Benedetto Cutrone
Deep Learning for Athletic Human Motion Classification and Generation: An Application to Tennis.
Rel. Marco Agostino Deriu. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
Abstract
Understanding and synthesising human motion from video is a central problem in computer vision, and sport is a demanding testbed for it. This thesis studies deep-learning methods for both the classification and the generation of athletic human motion, working from 3D skeleton sequences rather than raw pixels, with tennis as the concrete application. The input data are per-frame 3D skeletons of eighteen joints, the racket included, extracted from tennis highlight videos through a YOLO-based pose-estimation pipeline and then cleaned and normalised, so that clips of different players, viewpoints and lengths become directly comparable. The first part addresses stroke recognition. A graph convolutional network based on Channel-wise Topology Refinement (CTR-GCN) is extended with a dedicated racket stream and trained on a hierarchical objective: it predicts the fine stroke sub-type, forehand slice or kick serve for instance, and derives the coarse family (Forehand, Backhand, Serve, No-Shot) from it, which keeps the two levels consistent.
Two difficulties dominate
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
