Publications / 2020 Proceedings of the 37th ISARC, Kitakyushu, Japan

Development of a Workers' Behavior Estimation System Using Sensing Data and Machine Learning

Rikuto Tanaka, Nobuyoshi Yabuki and Tomohiro Fukuda
Pages 878-885 (2020 Proceedings of the 37th ISARC, Kitakyushu, Japan, ISBN 978-952-94-3634-7, ISSN 2413-5844)

Accurate information on workers' behavior is important for safety and productivity management on construction sites. In recent years, some methods for estimating construction workers' behavior using sensing data have been proposed to collect the data based on scientific evidence. Due to the limitations of previously proposed methods that usually relied on expensive devices such as motion capture systems, the huge amount of investment on the system installation and human resource costs are required. This paper proposes a method for estimating workers' posture with Long Short-Term Memory (LSTM) by using terminals that have already been introduced to construction sites, taking into consideration the operational cost and issues in the previous studies. Moreover, we also propose and evaluate a data augmentation method for utilizing limited training data sets. The experiment results for a reinforcing bar worker indicated that the proposed method could estimate not only the forward-leaning and squatting postures with 79% or more F-measure but also the number of rebar binding points by the acceleration data. Besides, we confirmed that the data augmentation method improved the accuracy of posture estimation by 5%.

Keywords: Construction worker; Sensing data analysis; Machine Learning; Behavior estimation