Re-identificação de pessoas em múltiplas câmeras por meio de imagens digitais

• Ano de defesa: 2021
• Autor(a) principal: Cordeiro, Alexssandro Ferreira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Medianeira; Brasil; Programa de Pós-Graduação em Tecnologias Computacionais para o Agronegócio; UTFPR
• Palavras-chave em Português: Pessoas – Identificação; Vigilância eletrônica; Redes neurais (Computação); Persons - Identification; Electronic surveillance; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/29695

The detection of people is still a key issue in computer vision, along with tracking and REID (re-identification of people), especially in environments with multiple cameras which have variations in the monitored views. The advances in convolutional neural networks for these tasks implicated in performance gains in recent years, since there is an increase in the amount of public data recorded for supervised training, thus allowing for more efficient training. However, considering the difficulties present in the tasks of detection, RE-ID and tracking of people, such as occlusions, variations of lighting and the quality of the captured image, studies are still needed in order to increase the accuracy of these tasks. Using a platform for analyzing the flow of people, it is possible to provide information which can optimize security, surveillance and other sectors of the agro-industry, a field with a strong human presence. The study presented here used techniques from digital image processing and convolutional neural networks like Darknet-53 with the YOLOV3 framework, which showed satisfactory results in the state of the art of research in object detection and will be used to perform the detection of people and, based on this detection, it will be possible to carry out the re-identification based on the area of interest detected. For the RE-ID problem, the RESNET50 neural network was used in the Siamese and triplet loss architectures with the public datasets CUHK03 and Market15-01 for training, the first being used for training and the second for validation. The results presented in RE-ID training with the triplet loss architecture in the CUHK03 dataset were compared to the state of the art and presented superior results in the mAP requirements with 81.40% in relation to at 80.70% of the benchmark and CMC RANK-1 of 84.80% in relation to 84.30% of the benchmark in the CUHK03 dataset, in the validation with the Market-1501 dataset, the triplet loss network presented CMC RANK-1 of 84.50% in relation to 85.00% of the benchmark, demonstrating the ability to generalization in environments never seen by the network. These results confirm that it is possible to obtain information on the flow of people in order to assist sectors of the agribusiness such as security, surveillance and personnel management.