Compressed Representation for 3D Human Pose Estimation using WiFi signal

Qijia Zheng,Bin Liu,Jinyang Huang

DOI: https://doi.org/10.1109/ICCBD-AI62252.2023.00047

2023-01-01

Abstract:Human pose estimation technology has been increasingly applied in many fields, such as user authentication, activity recognition, and health monitoring. The existing human pose estimation methods are mainly based on cameras and wearable sensors, while some works try to use WiFi signal to perceive finegrained human joints in recent years. WiFi antennas collects 1D WiFi signal, while final outputs are 3D images containing human joints. However, directly mapping 1D WiFi signal to 3D image with human skeletons has limitations such as degree of accuracy and high costs of computation resources. In this paper, we propose a novel approach to bridge the gap between WiFi signal and images for single person 3D human pose estimation using commercial WiFi devices, named CRPose. CRPose takes the 1D WiFi data as input and supervised by a teacher network, which outputs the dense representations of 3D human postures. Our experimental evaluation shows that CRPose can localize 14 joints on the human joints skeleton with an average error of 5.1cm, achieving a 55% improvement in accuracy and a 64% time saving in inference stage when compared to widely recognized methods on RF-based 3D human pose estimation.