An inertial sensor-based Motion capture pipeline for movement analysis

Bian, Qingyao ORCID: 0000-0002-2070-0931 (2023). An inertial sensor-based Motion capture pipeline for movement analysis. University of Birmingham. M.Sc.

Preview

Bian2023MScbyRes.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (2MB) | Preview

Abstract

Advanced human motion capture technologies have benefited both clinical diagnosis and rehabilitation in the past decades, which provided scientists and clinicians with a comprehensive understanding of human motion. So far, various motion capture methods based on different sensors, such as IMU sensors and high-speed cameras, have appeared and boosted the development of joint angle estimation methods. At present, the utilization of musculoskeletal models has enabled biomedical communities to calculate joint angles by applying a standard protocol. However, it costs great expense when we try to develop a human motion intention prediction method because few studies predict joint angles without EMG signals, which is a sort of neuro signal several ten microseconds ahead of the movement. Thus, this thesis presents works aiming to develop a lower limb human motion intention prediction method with pure IMU sensors. To improve the calculation efficiency of joint angle estimation, we developed an advanced algorithm based on Riemannian distance. A new comprehensive dataset, including both single-joint and multi-joint trials, was also collected in this part to find out suitable experimental parameters for IMU-based measurements and validate the R-distance-based joint angle estimation method. Moreover, we collected data from 6 healthy subjects to validate the motion intention prediction method proposed in the study. The subjects were asked to perform the 20s static calibration and the sit-stand-sit-walk task in the experiment. The human motion intention prediction method gives out a novel solution to low-cost motion intention prediction based on pure IMU data by fusing both musculoskeletal modeling technologies and Long Short-term memory(LSTM) neural networks. This thesis reveals that: (1) Motion in the horizontal plane performs worse when compared with those in the other two planes. The Root Mean Squared Error(RMSE) increases when movement range increases and motion speed increases except for motion in the horizontal plane, where measurement performance gets better at either a high speed or a low speed. (2) The R-distance-based method outperforms the Euler method when it comes to calculation efficiency. (3) The comprehensive motion intention prediction method outperforms both the pure LSTM method and ANN fusing MSK model when it comes to prediction accuracy.

Type of Work:

Thesis (Masters by Research > M.Sc.)

Award Type:

Masters by Research > M.Sc.

Supervisor(s):