Estudo e avaliação da plataforma computacional embarcada NVIDIA Jetson NANO para aplicações de redes neurais convolucionais na borda
Data
Autores
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Este trabalho propoê avaliar e explorar a integração entre aplicações de aprendizado de máquina baseados em redes neurais convolucionais (RNC) e sistemas embarcados, no contexto da visão computacional e da metodologia da computação de borda e processamento de dados distribuído. De forma a poder se explorar adequadamente os tópicos levantados, serão abordados os conceitos, tecnologias e metodologias funda- mentais à compreensão e aplicação de redes neurais artificiais, e como isto interage com as particularidades de sistemas embarcados. Será apresentados os fundamentos de redes neurais artificiais e RNC, quais seus subcomponentes e particularidades relevantes. Em seguida será descrito as principais plataformas de hardware utilizadas com RNCs e introduzido a plataforma selecionada para esta comparação, a NVIDIA Jetson NANO, além de outras relevantes também consideradas. Também será des- crita a metologia chamada de “computação em borda” e como ela complementa os conceitos previamente explorados. Para poder adequadamente comparar as platafor- mas escolhidas, serão caracterizados critérios como consumo de energia, tempo de execução, uso de memória, etc e sua relevância. Na implementação das aplicações teste, foi utilizado a biblioteca FastAI e a linguagem de programação Python para poder rapidamente desenvolver aplicações utilizando redes neurais convolucionais baseadas em arquiteturas bem documentadas. Em seguida, é apresentado os bancos de dados selecionados como ferramenta de comparação e suas características e composição. Finalmente, são expostos os resultados obtidos para cada banco de dados selecionado e em seguida estes são discutidos em relação aos critérios de avaliação levantados anteriormente. Pode-se concluir que, apesar da complexidade inerente da plataforma, é viável se utilizar ferramentas de aprendizado de máquina em conjunto de sistemas embarcados para atender demandas reais e que existe grande potencial a se explorar neste tipo de implementação.
This work proposes to evaluate and explore the integration between machine learning applications based on convolutional neural network (CNN) and embedded systems through the lens of computer vision and the edge computing methodology. In order to be able to adequately explore the previously brought up topics, will be addressed the main concepts, technologies and methodologies relevant to the understanding and application of artificial neural networks (ANN) and how it interacts with embedded systems. It will be presented the core concepts ANN’s and CNN’s, their subcomponents and relevant particularities. In sequence, will be enumerated the main relevant hard- ware platforms used with CNN’s and introduced the selected embedded platform for this comparison, the NVIDIA Jetson NANO, besides the other relevant platforms also considered. Also, it will be outlined the computing methodology called “edge computing” and how it complements the previously brought up concepts. To be possible to ade- quately compare the selected platforms, will be characterized criteria such as power consumption, execution time, memory usage, etc and it’s relevance for this application. In the implementation of the test applications, it was used the FastAI library and Python programming language to be able to rapidly develop applications using CNN’s based on well documented architectures. In sequence, is presented the datasets selected as a comparison tool and their characteristics and composition. Finally, the obtained results for each selected dataset are displayed and in sequence they are discussed in relation to the aforementioned evaluation criteria raised previously. It can be concluded that, besides the inherent complexity associated with the platform, it is viable to use machine learning tools in conjunction with embedded systems to meet real demands and that there is great potential to be explored in this kind of implementation.
This work proposes to evaluate and explore the integration between machine learning applications based on convolutional neural network (CNN) and embedded systems through the lens of computer vision and the edge computing methodology. In order to be able to adequately explore the previously brought up topics, will be addressed the main concepts, technologies and methodologies relevant to the understanding and application of artificial neural networks (ANN) and how it interacts with embedded systems. It will be presented the core concepts ANN’s and CNN’s, their subcomponents and relevant particularities. In sequence, will be enumerated the main relevant hard- ware platforms used with CNN’s and introduced the selected embedded platform for this comparison, the NVIDIA Jetson NANO, besides the other relevant platforms also considered. Also, it will be outlined the computing methodology called “edge computing” and how it complements the previously brought up concepts. To be possible to ade- quately compare the selected platforms, will be characterized criteria such as power consumption, execution time, memory usage, etc and it’s relevance for this application. In the implementation of the test applications, it was used the FastAI library and Python programming language to be able to rapidly develop applications using CNN’s based on well documented architectures. In sequence, is presented the datasets selected as a comparison tool and their characteristics and composition. Finally, the obtained results for each selected dataset are displayed and in sequence they are discussed in relation to the aforementioned evaluation criteria raised previously. It can be concluded that, besides the inherent complexity associated with the platform, it is viable to use machine learning tools in conjunction with embedded systems to meet real demands and that there is great potential to be explored in this kind of implementation.
Descrição
Citação
MENDES, Lucas Martins. Estudo e avaliação da plataforma computacional embarcada NVIDIA Jetson NANO para aplicações de redes neurais convolucionais na borda. 2022. Trabalho de Conclusão de Curso (Bacharelado em Engenharia Mecatrônica) – Instituto Federal de Santa Catarina, Florianópolis, 2022
