Research On Accurate Recognition Algorithm Of Video Smoke Detection In Complex Scenes

Nowadays,the fire has become one of the most dangerous disasters to the public security and social development.In order to avoid the fire effectively,the staff must detect the smoke timely and accurately when the smoke appears,so as to "prevent the fire before it burns".However,it is challenging to detect smoke with changeable features(shape,texture,density,etc.)in complex environment.At the same time,if we want to carry out real-time fire monitoring,we must deal with the smoke video.Therefore,real-time detection of video smoke with changeable characteristics in complex scenes is a very challenging task.This paper mainly uses computer vision technology,combined with some traditional algorithms to detect video smoke in complex scenes.The main research contents are as follows:(1)The arrangement of datasets: Since there is no uniform public datasets for smoke detection at present,we collected the smoke video and image data through various website in the early stage.And after arrangement,four smoke video datasets(16 videos)for testing the detection algorithm and 7000 static smoke pictures for training and testing the depth network are obtained,and labeled.The video datasets contains many kinds of interference factors,which can verify the robustness of the algorithm and provide data resources for the research of smoke detection in complex scenes.(2)The extraction of motion and color feature: Analyzing the motion and color characteristics of video smoke,and the conclusions are as follows: smoke generally has the trend of upward motion;in HSV color space,the saturation S of video smoke image is low,and when the smoke appears,the brightness V of image background has the trend of increasing.In view of the above characteristics,this paper combines Gaussian mixture model(GMM)and HSV color feature analysis to extract suspected smoke areas in the video.(3)Smoke detection based on the deep convolution neural network: in this paper,the deep convolution neural network YOLOv2 is used as the primary video smoke detector to detect smoke area,so as to avoid the problems of single feature selection and poor robustness of the algorithm.In a word,the algorithm in this paper can realize the early warning of video fire smoke in complex scenes.Through several groups of comparative experiments,the effectiveness of the algorithm designed in this paper is verified.Experimental results show that this method can greatly reduce the false detection of fire smoke in complex scenes,and has higher accuracy than other methods.