Otimização de técnicas de visão computacional para estimativa do fluxo multidirecional de pedestres através de dispositivos embarcados

• Ano de defesa: 2019
• Autor(a) principal: Souza, Wanderson Gomes de
• Tipo de documento: Tese
• Tipo de acesso: Acesso embargado
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Engenharia Mecânica; Programa de Pós-Graduação em Engenharia Mecânica; UFPB
• Palavras-chave em Português: Visão computacional; Gestão de tráfego de pedestres; Dispositivo embarcado e trens urbanos; Computer vision; Pedestrian traffic management; Embedded devices and urban trains; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/20154

The lack of adequate monitoring for pedestrian flow control in public areas is one of the major problems arising from the absence of urban planning policies. Controlling pedestrian traffic in crowded environments has become essential to promote safety strategies and the adequate provision of resources, as the use of conventional estimation methods such as electronic turnstiles or quadrant counting may not be sufficient to ensure fluidity of data acquisition and making decisions in real time. This works aims to elaborate an estimation system for pedestrian multidirectional flow control through optimized computer vision techniques. The use of these techniques in a subway-rail enables the traffic controller to continuously, automatically and real-time monitor, generating accurate, efficient, robust and cost-effective data without directly interfering with the operation of trains. As a case study, the tests were performed at the Brazilian Urban Trains Company of João Pessoa. For this, an ethernet network consisting of three embedded devices coupled to three IP cameras installed on the ceiling of the boarding platform, near the door of the trains, was set up. A set of computer vision techniques was used, divided into two stages: the first stage consists of image acquisition, normalization and training; In the second stage, classification, delimitation by ROI, tracking and crossing of the judgment line were performed. For training and classification of images, we used the convolutional neural network SSD added to the descriptor MobileNet. However, deploying deep neural networks on embedded devices can compromise tasks that require high processing power, including real-time processing capability. For this, a hardware device, known as VPU, was added to supplement the processing power of the images, thus achieving rates of 22.98 frames per second, accuracy of 94.36%, with mean absolute and squared error at 7.8 and 8.0, respectively. The research also contributed to sharing a dataset of over 10,000 samples, including annotations with images organized at low, medium and high user density. After the tests and analysis of the collected data, it was possible to detect passenger density levels in the station, as well as to identify, through graphs, the peak times in both directions of the traffic path.