Arquiteturas de redes neurais para condução de veículos autônomos terrestres em estradas brasileiras simuladas

• Ano de defesa: 2020
• Autor(a) principal: Fernandes, Victor Miranda
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Veículos autônomos; Deep Learning; CNN; Processamento de imagem; Autonomous vehicles; Image processing; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/18149

Autonomous vehicles are those that operate without a human driver, being one of the products that receive more investment from the automobile industry in its development, with more than 50 billion dollars invested in the last 5 years. One of the ways to implement them is using artificial intelligence techniques and, although there are already several implementations, few studies focus on driving on roads with infrastructure problems, such as: potholes, mud, etc., which can result in traffic accidents and vehicle damage. This becomes relevant, taking into account data on Brazilian highways, where, currently, only 13.7 % of them are paved. In this sense, a convolutional neural network architecture to drive autonomous vehicles on degraded roads is proposed in this work. The architecture was trained with data extracted from the game Euro Truck Simulator 2 with a realistic Brazilian map that considers the infrastructure problems mentioned above. The modeling and training of architecture was carried out in two phases: the first using data collected via the keyboard and the second using data collected from a joystick steering wheel and with the addition of image processing techniques. The results showed that the architecture trained in the second phase obtained a superior performance in average time for human intervention, that is, the time for the driver to intervene in the architecture trained by the steering wheel data was about 9 times greater when compared to the time for the architecture trained with the data collected via the keyboard.