Environment Adaptive RFID-Based 3D Human Pose Tracking With a Meta-Learning Approach

With the development of Radio-Frequency (RF) sensing techniques, RF based 3D human pose estimation has attracted increasing interest recently. Unlike video camera based techniques, RF sensing has the unique strength of preserving user privacy. However, due to the complex wireless channels indoors, a well-trained RF sensing system is usually hard to generalize to new environments. In this paper, we propose an environment adaptive solution for Radio-Frequency Identification (RFID) based 3D human skeleton tracking systems. We first analyze the challenges in environment adaptation for RFID based sensing systems. Following the analysis, we then propose a meta-learning approach for RFID-based 3D human pose tracking, termed Meta-Pose. The system is implemented with off-the-shelf RFID devices and can well adapt to new environments with few-shot fine-tuning, thus greatly simplifying the deployment of the trained system. We conduct extensive experiments in different indoor scenarios to validate the high adaptability and accuracy of the Meta-Pose system.