Potential of Convolutional Neural Network-Based 2D Human Pose Estimation for On-Site Activity Analysis of Construction Workers

Abstract

Vision-based 2D human pose estimation provides a non-invasive and effort-saving means of extracting human motion data to facilitate an automated activity analysis of construction workers, such as unsafe behavior monitoring, ergonomic analysis, and productivity estimation. However, it continues to suffer from inaccuracies and a lack of robustness, particularly under a dynamic and cluttered environment like a construction site where occlusions are prevalent. To address these issues, the authors apply convolutional neural network (CNN) to human detection and pose estimation on sequential images from site conditions. Using the benchmark training datasets that do not include any images taken from the site, the result of 2D pose estimation in testing data shows that this approach achieves a high level of accuracy and robustness considering the presence of partial occlusion. The potential of this human pose estimation method a under dynamic and cluttered construction environment is demonstrated, and its further applications for a worker activity analysis are discussed.