Inertial human motion capture with automatic IMU-to-segment position and orientation calibration using JointTracker and deep learning with uncertainties

<ns3:p>This paper introduces a novel approach to self-calibrating online inertial human motion capture designed for dense IMU configurations. By combining model-based sensor fusion with deep learning, the proposed method eliminates the need for static pose or functional calibration, enabling automatic IMU-to-segment (I2S) orientation and position calibration during dynamic activities such as walking. The approach integrates the previously established JointTracker framework with a novel probabilistic deep learning module for I2S orientation prediction. The JointTracker estimates global IMU orientations and joint positions from synchronized IMU measurements using recursive filtering, while the calibration module extracts kinematic features, employs probabilistic sequence-to-sequence prediction with an ensemble network trained in a self-supervised manner, and applies active selection of I2S updates based on predicted uncertainties to ensure robust performance under unseen motions. The evaluation demonstrates that the proposed method achieves comparable accuracy to existing static pose or functional calibration methods performed under supervised conditions. Notably, it performs on par or surpasses previous self-calibration approaches in accuracy on simulated and real IMU data and is, to the best of our knowledge, the first method to estimate I2S orientations and positions on the fly. The training works on kinematic data from real human motions, no real IMU data needed, making it easily adaptable to different motion scenarios. Qualitative results on newly collected motion data with simulated gait deviations further validate the method’s practical applicability. This research advances inertial motion capture by offering a robust and flexible solution broadening the potential for deployment in diverse applications where dynamic calibration is essential.</ns3:p>