Wearable Low-cost Motion Trackers for Dynamic Stability Parameters – Calibration based on Artificial Intelligence
Abstract
Background: Low-cost inside‑out wearable trackers can be deployed at scale to measure body motion. However, errors in the estimated sensor position and orientation can propagate into coordinate transformations and affect derived metrics such as gait parameters and measures of dynamic stability.
Methods: Healthy adults walked on a treadmill at 0.5–2.0 m/s while VIVE Ultimate Tracker (VUT) and Vicon data were recorded. Data-driven calibration models were developed to correct tracker coordinates and estimate the full-body centre of mass (CoM) from a sacrum-only configuration. Agreement with Vicon was assessed using root mean square error (RMSE), mixed-effects Bland-Altman limits of agreement, mean absolute error (MAE), and intraclass correlation coefficients (ICCs).
Results: Data-driven calibration models improved coordinate-level agreement. For gait parameters, model-corrected VUT showed small errors against Vicon, with MAE of 0.24–0.71 mm for step height, 1.73–4.63 mm for step length, 0.15–0.95 mm for step width, and 0.26–0.88 mm for foot clearance across speeds. Proxy CoM-derived XCoM and MoS showed excellent agreement with Vicon. For the sacrum-only CoM pipeline, data-driven calibration reduced CoM RMSE from 103.65–104.04 mm to 7.55–8.95 mm across speeds. It also markedly reduced systematic error in CoM-derived stability outcomes, with XCoM bias decreasing from 172.92 mm to 0.29 mm and MoS bias decreasing from -75.09 mm to -3.54 mm.
Conclusion: Data-driven calibration improved the measurement utility of low-cost VUTs under treadmill walking. A sacrum-only configuration estimated CoM-derived XCoM and MoS with near-reference agreement, supporting inexpensive and relatively simple gait and stability measurement in controlled settings.
Comments
Step width
Thank you for the nice poster, interesting results!
I am wondering how you determined step width from the accelerometer data, as it requires information about the position of both sensors on the feet relative to each other in a global system, right? How did you obtain this infomation? Can AI estimate these positions for instance?
RE: Step width
Thank you for your interest!
The sensors we use have interfaces on the SteamVR platform. Therefore, we can read the timeframes from the SteamVR coordinate system to determine the moment when both left and right foot sensors simultaneously touch the ground. Since our model can only output coordinate values, the step width is calculated by applying a secondary transformation to the output coordinates, rather than directly from the model output.
RE: Step width
Thank you for your interest!
The sensors we use have interfaces on the SteamVR platform. Therefore, we can read the timeframes from the SteamVR coordinate system to determine the moment when both left and right foot sensors simultaneously touch the ground. Since our model can only output coordinate values, the step width is calculated by applying a secondary transformation to the output coordinates, rather than directly from the model output.