A hybrid method for fire detection based on spatial and temporal patterns

Carregando...
Imagem de Miniatura

Data

Título da Revista

ISSN da Revista

Título de Volume

Editor

Universidade Federal de Minas Gerais

Descrição

Tipo

Artigo de periódico

Título alternativo

Primeiro orientador

Membros da banca

Resumo

Fire detection is a vital task for social, economic and environmental reasons. Early identification of fire outbreaks is crucial in order to limit the damage that will be sustained. In open areas, this task is typically performed by humans, e.g., security guards, who are responsible for watching out for possible occurrences. However, people may get distracted, or may not have enough eyesight, which can result in considerable delays in identifying a fire, after much damage has occurred. Thus, the idea of having machines to automatically detect fires has long been considered an interesting possibility. Over the years, different approaches for fire detection have been developed using computer vision. Currently, the most promising ones are based on convolutional neural networks (CNNs). However, smoke and fire, the main visual indicators of wildfires, present additional difficulties for the vast majority of such learning systems. Both smoke and fire have a high intra-class variance, assuming different shapes, colors and textures, which makes the learning process more complicated than for well-defined objects. This work proposes an automatic fire detection method based on both spatial (visual) and temporal patterns. This hybrid method works in two stages: (i) detection of probable fire events by a CNN based on visual patterns (spatial processing) and (ii) analysis of the dynamics of these events over time (temporal processing). Experiments performed on our surveillance video database show that cascading these two stages can reduce the false positive rate with no significant impact either on the true positive rate or the processing time.

Abstract

Assunto

Incêndios florestais

Palavras-chave

deep convolutional neural network, hybrid wildfire detection method, temporal analysis, According to Brazil’s National Institute for Space Research (INPE), approximately 600,338 fires were recorded in South America in 2021 and the first 10 months of 2022. Of these, about 338,580 cases were in Brazil. A system that enables early fire detection followed by immediate intervention has the potential to significantly decrease such high numbers, reducing the damage caused to the environment.

Citação

Curso

Endereço externo

https://link.springer.com/article/10.1007/s00521-023-08260-2

Avaliação

Revisão

Suplementado Por

Referenciado Por