AUTOMATIC VISUALIZATION OF SOCCER MATCH STATISTICS FROM BROADCAST VIDEOS USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Visualizing soccer match statistics from broadcast videos is a rising topic in terms of machine learning and deep learning. It is the process of extracting and presenting data about a soccer game, such as ball possession and pitch control, using video processing and feature extraction techniques. In this thesis, we investigate methods for extracting and representing data from televised soccer footage using Convolutional Neural Network architectures. Particularly, YOLO11 is used for Entity detection and Keypoints detection, the combination of SigLIP, UMAP, and K-Means is used for Team classification. The experiments are performed on datasets provided by Roboflow. The evaluation metrics used are mean Average Precision (mAP) at Intersection over Union threshold of 0.50 (mAP50) and at varying thresholds, ranging from 0.50 to 0.95 (mAP50-95) for entities detection and keypoints detection models. The mAP50 and mAP50-95 of the players detection model are 0.924 and 0.648 across all classes, respectively. The mAP50 and mAP50-95 of the keypoints detection model are 0.995 and 0.615, respectively.