Musculoskeletal Inverse Kinematics Tool for Inertial Motion Capture Data Based on the Adaptive Unscented Kalman Smoother: An Implementation for OpenSim.

Abstract

Purpose: Conventional tools for human kinematics estimation presume that observations are subject to uncorrelated, zero-mean Gaussian noise, and they provide no estimate for the uncertainty of their solutions. This paper presents AUKSMIKT-a tool for whole-body kinematics estimation in the Bayesian framework to account for these shortcomings.

Methods: We implemented AUKSMIKT as a C++ class that extends the OpenSim (v4.5) application programming interface. AUKSMIKT is based on the unscented Kalman filter combined with a run-time estimator of process and observation noises, and a fixed-lag Rauch-Tung-Striebel smoother. We tested the performance of AUKSMIKT using data from a public dataset consisting of both optical and inertial motion capture data recorded from overground walking subjects. We computed the mean absolute errors of estimated angular positions, velocities, and accelerations with respect to the gold standard optical motion capture estimates, and compared these metrics to those obtained from the least squares estimation-based tool native to OpenSim.

Results: AUKSMIKT produced smaller errors than the native tool for the angular position of three joints (0.8-1.9%), the velocities of six joints (0.7- $-$ 7.6%), and the accelerations of seven joints (3.0-13.7%). AUKSMIKT produced larger errors in the angular positions of five joints (1.3-7.6%), and the velocities of three joints (4.4-8.3%).

Conclusion: With respect to the optical motion capture solution, AUKSMIKT can estimate lower-body kinematics from inertial motion capture data with comparable or higher accuracy than the native OpenSim least squares estimator.