ThermalEdge: Uma solução em hardware para o reconhecimento de embriaguez em tempo real a partir de imagens térmicas utilizando edge computing

• Ano de defesa: 2022
• Autor(a) principal: SIlva, Públio Elon Correa da
• Orientador(a): Felipussi, Siovani Cintra
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus Sorocaba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC-So
• Palavras-chave em Português: Compressão de modelos; Visão computacional; Redes neurais profundas; Imageamento térmico; Classificação de imagens em tempo real
• Palavras-chave em Inglês: Edge computing; Model compression; Computer vision; Deep neural networks; Thermal imaging; Real-time image classification
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/17121

Alcohol consumption has presented problems for road safety, traffic enforcement can benefit from the development of methods for classifying the individual's status. In this way, recent advances in thermography associated with the advent of 5G, resulted in new opportunities for study, such as computing at the performed at the edge of the network and the use of deep neural network algorithms in microcomputers. The objective of this master's project is the development of a framework that allows the classification of thermal images, for the labeling of the individual's state in real-time, captured by a thermal camera coupled to a cellular device that sends the images to a server at the edge of the network that has a convolutional neural network model trained to recognize the individual's state from thermal images. In addition, for detecting drunkenness, hardware accelerators are used in embedded devices. In addition, an application for mobile devices was developed to allow the sending of images in real-time for classification, using the UDP protocol. Because of this, for the proposed framework to recognize the drunken state, a convolutional neural network model was trained using transfer learning to abstract characteristics related to drunkenness. The model trained for the Edge TPU obtained the best accuracy in the classification of inebriation, which was 94.33% in the test set.