Fire Detection Algorithms Using Multimodal ... - Bilkent University

CHAPTER 4. SHORT RANGE SMOKE DETECTION IN VIDEO 58a warning may be issued.It is also well-known that wavelet sub-images contain the edge informationof the original image. Edges produce local extrema in wavelet sub-images [10].Wavelet sub-images LH, HL and HH contains horizontal, vertical and diagonaledges of the original image, respectively. If smoke covers one of the edges ofthe original image then the edge initially becomes less visible and after sometime it may disappear from the scene as smoke gets thicker. Let the waveletcoefficient I HL (x, y) be one of the wavelet coefficients corresponding to the edgecovered by smoke. Initially, its value decreases due to the reduced visibility,and in subsequent image frames it becomes either zero or close to zero wheneverthere is very little visibility due to thick smoke. Therefore locations of the edgesof the original image is determined from the significant extrema of the wavelettransform of the background image in the proposed method. Slow fading of awavelet extrema is an important clue for smoke detection. If the values of agroup of wavelet coefficients along a curve corresponding to an edge decrease invalue in consecutive frames, then this means that there is less visibility in thescene. In turn, this may be due to the existence of smoke.An instantaneous disappearance or appearance of a wavelet extremum in thecurrent frame cannot be due to smoke. Such a change corresponds to an ordinarymoving object covering an edge in the background or the boundary of a movingobject and such changes are ignored.Color information is also used for identifying smoke in video as the third step.Initially, when smoke starts to expand, it is semi-transparent. Consequently, colorvalues of the pixels within semi-transparent smoke covered image regions decreasewith comparable rates. This in turn results in the preservation of the direction ofthe vector defined by the RGB channel values. This is another clue for differentiatingbetween smoke and an ordinary moving object. By itself, this informationis not sufficient because shadows of moving objects also behave similarly. Assmoke gets thicker, however, the resemblance of the current image frame and thebackground image decreases. The chrominance values U and V of the candidatesmoke pixels in the current frame gets smaller values than their corresponding