Research Of Flame Detection Algorithm Based On Video

Fire is one of the disasters with high frequency and great losses at home and abroad.Its existence seriously threatens human life and property safety,and also causes economic losses of the whole society.So a timely and accurate detection of the occurrence of a fire is a very meaningful thing for social progress.However,the traditional fire detectors based on photosensitive smoke are not only affected by the environment,but also have the problems of slow response speed and high false detection rate.With the deepening of the research of artificial intelligence and image processing theory and the popularization of video surveillance equipment,more and more attentions have been paid on the way of detecting the video flame by computer vision.Aiming at the shortcoming of the traditional flame color detection model,such as the simple model and the single flame detection type,we propose a Gauss distribution model based on RGB color channel,assuming that the channels of the flame regions are independent of each other,and based on the probability of knowledge derived metric function of the model.The skewness of a flame image is defined according to the asymmetry of the R channel pixel values distributed around the average value of the flame image.According to the selective attention mechanism of human visual system,we can quickly locate the salient objects in images without training.This paper introduces the concept of visual saliency from bottom to top.First,the two initial stage image features of the image are obtained from the video frame sequence,which are local saliency and global saliency.Then the two characteristic maps are weighted together,and the comprehensive flame saliency map is obtained to represent the candidate region of the flame.Then,we use the improved flame color characteristic model to estimate the candidate area,and use the accumulative frame difference algorithm to exclude other significant targets that have the same color with the flame.Finally,the space structure of the flame and the characteristics of the region geometry are extracted,and the final classification and discrimination are carried out by the support vector machine.This method takes the place of the previous method of identifying the candidate flame region by testing the moving target first according to the dynamic foreground algorithm.The experimental results show that the proposed algorithm integrating local saliency,global saliency and rarity can extract the flame candidate area accurately under complex background,and improve the recognition efficiency.The confidence and skewness of the flame image presented in this paper can also describe the flame characteristics more accurately.The use of support vector machines for flame classification under multiple features is also effective.In conventional flame video,the interference can be eliminated and the flame area is correctly identified.