Evaluation of Models for Estimating of Handball Game Flow

Abstract: This thesis presents and evaluates different models aimed at tracking the play in handball matches. The tracking is designed to be used with automatic broadcasting of handball matches without any camera operator present. To solve this problem the Kanade-Lucas-Tomasi tracker is used to generate data of player movements on the handball court. From the tracker data, features are extracted and used as input to the models being evaluated. Three different types of models are evaluated. One of the models are an artificial neural network (ANN) model, created from machine learning algorithms. The result shows that using an ANN model is the best approach of the models tested. They also show that the features chosen to describe the game flow used for making estimations are more important than the structure of the ANN. When using ANN, the model estimates the game play in almost all situations but some key events may still be missed.