Development of a Machine Learning-based Marker-less Gait Analysis System for Clinical Applications

This thesis explores the development of a novel markerless gait analysis (GAn) system utilizing machine learning (ML) and deep learning (DL) techniques to enhance clinical applications. Traditional gait analysis methods, while precise, often rely on costly, laboratory-bound setups and invasive marker-based (MB) systems, limiting accessibility. To address these challenges, this research introduces a portable, non-invasive, and cost-effective markerless-based (MLB) system capable of analyzing gait dynamics using commercial camera systems and advanced DL algorithms. By leveraging computer vision techniques and ML models, the system automates the detection of lower-limb joint angles and spatiotemporal gait parameters, offering a practical alternative to MB systems. This innovative approach significantly reduces operational costs and facilitates broader adoption in diverse clinical settings, including remote or resource-limited environments. The study's methodology encompasses data acquisition campaigns, advanced preprocessing techniques, and the development of custom DL algorithms for error correction in joint angle estimation. Results demonstrate the system's effectiveness in delivering clinically relevant gait parameters, aligning closely with traditional IMU-based methods. Additionally, the research addresses critical questions regarding the precision and applicability of 2D MLB systems in clinical diagnostics and rehabilitation. This thesis contributes to the field by bridging the gap between clinical needs and current technological limitations in gait analysis. The developed system empowers healthcare professionals with an accessible tool for early diagnosis, disease monitoring, and treatment evaluation, particularly for neurological and musculoskeletal disorders. The findings underscore the potential of ML-driven MLB systems to revolutionize clinical gait analysis, making it more efficient, scalable, and widely accessible.