Production-based progress monitoring of rebar tying using few-shot learning and kernel density

Abstract

Real-time monitoring of construction progress is crucial for the success of projects. Rebar-related activities are costly, labor-intensive and vital for the structural integrity of buildings. Traditional vision-based monitoring methods focus on appearance, often misjudging completion status by ignoring worker activities after visual completion. To address this problem, this paper proposes a production-based framework for monitoring the progress of rebar tying through worker activity recognition. A novel few-shot learning-based method is employed for worker activity recognition, including zero-shot worker region proposals and few-shot worker activity classification, thereby avoiding the high costs associated with traditional supervised learning data labeling. Furthermore, a progress-production model based on kernel density is proposed to factor in the impact of worker activities on construction progress and assess real-time progress in the rebar tying process through cumulative density. Two publicly available datasets, along with a custom dataset, are used to validate the method's effectiveness and generalizability, demonstrating superior performance over traditional zero-shot methods. Furthermore, real-world projects are utilized to assess the framework's applicability, consistently achieving a progress monitoring error of less than 5 % across multiple scenarios. This research contributes to the digitization of construction sites and provides a paradigm for inferring construction progress from production activities.

Keywords

• Construction progress monitoring
• Rebar tying
• Few-shot learning
• Worker activity recognition
• Computer vision