Informação de contexto aplicada à detecção de pedestres

• Ano de defesa: 2019
• Autor(a) principal: Cândido, Jorge
• Orientador(a): Marengoni, Maurício
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Palavras-chave em Português: detecção de pedestres; rede neural; árvore de decisão
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA::ELETRONICA INDUSTRIAL, SISTEMAS E CONTROLES ELETRONICOS::AUTOMACAO ELETRONICA DE PROCESSOS ELETRICOS E INDUSTRIAIS
• Link de acesso: http://dspace.mackenzie.br/handle/10899/24298

The detection of objetcs in digital images is one of the most studied and developed subjects within the computer vision field. Unlike the problem of object identification, where the basic task is classify objects into predefined classes, object detection has the difficult task of searching the entire image and answering the following questions: how many objects are in the image and what is the location of these objects? When the object being searched is a pedestrian, it is characterized the problem of pedestrian detection. In this research project we evaluated the use of additional information in the scene, here called context information, aiding the task of pedestrians detection. The context information explored in this research were the presence of floor area and the relationship between the pedestrian height and the pedestrian vertical position within the image. The information of floor area presence is obtained by means of an artificial neural network that classifies a region in the image as belonging or not to a floor area. The neural network is applied in an area below the bounding box that delimits a pedestrian detection candidate. The relationship between the pedestrian height and the pedestrian vertical position is obtained by the bounding box at the output of the pedestrian detection algorithm. Based on a statistical model, this relationship represents additional information that may indicate that the pedestrian detected by the algorithm represents a false positive and can be eliminated from the final result. This additional information is incorporated into the detector information improving its accuracy. Based on the tests performed in this thesis, we can say that the use of this additional information considerably improves the precision in the pedestrian detection algorithms proposed in the literature, considerably reducing the number of false positives