| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 8.18 MB | Adobe PDF |
Advisor(s)
Abstract(s)
Informação e dados estatísticos sobre jogos de futebol são importantes para qualquer equipa, mas nem todas possuem os recursos, monetários e humanos para obter os mesmos. Sendo que atualmente grande parte das soluções requer serviços e dispositivos dispendiosos ou esforço humano para a recolha manual de dados, surgiu a necessidade de desenvolver uma solução simples e sem a necessidade de grandes recursos por parte dessas equipas. Esta dissertação visa apresentar todo o processo de estudo, experimentação, desenvolvimento e avaliação de uma solução de reconhecimento de momentos de golo em jogos de futebol. Desde a definição de momento de golo em termos de eventos à avaliação de diferentes abordagens e fluxos para a obtenção dos mesmos para a catalogação dos golos. Durante o desenvolvimento foram aplicados conhecimentos obtidos na fase de estudo e técnicas de visão computacional. Nesta tese pode ser encontrado um estudo e comparação da aplicação das arquiteturas ConvLSTM e LRCN para classificação de eventos, assim como um estudo sobre o impacto nos modelos das configurações e parâmetros na arquitetura e treino.
Information and statistical data about football matches are important for any team. However, not all of them have the resources, monetary and human, to obtain these. Since currently most of the solutions require expensive services and devices or human effort to manually collect data, the need for a simple solution without the usage of large resources arose. This dissertation aims to present the entire process of study, experimentation, development and evaluation of a solution that recognizes goal moments in football matches, from the definition of the goal moment in terms of events to the evaluation of different approaches and flows to obtain the necessary events. The knowledge obtained in the study phase is considered, as well as the application of computer vision techniques. A study and comparison of the application of the ConvLSTM and LRCN architectures for event classification can be found in the document, as well as a study of the impact of the different possible configurations and parameters in the models, regarding architecture and training.
Information and statistical data about football matches are important for any team. However, not all of them have the resources, monetary and human, to obtain these. Since currently most of the solutions require expensive services and devices or human effort to manually collect data, the need for a simple solution without the usage of large resources arose. This dissertation aims to present the entire process of study, experimentation, development and evaluation of a solution that recognizes goal moments in football matches, from the definition of the goal moment in terms of events to the evaluation of different approaches and flows to obtain the necessary events. The knowledge obtained in the study phase is considered, as well as the application of computer vision techniques. A study and comparison of the application of the ConvLSTM and LRCN architectures for event classification can be found in the document, as well as a study of the impact of the different possible configurations and parameters in the models, regarding architecture and training.
Description
Keywords
Visão computacional Golo ConvLSTM LRCN TensorFlow Python Computer Vision Goal