Generative AI for 3D Human Pose Completion Under RFID Sensing Constraints

Accurate 3D human pose estimation (HPE) from wireless signals is highly challenging due to wireless sensing constraints. A functional platform typically requires a comprehensive setup of transceivers and antennas, such as WiFi devices, FMCW radars, or RFID tags, all of which come with their own limitations. RFID-based 3D HPE often suffers from tag interference and sparse readings, resulting in incomplete skeletal pose estimations, as current RFID systems capture information from only up to 12 joints. To address this challenge, we propose a novel latent diffusion transformer (LDT) framework with a cross-attention conditioning method, termed PoseCompLDT, to accurately complete 3D poses by generating the missing joints, enabling full 25-joint configurations from partial 12-joint inputs. This marks a significant advance, as it is the first approach to achieve over 20 distinct skeletal joints for wireless sensing-based continuous 3D human pose estimation (HPE) using generative AI. Extensive experiments validate the effectiveness of PoseCompLDT in preserving motion fidelity, supported by rigorous qualitative and quantitative studies. The framework offers scalable solutions beyond RFID sensing applications, such as pedestrian tracking and health monitoring in occluded or constrained sensing environments.