Date of Award

10-16-2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Systems Engineering

First Advisor

Mariofanna Milanova

Abstract

Human gait is essential for long-term health monitoring as it reflects physical and neurological aspects of a person's health status. In this thesis, we proposed a non-invasive video-based gait analysis system to detect abnormal gait, and record gait and postural parameters on a day-to-day basis. The system does not require subjects to wear any sensors or attach markers on their body, therefore it is convenient to use in daily life. It takes videos captured from a single camera as input. OpenPose is used to localize skeleton and joints in each frame. Angles of body parts form multivariate time series. Then we employed time series analysis for normal and abnormal gait classification. There are two methods implemented and compared in this thesis: BOSSVS (Bag-of-SFA-Symbols in Vector Space) based and DTW (Dynamic-Time Wrapping) - SVM (Support Vector Machine) based. They classify normal and abnormal gait by characterizing subjects' gait pattern and measuring deviation from their normal gait. In the experiment, we captured videos of our volunteers showing normal gait as well as simulated abnormal gait to validate the proposed methods. Both methods show promising results in intra-subject and inter-subject cross validation. From the gait and postural parameters, we observed distinction between normal and abnormal gait group. It shows that by recording and tracking these parameters, we can quantitively analyze how gait has changed over time.

Download

Included in

Engineering Commons

Share

COinS