Keywords

Wearables, IMUs, Deep Learning, Joint Moment, Joint Kinematics, Ground Reaction Forces

Abstract

Acquiring joint kinematics, joint moments, and ground reaction forces (GRFs) during walking is essential for assessing disease progression and health monitoring during rehabilitation. However, spatial and temporal constraints, expert processing, and high costs limit the current gold standard methods, such as optical motion capture systems and floor-embedded force plates. Experts have suggested wearables with machine learning methods to address this issue, but their large sensor count renders them impractical for daily use, and the use of generic algorithms limits their accuracy. As a result, learning kinematics and kinetics in everyday life outside of laboratory settings is challenging. Thus, there is a need for an inexpensive, near-real-time system and an accurate method for estimating kinematics, joint moments, and GRFs. This dissertation proposes using shoe-mounted IMU sensors and deep learning to estimate these parameters across various locomotion modes, reflecting everyday walking conditions. Four different approaches are explored. The first approach uses shoe-embedded IMU sensors with novel deep learning models, DeepBBWAE-Net, Kinetics-FM-DLR-Ensemble-Net, and DL-Kinetics-FM-Net, which outperform state-of-the-art models but are computationally expensive. The second approach introduces Kinematics-Net and Kinetics-MMF-Net, which are lightweight yet maintain similar performance. Sparse IMU sensors on the feet may miss critical walking dynamics, so the third approach proposes a sensor distillation technique, transferring knowledge from a teacher model (trained with full sensors) to a student model (trained with sparse IMUs), enhancing estimation accuracy. Although our models are trained on a substantial number of subjects, deep learning models tend to perform better with larger datasets. Collecting extensive subject data is resource-intensive and time-consuming. Additionally, public datasets often differ in sensor types, locations, and protocols. Our fourth approach addresses this by proposing a domain adaptation technique that transfers knowledge from source datasets to the target dataset, improving performance in the target domain.

Completion Date

2024

Semester

Summer

Committee Chair

Choi, Hwan

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

ECE

Degree Program

Computer Engineering

Format

application/pdf

Identifier

DP0028553

URL

https://purls.library.ucf.edu/go/DP0028553

Language

English

Release Date

8-15-2024

Length of Campus-only Access

None

Access Status

Doctoral Dissertation (Open Access)

Campus Location

Orlando (Main) Campus

Accessibility Status

Meets minimum standards for ETDs/HUTs

Share

COinS