Extrinsic Calibration of Motion Tracking Sensors in Wearable Exoskeletons: A Preliminary Study

In recent years, wearable Inertial Measurement Units (IMUs) have become widely adopted, leading to the development of various calibration techniques for both single and multi-sensory tracking systems. With IMU integration into devices like stereo cameras and optical trackers becoming more common, optimizing calibration parameters and establishing generalized multi-sensory calibration frameworks, particularly for extrinsic calibration, are essential for accurate sensor positioning relative to body segments or other sensors, especially when manual alignment is impractical. This article addresses the challenge of extrinsic calibration in wearable robotics, where the final on-the-field accuracy of wearable sensors and their integration with exoskeleton devices are ongoing areas of study. The proposed hand-eye calibration technique, commonly used in robotics and computer vision, was initially employed to assess the feasibility of using pairs of tracking sensors and was then adapted in a case study to align external motion tracking sensors with a 4-degree-of-freedom wearable exoskeleton. For the feasibility analysis, a UR5e cobot was utilized to provide standard $180^{\circ}$ rotations, and the accuracy of position and orientation was evaluated using an optical tracker as a reference sensor. For the case study, this technique involved capturing multiple configurations and positions of the exoskeleton and the corresponding poses of the external wearable sensors. The algorithm computed the roto-translation alignment matrices between the exoskeleton and the external sensors. The results of the feasibility analysis demonstrated that the hand-eye calibration proved to be suitable and accurate, with a root mean square error of 1° for orientation. The case study results showed that the standard deviation of the residuals remain below $0.5^{\circ}$, indicating accurate calibration, despite the subject's natural movements being observed instead of algorithm-specific movements.