Sistema de Visão Computacional para Análise de Jogadas em Goalball

Abstract

The technologies of computer vision and machine learning have been transforming the way sports are analyzed. With these tools, it is possible to obtain more accurate in sights into athletes’ actions, optimizing strategies and improving the understanding of game dynamics. This work proposes a system for analyzing plays in goalball matches, using computer vision and machine learning techniques to recognize player actions, de tect ball bounces, and calculate its average speed. The methodological approach included the creation of a custom dataset, the use of the Yolov8 model to detect players and the ball, convolutional neural network (CNN) models and Long Short-Term Memory (LSTM) networks for action recognition, and machine learning algorithms to detect bounces in the ball’s trajectory. The results indicated that the 3D CNN model achieved an average accuracy of 0.8585 and an F1-score of 0.8569, outperforming the LSTM + 2D CNN com bination. In bounce detection, the KNN algorithm outperformed other tested algorithms, achieving an average accuracy of 0.9298 and an average F1-score of 0.8365. In estimating the ball’s average speed, the system identified 66 of the 76 plays that occurred. Of these, 36% showed an error below 10%, 38% had an error between 10% and 40%, and 26% showed an error above 40%. These results are promising, and with improvements to the models, the system has the potential to be used for real-time match analysis.