Green telemedicine: energy efficient healthcare monitoring

Human Motion Analysis (HMA), which involves locating and evaluating the movements of the body's joints, plays a key role in several fields, including medical and industrial applications. Traditionally, this task has relied on marker-based systems or wearable devices that, while accurate, are often invasive, expensive, and require extensive and dedicated setup. Advances in computer vision and artificial intelligence have enabled markerless Human Pose Estimation (HPE), offering a nonintrusive, scalable, and cost-effective alternative. This paradigm facilitates the advancement of telemedicine and remote monitoring allowing small medical centers to leverage advanced but portable and low-cost AI technologies. Consequently, patient care can become more sustainable by minimizing the need for transportation to specialized facilities, optimizing appointment scheduling, and ensuring prompt and effective treatment delivery. However, implementing markerless HPE in real-world scenarios presents several challenges, such as trade-offs between efficiency and accuracy, resistance to environmental variability, and domain-specific customization. This thesis addresses these challenges by focusing on three interconnected pillars: efficiency, accuracy, and applicability of HPE in real world scenarios. We leverage edge computing to improve efficiency, introducing a collaborative edge–cloud inference technique tailored for telemedicine. In addition, we propose a novel runtime domain-adaptive fine-tuning method that allows a lightweight edge model to adapt to dynamic scenarios. To deal with the accuracy challenges this thesis presents a framework designed to automatically evaluate the accuracy of an HMA pipeline by injecting synthetic errors. Additionally, it proposes two novel filtering techniques to overcome the limitations inherent in HPE pipelines. Finally, we demonstrate the applicability of these methodologies through domain-specific solutions. The primary application pertains to the medical domain, where we present a portable 3D HPE platform designed for treadmill-based gait analysis, subsequently extending its applicability to overground scenarios. The platform’s accuracy is validated through comparative analysis against marker-based systems. Additionally, we introduce automated machine learning-based tools for clinical gait assessment. Furthermore, we propose an evaluation framework for upper limb motion analysis in stroke patients. To extend the capability of our methodologies, we develop a multicamera distributed HPE system to provide robust, privacy-preserving monitoring in large and complex environments. We applied these techniques to the industrial field to predict human-robot collisions and monitor worker behavior in smart factories using wearable-integrated HPE solutions. In summary, this thesis investigates the key challenges of markerless HMA, emphasizing its medical implications. Furthermore, it extends these methodologies to another human-centric domain—workplace safety—demonstrating their applicability across a broader range of scenarios.