Occlusion-aware 3D Human Pose Estimation and Mesh Recovery for Remote Support of Disaster Medicine (1st-report)

Evaluation of Deep-Learning-Based Human Pose Estimation and Mesh Recovery Performance in High Occlusion Environment

抄録

This research aims to develop a disaster medical tele-support system that can rapidly provide a telemedicine team with a digital twin reflecting the situation of a victim buried under rubble at a disaster site, such as an earthquake. For this purpose, we are developing a method to fit the size, pose, and joint positions of a standard 3D human mesh model to images of victims in the rubble taken by smartphones and 3D measurement point clouds and to transmit this information to a telemedicine team. In this report, as the first step, we experimentally evaluate the method's effectiveness by extending the training dataset so that the conventional 3D human body pose estimation and mesh restoration method based on deep learning can work stably even under high occlusion conditions such as debris.