Generating Human Pose and Activity Videos from Wearable Ultra-Wideband Sensors

Abstract

Video is the most interpretable representation for human activity monitoring, but continuous camera deployment in homes is often constrained by privacy, coverage, lighting, viewpoint, and occlusion. This paper explores whether cameralike activity videos can be generated from wearable motion measurements without using a runtime camera. We present UWB2Video, a wearable sensing-to-video pipeline that uses body-worn ultra-wideband (UWB) tags and fixed environmental anchors to estimate sparse 3D trajectories of body landmarks. These landmarks are converted into full-body pose or human mesh sequences and placed inside an offline 3D reconstruction of the room built from multi-view images. The system then renders avatar-based activity videos from arbitrary virtual viewpoints, preserving the interpretability of the video while avoiding continuous visual recording of the person. We evaluate an initial prototype using tag localization, pose and mesh reconstruction, scene reconstruction, and qualitative analysis of activity videos. Preliminary results suggest that wearable UWB can generate useful activity videos in controlled settings, while revealing open challenges in robust ranging, sparse pose completion, human-scene composition, and real-time deployment.