The Research On The Algorithm Of Video-Based Flame Detection

Fire is the biggest disaster all over the world,which is a leading unintentional cause of injury and death.Fire is also the most common misfortune to affect businesses and households.The goal of designing efficient flame detection algorithms with high accuracy is to reduce the risk of fire hazard.Developing robust fire alarm system is meaningful for both our communities and the nature environment.In this paper,we propose a method to detect fire and flames.Our research is structured as following: 1)Image pre-processing;2)Flame feature Analysis;3)Image Segmentation;4)Model training and optimization;5)Result discussion and conclusion.For image pre-processing part,we use Gaussian filtering and bilateral filtering methods to filter noises,and strengthen image features of hue,saturation,and lightness.We analyze flame feature in two ways,1)color moment feature from RGB,HSI,and YCbCr color space,2)image texture feature extracted using GLCM matrix method.At the image segmentation step,we use Gaussian mixture model to extract video foreground,filter the foreground by color flame model,optimize the foreground by Mathematical Morphology Closing Algorithms,and use Seed-Filling Connected Component Analysis method to conduct foreground object segmentation.By following this procedure,we come to a flame region result and input these images as training data set for SVM model.To build a well-trained flame detection model,we parameterize and optimize SVM model using different kernel functions with experiments.As a result,our model is well-trained based on color and texture feature using an SVM algorithm.Evaluation experiments with different scenes show that our approach is capable of fire detection.At the last part of this paper,we discuss pros and cons of our model with possible future directions.