational activity.