Deep Learning-based methods comparison for unconventional human pose estimation in sport analysis solutions

Artificial Intelligence is changing the way we approach significant decisions in all fields, among which healthcare, industry, and sport. In this context, Deep Learning (DL) is gaining more and more importance thanks to technological development, large dataset availability and cutting-edge results, making it suitable as a support for decision-making and performance analysis. This thesis aims to compare different DL-based methodologies for single person Human Pose Estimation (HPE) from monocular camera videos in gymnastics; in particular, some of the considered exercises include Balance Beam, Uneven Bars, Floor and Vault. This way, an athlete can be monitored in order to assess the performed exercises both for evaluation and training purposes. HPE in such a dynamic environment requires high accuracy and real-time responsiveness; consequently, the training phase results to be complicated due to the lack of available images depicting unconventional poses. In this sense, the current work focuses on two aspects: on the one hand it is necessary to expand the available dataset of labeled images to properly feed DL algorithms; on the other hand the assessment of existing model has been performed, trying to reduce the pose loss and the Object Keypoint Similarity (OKS) metric, while keeping the inference speed. This study opens the door towards a wider solution to automatically assess gymnastics exercises through the accurate identification of the athlete pose, and the evaluation of the elements composing an exercise. The thesis is structured as follows: the first chapter identifies the problem statement; the second chapter frames the theoretical background on DL; the third chapter describes the proposed methodology; the fourth chapter shows the obtained results; the final chapter draws the conclusions.