A hybrid method for fire detection based on spatial and temporal patterns
de Venâncio, Pedro Vinícius A. B.,Rezende, Tamires M.,Lisboa, Adriano C.,Barbosa, Adriano V.
DOI: https://doi.org/10.1007/s00521-023-08260-2
2023-02-07
Neural Computing and Applications
Abstract: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.
computer science, artificial intelligence