Abstract
AbstractIdentifying workers’ activities is crucial for ensuring the safety and productivity of the human workforce on construction sites. Many studies implement vision-based or inertial-based sensors to construct 3D human skeletons for automated postures and activity recognition. Researchers have developed enormous and heterogeneous datasets for generic motion and artificially intelligent models based on these datasets. However, the construction-related motion dataset and labels should be specifically designed, as construction workers are often exposed to awkward postures and intensive physical tasks. This study developed a small construction-related activity dataset with an in-lab experiment and implemented the datasets to manually label a large-scale construction motion data library (CML) for activity recognition. The developed CML dataset contains 225 types of activities and 146,480 samples; among them, 60 types of activities and 61,275 samples are highly related to construction activities. To verify the dataset, five widely applied deep learning algorithms were adopted to examine the dataset, and the usability, quality, and sufficiency were reported. The average accuracy of models without tunning can reach 74.62% to 83.92%.
Funder
Research Grants Council, University Grants Committee
Tsinghua University
Publisher
Springer Science and Business Media LLC
Subject
Library and Information Sciences,Statistics, Probability and Uncertainty,Computer Science Applications,Education,Information Systems,Statistics and Probability
Cited by
6 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献