Vector-based Inertial Poser: Human pose estimation with high gain observer and deep learning using sparse IMU sensors

A. Nouriani, R. McGovern, R. Rajamani

Research output: Contribution to journalArticlepeer-review

Abstract

This paper introduces a novel human pose estimation method using inexpensive inertial sensors, by combining a nonlinear high-gain observer with deep learning and kinematic modeling. Our approach offers superior estimates of sensor orientations in 3D space. It uses a deep learning model trained on vector-based representations derived from motion capture data to estimate the orientations of various segments of the body. The application of a kinematics model on the output of the deep learning network localizes the joints of the body. Finally, an optimization problem with physical constraints estimates the internal torques and increases the accuracy of the results of the kinematics model. Our algorithm presents a viable solution for in-home patient monitoring while capturing natural behaviors outside the lab. The method demonstrated superior performance across diverse activities in estimating joint positions and inter-limb angles compared to the literature. The vector-based approach reduces redundancy in human pose representation and minimizes body shape reliance. In an experiment with a dataset consisting of measurements by a motion capture system and IMUs, the proposed solution showed superior accuracy in position and orientation estimation compared to all previous methods in the literature.

Original languageEnglish (US)
Article number106432
JournalBiomedical Signal Processing and Control
Volume95
DOIs
StatePublished - Sep 2024

Bibliographical note

Publisher Copyright:
© 2024 Elsevier Ltd

Fingerprint

Dive into the research topics of 'Vector-based Inertial Poser: Human pose estimation with high gain observer and deep learning using sparse IMU sensors'. Together they form a unique fingerprint.

Cite this