Research On Video-based Fire Smoke Detecting Algorithms

With the rapid development of society and economy, the high buildings,underground tunnels, supermarkets and other complex buildings are increasing day byday. Once fire occurs in these places, it usually causes economic damage and the lossof human lives. So fire warning system has become more appealing in surveillancesystems. Conventional point-type thermal and smoke detectors are widely usednowadays, but they typically take charge of a limited area in space. In large rooms andhigh buildings, they inherently suffer from the transport delay of smoke from fire tosensor. Thus such detectors need to be in a close proximity of the fire. Along with theprogress of computer vision and image processing, video-based fire detection iscurrently a fairly common technology, which has remarkable advantages overtraditional methods, such as fast response, wide detection area, and little environmentpollution, and it can be used for the fire detection in high buildings, and even outdoorenvironment.Smoke is the forecasting symbol of fre. As a result, smoke detection is highlyattractive for early fre-alarm systems. In order to improve the efficiency of thevideo-based smoke detection, a novel smoke detection method using both texture andcolor features is presented. Texture features contains both local contrast and spatialstructure information. A histogram sequence of local binary pattern variance (LBPV)is used to characterize the local contrast information of smoke. Completed LBP(CLBP), which contains both sign and magnitude information are adopted to identifyspatial structure. To obtain discriminative patterns of CLBP, a three-layered learningmodel is used, which can estimate the optimal pattern subset of interest bysimultaneously considering the robustness, discriminative power and representationcapability of features to derive new discriminative completed LBP (disCLBP) features.The integration of disCLBP and LBPV can lead to remarkable improvement onsmoke texture classification performance. Furthermore, color feature is extracted bothin RGB and HIS color space. Each feature is trained by a support vector machine(SVM), then combined by Adaboost to make a final decision. Experiments show thatproposed algorithm can detect smoke effectively with a lower false-alarm rate andhigher reliability in open and large spaces. The study of video-based smoke detection technology is still in the developmentstage. Smoke detection in video is still an open challenge for computer vision andpattern recognition fields. The study of this thesis will provide theoretical basis andtechnological support to the extensive use of video-based smoke detectiontechnology.