Classificação de embriaguez a partir de imagens faciais usando redes neurais convolucionais: uma abordagem baseada em respostas fisiológicas induzidas pelo álcool

• Ano de defesa: 2021
• Autor(a) principal: Silva, Washington Rodrigo Dias 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: Computação forense; Farmacocinética do álcool; Segurança viária; Classificação de embriaguez; Aprendizado profundo; Redes neurais convolucionais; Análise de imagens térmicas
• Palavras-chave em Inglês: Forensic computing; Alcohol pharmacokinetics; Road safety; Drunkenness classification; Deep learning; Convolutional neural networks; Thermal imaging
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/15207

Due to population growth and an increasing motorized vehicle fleet worldwide, road safety is still a global challenge that persists despite the awareness campaigns and the legal enforcement instruments. Most of the road safety risk factors are completely preventable, and in this sense the drunk-driving needs a special attention since a few doses of alcohol can be enough to impair motor and sensory functions required for safe driving. Hence, an individual in such condition may represent a road safety risk even presenting a blood alcohol concentration within the local legal limits. Since drunk-driving is a recurrent problem, in this master thesis we proposed a complementary method for drunk drivers enforcement considering the physiological responses induced by alcohol as intoxication indicatives. We addressed the drunkenness classification problem by identifying facial temperature distribution behaviors related to alcohol consumption. Such features were extracted from individuals facial thermal images in both drunk and sober states by a Convolutional Neural Network, which enabled the automatic pose invariant feature extraction and the model inference interpretation. This approach allowed us to identify temperature variations between the forehead and nose regions as a drunkenness indicative and also to identify an asymmetric facial temperature distribution in inebriated individuals that was not showed by previous studies. Hence, we were able to demonstrate a relation between facial temperature distribution patterns and the alcohol metabolic cycle; evidencing the viability of using images to recognize physiological effects caused by alcohol.