Re-Identificação de veículos em uma rede de câmeras não sobrepostas

• Ano de defesa: 2021
• Autor(a) principal: Oliveira, Icaro Oliveira de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• 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; Redes neurais (Computação); Visão por computador; Processamento de imagens; Sistemas de reconhecimento de padrões; Algoritmos; Vehicles - Identification; Neural networks (Computer science); Computer vision; Image processing; Pattern recognition systems; Algorithms; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA::TELECOMUNICACOES::SISTEMAS DE TELECOMUNICACOES; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/26663

The vehicle re-identification through a network of non-overlapping cameras is an important task to assist surveillance activities, travel time estimation, enforcement of speed limits, criminal investigations, and traffic flow. The challenges of this problem come from the high inter-class similarity, caused by distinct vehicles manufactured in the same color, model and brand that often look exactly the same, and from the high intra-class dissimilarity, caused by abrupt illumination changes or camera viewpoints, that makes two instances of the same vehicle have no significant resemblance. In this thesis we investigated the main features that could be extracted and combined só as to solve the vehicle re-identification problem. For that purpose, we developed a two-stream Convolutional Neural Network (CNN) that simultaneously uses two of the most distinctive and persistent features available: the vehicle’s appearance and its license plate. In the first network stream, shape similarities are identified by a Siamese CNN that uses a pair of low-resolution vehicle patches recorded by two different cameras. In the second stream, we use a CNN for Optical Character Recognition (OCR) to extract textual information, confidence scores, and string similarities from a pair of high-resolution license plate patches. Then, features from both streams are merged by a sequence of fully connected layers for decision. State-of-the-art algorithms for this task usually take advantage of one of these attributes, however, in many cases it is difficult to distinguish one vehicle from another by considering only one type of feature. As another important contribution, we introduce a novel dataset for vehicle identification, called Vehicle-Rear, that contains more than three hours of high-resolution videos, with accurate information about the brand, model, color and year of nearly 3,000 vehicles, recorded in two roadway intersections of Curitiba city — whose infrastructure was carried out with support of the City Hall — and that has the unique characteristic of having high quality images of vehicle’s rear end, which allows its identification by the license plate. The main issue of existing datasets for vehicle re-identification is the fact that the license plates in many other countries are linked/related to the owner, not to the vehicle as in Brazil, and therefore they are intentionally redacted to respect privacy restrictions. In our experiments, our architecture achieved an accuracy, recall and -score of 99.3 %, 98.5 % and 98.9 %, respectively. As a final contribution, we also discuss and compare three alternative architectures that also explore features over the vehicle’s shape and license plate, but using additional streams and temporal information.