Study On Method Of Multi-Feature Fusion Based Video Flame Detection

Considering the shortage of traditional fire detectors in special facilities such as high raise building, it is now believed that as a substitution video fire detection methods based on surveillance system and machine vision will be widely used. However, most of the present video fire detection methods cannot eliminate interferences when the scenes become complex, which means there will be so many false and failed alarms when the region under surveillance turns to be complicated. As a consequence, we in the present thesis first analyzed the characteristics of video fire, and then discussed detailed moving features, color distribution and frequency property by constructing mathematical models. We at last formulated a multi-feature fusion based video fire detection algorithm.Firstly, we study the method of real time video signal acquisition to get the scenes of surveillance region by using a CCD camera and an image card. After that, the video sequences are then pre-processed to get high quality images for fire detection.Secondly, we proposed a fast Gauss mixed method based on Gauss mixed model to detect moving objects and eliminate static objects that has almost the same color with fire. Considering the correlation in space of the image pixels, we divide them into active pixels and inactive pixels and only the active pixels instead of all the pixels in the image are modeled by Gauss mixed model to determine weather the pixel is foreground or background while the property of an inactive pixel is determined by its four neighbors. This method speeds up the algorithm efficiently. We further improved the update methods for the modified self-adaptive model.Thirdly, we perform color discrimination and elimination of non-fire colored moving objects, and mark the fire colored region to be fire candidate region. A lot of fire images were collected to form a database of fire pixels under different circumstance. These fire pixels were then decomposed into different color channels, and then reconstructed in two and three dimension color space. By combining theoretical analysis and empirical fitting, a new criterion was constructed in the RGB and HIS color space.At last, we constructed a fire flicker model based on statistical analysis of the frequency characteristic to differentiate fire flames from other moving light sources in the fire candidate region. We extracted time series of image characteristics for burning experiments, and then transformed them into Fourier domain. We found the main frequencies are in good agreement with the ones calculated by Pagni's law, and the frequencies of fire are almost below 10Hz. Based these findings, we constructed a fire flicker model to describe the variance of the intensity of pixels.To realize the proposed multi-feature fusion based video fire detection method, we built a fire detection system, including hardware and a software entitled'Firedetect,'which is programmed in C language. The performance of the system is then evaluated by using different video clips of different scenes as well as online experiments performed in Tibet. Speed of detection, rate of positive fire alarm and processing speed are focused. Results reviewed that the'Firedetect'is good at self-adaption and has high enough positive alarm rate. When planted on a computer with an AMD 2.04GHz CPU and 1G memory, the'Firedetect'system can process a 320*240 pixels video scene with a speed of 24 frames per second.