Reconhecimento de marca e modelo de veículos a partir de imagens

• Ano de defesa: 2019
• Autor(a) principal: Wille, Renan Barcik de Castro
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Veículos - Identificação; Sistemas de comunicação móvel; Veículos - Imagem; Processamento de imagens; Sistemas de reconhecimento de padrões; Algorítmos; Imagens digitais; Engenharia elétrica; Vehicles - Identification; Mobile communication systems; Vehicles - Imaging; Image processing; Pattern recognition systems; Algorithms; Digital images; Electric engineering; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::PROCESSAMENTO GRAFICO (GRAPHICS); Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/4165

Automatic vehicle make and model recognition can reduce costs for automated parking systems, as well as assist public entities such as the police in identifying and restraining vehicular tampering. This work aims to extract the make and model of vehicles through images. First, a study was done to list the viable characteristics of being obtained through image processing. Among them are the color, the vehicle license plate, the make and model of the vehicles. Then, it was decided to classify the vehicle make through its logo. To find it, it was used the following techniques: extraction of edges, binarization and morphology. After that, with a SVM classifier and a HOG descriptor the region containing the logo is categorized. Experimenting to improve the approach, we used the technique of locating the logo through sliding window also using SVM and HOG descriptor for classification. As the presented methods depend on local information and with the objective of improvement in relation to these methods, the finetunning of convolutional neural networks was studied. By using MobileNets and other architectures for the global classification of the image, it became possible with this method to extract not only the make but also the model of the vehicle. Finally, tests were performed on two Brazilian vehicle image datasets: The first one, called Pre-jcars-test, was used to measure the accuracy of vehicle make classification and compare the developed approaches. The best result was 79.67 % in top-1 by using convolutional neural networks. The second dataset, called Jcars-test, was used to measure the accuracy of the classification of vehicle make and model, and the best approach reached 96.89 % accuracy in the top-5, allowing to classify 354 models from 61 vehicle makes.