Data Augmentation for RFID-based 3D Human Pose Tracking

Interest in Radio Frequency (RF) based 3D human pose tracking has skyrocketed in the age of Artificial Intelligence of Things (AIoT). Compared to Computer Vision (CV) based methods, RF-based approaches are more resilient to lighting and non-line-of-sight conditions, and can better preserve user privacy. However, the majority of the current RF-based methods rely on a vision-aided multi-modal learning approach. An extensive amount of paired training data, i.e., Radio-Frequency Identification (RFID) data and vision data, must be collected, to achieve an adequate performance with the supervised-learning network. In order to mitigate such time-consuming and costly tasks, we propose a data augmentation method based on Generative Adversarial Network (GAN), named RFPose-GAN, to generate synthesized RFID data to alleviate the complications of using commodity RFID tags and receivers. In this paper, a forward kinematic layer is incorporated to generate simulated vision pose data, thus eliminating the need of using a Kinect 2.0 device in RFPose-GAN. Experiments conducted demonstrate that the synthesized data achieves accurate pose estimation performance.