A vision-based
approach to
fire detection

Pedro Gomes
Pedro Santana
Jose Barata


Finding

classified image pixel according to an appearance model of fire.

typical fire’s dynamic texture

RGB [1–4],

YCbCr [5], .

], CIE Lab* [6] .

HSI [7] colour spaces.

spatio-temporal wavelet analysis

vision systems need

to exploit the well-known flickering

and textured characteristics of flames for their detection.

( 18 June 2014)

handle exceptions

manage the speed-accuracy trade-off

avoid perceptual aliasing situations

embedded with seamless calibration procedures

challenges

handling sudden background changes

determining when a computationally intensive frequency analysis is worth applying

detecting and tracking potential
distractors

people with fire-coloured clothing

automatically learning the camera-world coordinates
mapping

technique

employing a dynamic threshold to the magnitude of each pixel’s intensity variation across three magnitude of each pixel’s intensity variation across three

fire detection pipeline

segmenting fire regions according to a colour model

determining which of the segmented regions present
a dynamic texture

filtering out the regions with dynamic texture that do not exhibit the spatio-temporal frequency signature of typical fire

fire confirmation pipeline
-reduce the fire false alarm rate

detects foreground objects invariant to the presence of shadows

tracks the objects across frames

recognizes the objects’ categories

coordiate with GPS

detect distractor

boundary for thescene and area of subject

fire detection

color based analysis

Spatio-temporal Frequency Analysis

Dynamic Textures Detection

color space

HSI

RGB

YCbCr

method

HYR method

HY method

more consistent

Parameter/ Situation

indoor

rural

urban

night

detect fire or non-fire binary image

main characteristics of fire is its flickering rate
at a frequency of around 10 Hz

presence of fire signature when 10 Hz

two conditions must be met

First, the ratio of the analysed pixels that were labelled as fire must be above a given threshold

he accumulated number of zero-crossings in
the filters’ outputs must be above a another given
threshold

stage filter bank