Generative AI-Empowered RFID Sensing for 3D Human Pose Augmentation and Completion

Abstract

Collecting paired Radio Frequency Identification (RFID) data and corresponding 3D human pose data is challenging due to practical limitations, such as the discomfort of wearing numerous RFID tags and the inconvenience of timestamp synchronization between RFID and camera data. We propose a novel framework that leverages latent diffusion transformers to generate high-quality, diverse RFID sensing data across multiple classes. This synthetic data augments limited datasets by training a transformer-based kinematics predictor to estimate 3D poses with temporal smoothness from RFID data. Most importantly, we introduce a latent diffusion transformer training stage with cross-attention conditioning and an inference design of two-stage velocity alignment to accurately infer missing joints in skeletal poses, completing full 25-joint configurations from partial 12-joint inputs. This is the first method to detect >20 distinct skeletal joints using Generative-AI technologies for any wireless sensing-based continuous 3D human pose estimation (HPE) task. The application is particularly important for RFID-based systems, which typically capture limited joint information due to RFID sensing constraints. Our approach can extend the applicability of wireless-based pose estimation in scenarios where collecting extensive paired datasets is impractical and achieving more fine-grained joint information is infeasible, such as pedestrian and health monitoring in occluded environments.