Title: Ultra Diffusion Poser: Leveraging Diffusion Models for Human Motion Tracking via Sparse Inertial Sensors and UWB-Based Inter-Sensor Distances

Abstract: Inertial measurement units (IMUs) offer a practical, wearable substitute for traditional camera-based motion capture systems. To counteract the drift inherent in inertial data, recent approaches to sparse inertial pose estimation have incorporated inter-sensor distances obtained through ultra-wideband (UWB) ranging. However, existing methods typically treat UWB distances merely as supplementary input features, failing to account for the physical constraints they establish regarding sensor placement. In contrast, these distances can be utilized to reconstruct the underlying three-dimensional sensor configuration, thereby generating more robust inputs for pose estimation. We introduce Ultra Diffusion Poser, a diffusion-based framework that explicitly incorporates these geometric constraints. The model features a Spatial Layout Module capable of analytically deriving 3D sensor positions from UWB data. These reconstructed positions, combined with IMU signals and UWB distances, serve as conditioning inputs during the diffusion process. Since network outputs may still deviate from measured inter-sensor distances, we propose UWB-Diffusion Guidance to ensure consistency between predicted poses and ranging measurements throughout diffusion sampling. These innovations allow our model to attain state-of-the-art results, decreasing joint position error by as much as 22% compared to previous methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC