Dynamic Hazardous Proximity Zone Design for Excavator Based on 3D Mechanical Arm Pose Estimation via Computer Vision
Zehui Tian, Yong Yu, Feng Xu, Zhipeng Zhang
Abstract
In the construction industry, collision accidents between excavators and workers can result in a significant number of fatalities. Hence, low-cost supervision systems are necessary to monitor the motion status of excavators and extract the dynamic hazardous proximity zones around these excavators. Recent computer vision–based methods are convenient and economical, and they also do not require the installation of additional sensors on excavators; however, most of these methods only focus on two-dimensional (2D) or static hazardous proximity zones, which are excessively rough for the fine management required at construction sites. Therefore, a framework for the three-dimensional (3D) dynamic hazardous proximity zone design of excavators based on real-time 3D pose estimation is proposed herein. For this purpose, a set of excavator 3D pose estimation algorithms containing the 2D keypoint detection module and 2D-to-3D lifting module were trained based on two corresponding data sets, one of which was collected on the Internet and the other was generated via Cinema4D software. The real-time motion status was calculated based on the 3D poses estimated from the video after postprocessing and coordinate transformation. Moreover, trajectories could be predicted to form the 3D dynamic hazardous proximity zone for an excavator. Notably, this study developed a computer vision–based pipeline for sensor-free excavator 3D pose estimation and applied the 3D pose estimation algorithm to a specific management issue, thus providing a reference for excavator–worker collision prevention and transforming the conventional static and general onsite construction safety management to a dynamic and specific one.