Tsai, Cheng-YunCheng-YunTsaiJACOB JE-CHIAN LINLiang, Ci-JyunCi-JyunLiang2025-11-252025-11-2520259780645832228https://www.scopus.com/record/display.uri?eid=2-s2.0-105016567677&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/734130Crew-level productivity analysis plays a crucial role in construction site management, as it provides a macro-level understanding of workforce performance. While traditional productivity measurement methods, such as work sampling, group timing technique, and five-minute rating, offer valuable insights, they rely heavily on manual observation, making them labor-intensive and prone to inaccuracy. In this study, we propose a deep learning-based framework for automated crew-level identification. The framework employs a graph-based approach that integrates visual feature similarity and spatial proximity of workers, combined with clustering algorithms, to detect and analyze worker groups. The proposed method is validated on a construction site dataset collected from a rebar installation task. Experimental results demonstrate the framework’s effectiveness, achieving high accuracy in group detection with robust performance across various evaluation metrics. This work highlights the potential of automated systems to enhance construction site management by reducing reliance on manual observation and providing real-time insights into crew-level productivity.trueClusteringGroup DetectionHuman Activity RecognitionImage Understandingconference paper10.22260/ISARC2025/01452-s2.0-105016567677