Efficient Refinement and Analysis of Human Pose Estimation for Healthcare and Industry 5.0

Human Pose Estimation (HPE) advancements have unlocked new possibilities in markerless human motion analysis. Unlike marker-based methods, markerless systems leverage computer vision algorithms to extract pose information from images or videos, making motion analysis more accessible. However, these systems are highly vulnerable to noise and occlusions, which can significantly degrade data quality. This thesis aims to develop and assess novel real-time filtering approaches that combine learned models with traditional filtering methods. This allowed us to apply hpe across healthcare and Industry 5.0 contexts, specifically in four real-world scenarios: gait analysis and post-stroke monitoring in healthcare, distributed 3D motion analysis and human-robot interaction in industrial applications. For gait analysis, the thesis presents MAEVE, a portable, low-cost platform implemented on a low-power computing device that ensures privacy by design. Extensive evaluations of MAEVE's accuracy and performance are conducted for treadmill and overground scenarios. The thesis also presents a Markerless Automatic Video-based Gait Analysis (MaGA) system, which leverages support vector machines to distinguish pathological walking patterns from physiological ones. For post-stroke monitoring, the thesis analyzes upper limb movements through the Finger-to-Nose Test (FNT), a standard assessment of coordination and motor control in stroke recovery. It demonstrates that markerless systems can effectively measure key kinematic features, highlighting that individuals with cerebellar lesions exhibit higher trial-to-trial variability in peak velocity and timing. The thesis proposes BeFine, a distributed real-time 3D HPE platform for industrial contexts. BeFine integrates 3D HPE nodes on edge devices to capture poses from multiple viewpoints, aggregating this data through a centralized network that applies real-time filtering, clustering, and association algorithms. Also, the thesis tests advanced deep-learning techniques, such as Active Learning and Knowledge Distillation, to increase the accuracy of such distributed systems. Finally, for human-robot interaction, the thesis investigates collision prediction by analyzing the behavior of robotic arms and the spatial occupancy of humans and robots. It proposes a filtering pipeline that enhances incomplete 3D human poses obtained from a single RGB-D camera, effectively mitigating the impact of occlusions during interaction. This thesis analyzes the state of markerless motion capture systems, offers robust solutions to address their limitations, and demonstrates their applicability in diverse real-world scenarios.