Χωροχρονική ανάλυση παικτών χειροσφαίρισης με τεχνικές βαθιάς μάθησης
Spatio-temporal analysis of handball players using deep learning
View/
Keywords
Χειροσφαίριση ; Υπολογιστική όραση ; Βαθιά μάθηση ; YOLOv8 ; DeepSORT ; Ομογραφία ; Αναγνώριση ενεργειών ; Τροχιές ; Θερμικοί χάρτες ; Multi-object tracking ; Random forest ; Extra trees ; Gradient boosting ; XGBoost ; Logistic regression ; Gaussian Naive BayesAbstract
Artificial intelligence, computer vision, and deep learning are transforming sports analytics. However, their adoption in handball remains limited. This thesis presents a practical and reproducible framework that operates directly on broadcast footage and covers the full workflow from detection to final reporting, transforming raw video into court-referenced trajectories, kinematic metrics, and visual summaries.
The system detects players using YOLOv8, maintains stable identities over time with DeepSORT, and projects movement onto a standardized 40 × 20 m handball court using homography. In this way, motion is expressed in meters rather than pixels, enabling more meaningful interpretation and comparisons across different clips. In addition, the framework produces heatmaps and concise movement profiles with consistent units and clear reading rules.
Next, action recognition is performed on the projected trajectories for seven actions: crossing, defence, dribbling, jumpshot, passing, running, and shot. The approach relies on interpretable motion features and evaluates multiple classifiers, including Random Forest, Extra Trees, Gradient Boosting, and XGBoost, aiming for robust performance when the available data are limited.
Finally, the framework is designed for real broadcast conditions and addresses challenges such as occlusions, high player density, and camera variations, while acknowledging key limitations, including the nature of monocular footage, the sensitivity of homography to point selection, and the absence of official player identities. Overall, the work delivers a modular and extensible basis for handball analytics with an emphasis on transparency, reliability, and reproducibility.
Postgraduate Studies Programme
Πληροφοριακά Συστήματα και ΥπηρεσίεςDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνNumber of pages
126Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα