Correspondência entre pessoas em uma rede de câmeras de vigilância

• Autor(a) principal: Raphael Felipe de Carvalho Prates
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/33533; https://orcid.org/0000-0003-2099-9256
• topic: Computer Vision; Smart Surveillance; Person Re-Identification; Computação - Teses; Visão por computador; Processamento de imagens; Aprendizado do computador

The number of surveillance camera networks is increasing as a consequence of the escalation of the security concerns. The large amount of data collected demands intelligent surveillance systems to extract information that is useful to security officers. In order to achieve this goal, this system must be able to correlate information captured by different surveillance cameras. In this scenario, re-identification of people is of central importance in establishing a global identity for individuals captured by different cameras using only visual appearance. However, this is a challenging task, since the same person when captured by different cameras undergoes a drastic change of appearance as a consequence of the variations in the point of view, illumination and pose. Recent work addresses the person re-identification by proposing robust visual descriptors or cross-view matching functions, which are functions that learn to match images from different cameras. However, most of these works are impaired by problems such as ambiguity among individuals, scalability, and reduced number of labeled images in the training set. In this thesis, we address the problem of matching individuals between cameras in order to address the aforementioned problems and, therefore, obtain better results. Specifically, we propose two directions: the learning of subspaces and the models of indirect identification. The first learns a common subspace that is scalable with respect to the number of cameras and robust in relation to the amount of training images available. we match probe and gallery images indirectly by computing their similarities with training samples. Experimental results validate both approaches in the person re-identification problem considering both only one pair of cameras and more realistic situations with multiple cameras.