Detecção automática de eventos em partidas de futebol

Abstract

Increasingly the world of sports has been modernized through the use of new technologies, in football is no different. The technology contributes, for example, to the statistical survey of the teams, which helps them to improve the tactical performance, as well as the collective and individual performance of the athletes. The present work aims to develop a tool to perform the automated collection of this information through the detection of video events, using machine learning techniques and computer vision. In the area of machine learning, techniques such as convolutional neural networks and k-means are used for the recognition and classification of objects. In the area of computer vision, a new method for event detection is proposed, based on a state machine defined by the positioning of the players and the ball. In order to obtain results for the proposed tool, different forms of evaluation were done in order to obtain results for each of the proposed techniques. It is worth noting the results obtained by the Faster R-CNN convolutional neural network model for object identification and the proposed technique for player classification, which reached a precision of over 90 %. The results obtained for event detection vary according to the event, for the detection of passes for example the results were over 90 % but for more difficult events such as stealing the ball accuracy drops to about 70 %.