Análise do treinamento em nuvem de redes neurais convolucionais para localização de traves em jogos de futebol de robôs humanoides

Abstract

Work aimed at researching cloud training methods in Convolutional Neural Networks for locating goalposts in humanoid robot soccer games. The need to develop these algorithms was observed in projects, barred by hardware issues with low processing power, so the objective is to evaluate solutions that already exist in the literature, but little disseminated, so that such structures can be trained without depending on hardware equipment, which often limit users by not having the necessary processing capabilities, the research was done by training two neural networks in a real database and evaluating their performance through test videos. Although the models quality and efficiency has been shown to be below the desired level, the possibility of optimizing and achieving a better result is observed.