Automatic identification of football players based on game video sequences for the production of metadata
Michel, Tom
Promoteur(s) : Wehenkel, Louis ; Van Droogenbroeck, Marc
Date de soutenance : 26-jui-2017/27-jui-2017 • URL permanente : http://hdl.handle.net/2268.2/2550
Détails
Titre : | Automatic identification of football players based on game video sequences for the production of metadata |
Titre traduit : | [fr] Identification automatique de joueurs de football basée sur des séquences vidéo de match pour la production de métadonnées |
Auteur : | Michel, Tom |
Date de soutenance : | 26-jui-2017/27-jui-2017 |
Promoteur(s) : | Wehenkel, Louis
Van Droogenbroeck, Marc |
Membre(s) du jury : | Geurts, Pierre
Barnich, Olivier |
Langue : | Anglais |
Nombre de pages : | 133 |
Mots-clés : | [en] Football [en] Deep learning [en] Machine learning [en] Computer vision [en] Player identification |
Discipline(s) : | Ingénierie, informatique & technologie > Ingénierie électrique & électronique |
Organisme(s) subsidiant(s) : | EVS Broadcast Equipment |
Institution(s) : | Université de Liège, Liège, Belgique |
Diplôme : | Master en ingénieur civil électricien, à finalité spécialisée en "electrical engineering" |
Faculté : | Mémoires de la Faculté des Sciences appliquées |
Résumé
[en] For the broadcast of live football games, a lot of people are required: the director chooses at any time the most interesting camera to show on TV while the so-called loggers encode the statistics of the match. More and more, we try to help those people by automating their tasks. To do so, artificial intelligence methods must be developed for computers to understand what happens on the pitch. In order to achieve this goal, many subproblems must be solved separately. Automatic identification of players is one of them. This master thesis tries to handle this specific problem by focusing on the players numbers. The whole work is separated into 3 main parts: player detection, digit detection and digit recognition. Those different tasks are tackled with the help of computer vision and deep learning techniques. In particular, a segmentation semantic method is implemented. This allows us to know whether a pixel in the image belongs to a player or a digit with an accuracy higher than 98%. Digit recognition also obtains very promising performances. On the whole, this master thesis offers an array of techniques for each individual task. With more training data, those techniques would give robust models that could be applied to any football game.
Fichier(s)
Document(s)
Description: -
Taille: 55.42 MB
Format: Adobe PDF
Annexe(s)
Description: Folder containing the source code developed in this master thesis.
Taille: 56.22 kB
Format: Unknown
Description: Folder containing some videos of the main results of this master thesis.
Taille: 78.37 MB
Format: Unknown
Description: One page summary with illustrations.
Taille: 1.75 MB
Format: Adobe PDF
Citer ce mémoire
L'Université de Liège ne garantit pas la qualité scientifique de ces travaux d'étudiants ni l'exactitude de l'ensemble des informations qu'ils contiennent.