Vision-Aided IMU Calibrator

Abstract

Inertial motion capture depends on expressing each IMU’s orientations and accelerations in a shared world frame and knowing how each sensor is rotated relative to the body segment it is attached to. However, in practice, this requires careful placement and manual calibration routines (T-pose) that do not scale to arbitrary sensor sets or in-the-wild use. We present VAIC, a visual–inertial calibration method that uses a few seconds of monocular video of arbitrary user motion plus uncalibrated IMU streams to jointly recover (1) each sensor’s body location (as a vertex on an SMPL mesh), (2) its sensor-to-body mounting rotation, and (3) the global reference alignment to the world frame. Our method operates without predefined poses, fixed sensor configurations, or pre-calibrated hardware. Instead, we treat calibration as an optimization problem: from video-based SMPL recovery, VAIC searches for the sensor-to-vertex alignment that best explains the measured IMU signals given these vertex kinematics.

Evaluated on TotalCapture (13 IMUs), VAIC localizes sensors with 95.9% correct bone assignment and 5.8 cm positional RMSE (97.9% and 5.1 cm when excluding low-excitation motions), and achieves up to 15.5° end-to-end orientation error, which outperforms all classical calibration methods such as T-pose or walking-based calibration. When sensor positions are known, VAIC improves over T-pose calibration, reduces body-mounting error by 5.3°, and lowers overall body-orientation error. These results suggest a practical path toward plug-and-play inertial motion capture, providing reliable cold-start calibration that can also initialize or complement downstream continuous tracking systems.