Studio e validazione di modelli basati su intelligenza artificiale per il riconoscimento della posa e analisi biomeccanica in atleti agonisti

Human Pose Estimation (HPE) has emerged as a crucial tool in sports biomechanics, offering objective insights into athlete motion analysis through markerless systems. This thesis investigates the feasibility and accuracy of artificial intelligence-based HPE models, specifically YOLO and MediaPipe, applied to two case studies: a swimmer in a controlled laboratory setting and a rhythmic gymnast in a natural training environment. By comparing markerless HPE systems with the marker-based OptiTrack system and assessing the visual outputs of these frameworks, the research evaluates the precision and applicability of AI-driven models in elite sports scenarios. The swimmer’s case study highlights the challenges of comparing markerless and marker-based systems, emphasizing accuracy discrepancies influenced by the controlled laboratory setup, occlusions, and motion dynamics. With an overall accuracy of 55% for YOLO, static targets, such as the head, knees, and pelvis, were identified consistently (MAE < 4 cm), while dynamic targets, such as wrists and ankles, showed errors of up to 25 cm. This analysis lays the groundwork for future studies in underwater environments, where complications such as bubbles and reflections present additional challenges for HPE systems. For the gymnast, the study examines the robustness of HPE algorithms under complex, real-world conditions, analyzing how exercise complexity affects performance metrics. Notably, MediaPipe demonstrated superior performance at lower resolutions when capturing movements involving occlusions and dynamic variability. The findings reveal the limitations of current markerless systems in capturing intricate athletic movements and demonstrate their potential for non-invasive, cost-effective biomechanical assessments. This research advances the integration of AI-based HPE models in sports by addressing challenges like environmental factors, movement variability, and algorithmic constraints. The outcomes provide actionable insights for improving markerless systems and enhancing their reliability in optimizing athlete performance and mitigating injury risks, paving the way for more effective applications in sports science.