A Wearable System for Articulated Human Pose Tracking Under Uncertainty of Sensor Placement

Abstract

To precisely track human motion, today's state-of-the-art employs either well-calibrated sensors tightly strapped to the body or high-speed cameras confined to a finite capture volume. These restrictions make such systems less mobile. In this paper, we aim to break this usability barrier around motion-capture technology through a wearable system that has sensors integrated directly into garments. We develop a pose-estimation approach based on classic kinematics and show that it is insufficient to analyze motion in such a system, leading to mean Euler angle errors of up to ±60° and standard deviations of 120°. Thus, we motivate the need for data-driven algorithms in this domain. Through a quantitative study, we attribute motion-estimation errors to the high-degree of sensor displacement (up to 118° standard deviation from the nominal value) with respect to the body segments that are present when human poses change. Based on controlled experiments, we develop a new dataset for such systems comprising over 3 hours of biomechanical motion recordings from 215 trials on 12 test subjects.