Um sistema automático de detecção de incêndios baseado em aprendizado profundo para dispositivos de baixo poder computacional

• Ano de defesa: 2021
• Autor(a) principal: Pedro Vinícius Almeida Borges de Venâncio
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA; Programa de Pós-Graduação em Engenharia Elétrica; UFMG
• Palavras-chave em Português: Aprendizado profundo; Detecção de incêndios em larga escala; Otimização multiobjetivo; Redes neurais convolucionais; Visão computacional; Engenharia elétrica; Incêndios; Visão por computador
• Link de acesso: http://hdl.handle.net/1843/39452

Large-scale fires have been frequently reported both in the national and the international press in recent years. The impacts resulting from such catastrophes comprise a series of irreversible consequences, such as biodiversity losses, greater emission of polluting gases into the atmosphere, deleterious effects on human health and destruction of rural properties and cultural heritage. From this perspective, it is essential to search for effective solutions for preventing and fighting fires. A potential solution to this dilemma is an autonomous computer vision system capable of quickly identifying fire outbreaks, enabling suppression to mitigate damages and, consequently, minimizing combat and restoration operating costs. The state of the art of these systems use convolutional neural networks to recognize the main visual indicators of wildfires: fire and smoke. However, deep learning algorithms such as these are computationally expensive, have thousands of parameters, consume a considerable amount of memory and require a large volume of labeled data for training. In this context, this work presents an automatic fire detection tool aimed at low end, mobile computing devices. The fire and smoke detection model is derived from the training of a convolutional neural network on a novel database, which comprises a variety of real fire events. Subsequently, less relevant convolutional filters in the model are identified and removed, in order to preserve the detection performance obtained in an optimized architecture. The experimental results show that it is possible to build a fire detector based on deep learning that is both robust and computationally efficient, in addition to instigating a large-scale environmental monitoring scenario with distributed local processing and low infrastructure cost.