Flame Detection In Videos Using Multiple Features

The early and real-time detection of fire flames based on video image sequence processing is mainly through the analysis of the video image sequence in the surveillance cameras, utilizing image processing technologies and changes in the characteristics of flame image in a continuous sequence. Comparing with the traditional fire detection technologies, this method could detect more information, such as the location, the size and the growth ratio of flames, etc. And it can be applied to multi-dimensions.In this paper, a thorough research has been done on the algorithms of image enhancement, noise filtering, edge detection, frequency detection and their applications in the fire video image processing. On the base of integration of algorithm above, a three-stage-based intelligent fire detection method was proposed.1) First stage is also known as the flame image segmentation stage. First, we analyzed and compared the various background updating algorithms and improved the median background updating algorithm that behaves robust. We optimized the calculation speed and significantly reduced the mean time complexity. Second, after analyzing a variety of statistical models on the color of fire flames and comparing the results, we used a compromise solution: integrating both RGB and HIS color statistical models, filtering and optimizing the extracted flame region of last step.2) Second stage is to detect other characteristics of the flame. First, we detected the flame shape. After comparing and analyzing the state-of-the-art detection methods, we proposed to use canny edge detection algorithm to extract the edges and then to calculate the perimeter and area ratio to make the classification. Second, we detected the flame frequency. We approximated to estimate the jittery frequency by the changes of flame edges. Third, by comparing the two frames between the changes of the S channel histogram to filter objects similar to flames, we detected the histogram changes. It enhanced the discriminant accuracy of the system.3) Third, the alarm system is introduced. Utilizing the fuzz logic theory, we integrated the detection results of the various features of the flames and smokes, and fuzzified them as input. Then we set the database of fuzz logic theory and output the alarm information.Finally, the simulation experiment has been done. The results suggest that our method can efficiently recognize the fire flames. And it is much faster and robust.