Research On Video Smoke Detection Algorithm Based On Computer Vision

Video smoke detection technology is widely used in the process of fire warning and fire rescue because of its high flexibility and low requirements for the surrounding environment.The traditional smoke detection method mainly extracts the rich texture,color and dispersion features from the smoke video sequence to realize the recognition and detection of the fire smoke image,which has some practical value and reference meaning,but the application scene is relatively simple and the accuracy needs to be improved.With the development of convolutional neural network,more advanced and abstract target features can be extracted by different network architectures,so that the detection and classification of video smoke frames can be improved to different degrees.Aiming at the problems of high false detection rate of traditional smoke detection methods and the inability of conventional convolution classification model to obtain the smoke location information,this paper studies the video smoke location information,edge information and future drift trend by using the deep learning method,aiming to improve the accuracy of video smoke detection,and provide a full range of smoke location and trend information guarantee for fire rescue.The main contents of this paper are as follows:(1)A fusion smoke detection method based on Gaussian Mixture Model and YOLOv2 is proposed.The improved Gaussian Mixture Model is used to eliminate the interference of video static target and reduce the scope of smoke detection.And YOLOv2 target detection network is used to realize the secondary screening of smoke area and the accurate detection and location prediction of multi scene video smoke.(2)In view of the complexity of smoke edge information,an end-to-end smoke region segmentation method is proposed.The deeplabv3+(encoder decoder with atoms separable revolution)network is built to realize pixel level detection and segmentation of smoke area in video frame,and the method of fully connected conditional random field is integrated to realize accurate positioning of smoke thin contour information.The smoke distribution heat map based on HSV-Gray is established to realize the visualization analysis of the smoke area.(3)Train and learn the context information of video smoke based on GAN(Generative Adversarial Network),so as to generate the video smoke frame in the future period by iterative prediction,and combine the smoke segmentation method to realize the prediction and analysis of the smoke future drift trend and thickness change,so as to provide positive suggestions and basis for the follow-up fire rescue and the development of evacuation plan.