Gait Analysis Using Markerless Pose Estimation: A Vision-Based Approach

Abstract

Gait disorders affect a considerable portion of the population and can significantly limit
the mobility, independence, and quality of life of those affected. Early and accurate gait
analysis is essential for medical diagnosis and resulting therapy planning. Conventional
methods, such as marker-based motion capture systems, are expensive, technically complex, and often limited to laboratory environments. This thesis investigates a markerless
approach to gait analysis using 2D video data and pose estimation. The aim is to develop
a prototype that does not rely on marker-based systems and can reliably distinguish
between normal gait and Trendelenburg gait, a pathological gait pattern characterized
by a pelvic tilt to one side while walking due to weak muscles on the opposite leg.
To achieve this, various models for markerless pose estimation were integrated into a
modular system and evaluated in terms of accuracy, robustness, and practicality.
The developed system utilizes video recordings from commercially available smartphone cameras, extracts relevant joint positions and gait-related features from them,
and classifies gait patterns using machine learning techniques. For the specific case of
Trendelenburg gait detection, no suitable, publicly available datasets are available to
date. Therefore, evaluation was based on recordings specifically created for this problem. The results show that, when using the right models for pose estimation, the system
is sufficiently accurate to distinguish between normal gait and Trendelenburg gait. Furthermore, the findings highlight the potential of this inexpensive and easily accessible
method for use in physical therapy and clinical diagnostics. At the same time, existing limitations, such as inaccuracies in pose estimation or influences caused by varying
lighting conditions, are discussed.

Date of Award	2025
Original language	English
Supervisor	Stephan Selinger (Supervisor)