On Early Stage Fire Detection Method In Large Space Based On Double-Band Image Processing

In every kind of disasters, fire happens most frequently. And with the high grow speed of the population and the number of cities, destructions and losses that fire causes become more and more great. Today, there are more and more skyscrapers and big buildings, how to detect fire threaten in these kinds of places as soon as possible and limit the loss to the lowest level become more important than ever.For its broad detect range and easy settlement features, much attention has been focused on a new style detection method - fire detection based on video images. This dissertation develops a new video images based fire detection system– double band image based fire detection system. In visible band, smoke was the detection object, and in the ultra red band, smolder is the detection object. Because smoke and smolder both are physical phenomenon in fire's early stage, system could indicate the existence of early stage fire if those phenomenon were detected.Smoke detection is the most important part of this dissertation. A whole process of detection was developed based on classic three steps method in pattern recognition. In suspicious region segmentation step, system uses the multi-Gaussian model to establish model of background, thus to district motive region in the current frame. In features extraction step, system extracts four eigenvalues like high/low frequency energy, color saturation energy and average boundary flicker frequency. The eigenvalues were based on the semitransparent and flickering characters of smoke. In the last step which was classifying the suspicious region to certain category, system uses the eigenvector to determine whether there is smoke by corresponding judge rules. In ultra red band, system uses area changing rate, number of sharp-angles and area similarity to judge if there is smolder in suspicious region.Methods that used in region segmentation and recognition are thoroughly discussed in this dissertation. Videos of smoke and flames in both visible and ultra red bands are used to testify those methods, and the result show that the system and methods we developed are effective.