Research

On-body IMU-based pose tracking systems have gained prevalence over their external tracking counterparts due to their mobility, ease of installation and use. However, even in these systems, an IMU sensor placed on a particular joint can only estimate the pose of that particular limb. In contrast, activity recognition systems contain insights into the whole body’s motion dynamics. In this work, we present ActivityPoser, which uses the activity context as a conditional input to estimate the pose of limbs for which we do not have any direct sensor data. ActivityPoser compensates for impoverished sensing paradigms by reducing the overall pose error by up to 17%, compared to a model bereft of activity context. This highlights a pathway to high-fidelity full-body digitization with minimal user instrumentation.