ISBN 978-3-8439-4664-3

978-3-8439-4664-3, Reihe Informatik

Markus Miezal
Models, methods and error source investigation for real-time Kalman filter based inertial human body tracking

149 Seiten, Dissertation Technische Universität Kaiserslautern (2020), Softcover, A5

Zusammenfassung / Abstract

Accurately capturing human motion is an enabling technology for many health and sports applications. This thesis proposes and evaluates biomechanical model representations and methods for precise real-time motion capturing based on extended Kalman filters using inertial measurement units (IMUs).

The first part of this thesis is dedicated to finding a robust model representation. A generic approach for tracking a minimally parameterized kinematic chain and different redundant models with biomechanical constraints are introduced. All approaches are evaluated w.r.t. their tracking accuracy, robustness in the presence of systematically induced model errors. The results show, that the redundant models outperform the kinematic chain and the impact of the induced errors is classified.

The best-performing model is enhanced with a probabilistic ground contact estimation based on anatomical foot points. This allows global translation estimation and considerably improves orientation estimation quality under agile locomotion.

To address problems related to the usage of magnetic measurements (e.g. disturbances through environmental effects), they are excluded from the tracking workflow (tracking, initialization, IMU-to-segment calibration). It is shown that magnetometer omission raises the need for accelerometer bias treatment and that the inclusion of

biomechanical constraints helps to intrinsically align the body segments during motion, even without magnetometer information. The evaluation is carried out on 6 minute walking test sequences of 28 persons.

Finally, the full workflow developed in this thesis is compared as a whole to a typical setup with a marker-based optical system for certain biomechanical measures. The experiments show excellent correlations between the systems and range of motion errors below seven percent.