Deep learning for action recognition in basketball games

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

The aim of this thesis is to create a Deep Learning model for learning framing strategies for sporting events. To this end, I investigated different models and strategies for recognizing the actions taking place during a basketball match on the basis of video data provided by SportRadar. The first method explored is a Convolutional Neural Network (CNN), useful as a first model overview and for getting to grips with the data provided. The second method is a network composed of a Transformer and a Long Short-Term Memory (LSTM) Network capable of taking better account of temporal information. Both models commit a number of errors, particularly for fast actions, but the results show that the LSTM network dominates in terms of prediction and efficiency, despite its tendency to fragment labels. On the other hand, CNN struggles to predict actions correctly, achieving a much lower accuracy. However, another model very similar to the LSTM network, except for its layer-by-layer passing of information, shows better results. All in all, the LSTM-based model is promising, but cannot yet be used in real applications. However, it forms a solid basis for future improvements in the field of sports action recognition.