3D pose estimation of workers in complex radioactive environments using a single RGB-D camera

Abstract

Accurate 3D Human Pose Estimation (3DHPE) is necessary for computational dosimetry of workers near radioactive sources, but existing frameworks have difficulty dealing with single-view configuration due to heavily occluded environments in multiple camera views. Here, a 3DHPE framework is proposed that utilizes a single RGB-D (Red, Green, Blue, and Depth) camera. First, 3D poses relative to the body root are estimated by employing ground plane detection, after which the position of the body root is determined by matching key points of the 3D pose within a voxelized point cloud reconstructed from depth images. In experiments, the proposed framework achieved a 22.7 % reduction in the mean per joint position error and a 29.2 % error reduction for the wrists when applied to a challenging dataset featuring substantial occlusions from interventional radiology equipment and other workers. The proposed framework demonstrated comparable positioning accuracy as state-of-the-art techniques without requiring the collection of actual 3D pose data. It can be utilized as a supplementary tool for reducing uncertainty in computational dosimetry and optimizing the workflow in medical settings.