Sistema automático de contagem de audiência com uso de aprendizagem profunda

• Ano de defesa: 2020
• Autor(a) principal: Florentino, Caio Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Contagem de audiência; Aferição de bilheteria; Visão computacional; Aprendizagem de máquina; Audience couting; Computer vision; Machine learning; Box office records; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/18669

Counting objects or living things is a common necessity in many areas of industry, commerce and services. Automating this activity can promote an optimization of the process involved and, consequently, the reduction of time and costs. With this in mind, computer vision is an approach that offers new possibilities for digital image processing, giving the computer an increasingly similar interpretive capability to humans. This work compares the efficiency of volumetric counting techniques, both in traditional computational view and in deeplearning, in the counting of audiences in face-to-face events. As a case study, the investigation focused on the audience count of movie and / or theater sessions from the audience photos. Measuring billing automatically, accurately and transparently is a recurring need in the entertainment industry. For the accomplishment of the experiments, it was necessary to develop an image base with examples of audiences and the amount of people present. From the results it was possible to observe the great potential of the application of deep learning in this context. When compared to several automatic volumetric counting techniques available, deep learning was the strategy that presented the best results, reaching sensitivity and precision above 96%. It is proposed an Automatic Audience Counting System that contains the classification / counting modules (it uses deep learning for audience count), capture (continuous monitoring of images for better capture) and control (integrates, manages and operates the system). This work contributes with knowledge sharing in the following aspects: selection of neural networks that best perform the task of counting a large number of objects in images, development of two test image bases of detection of people in audience, specification of requirements to perform the counting task successfully and in enabling and developing an automatic audience counting system.