Research And Implementation Of Smoke Detection For Complicated Scenes

Fire is a kind of high-frequency and destructive disaster.It not only poses a great threat to human life,but also causes hundreds of millions of economic and property losses.Smoke often occurs in the early stage of fire,so timely smoke detection is crucial to prevent fire.Therefore,the early detection and warning of smoke based on computer vision is an important research topic in the field of intelligent monitoring.At present,existing video-based smoke detection methods include three main steps,which are motion detection,feature extraction and detection method.For the motion detection,well-known methods are frame difference and background difference.HSV(Hue,Saturation,Value)color feature,Local Binary Pattern and Histograms of Oriented Gradients features have been utilized in smoke detection.For classification methods,scholars usually select Support Vector Machine(SVM),Adaboost and Neural Network methods.However,the wide range of color and shape variations and other factors affect smoke detection accuracy.All these make smoke detection methods face great difficulties and challenges.The traditional visual features can not represent the content of smoke video effectively.In this paper,smoke detection based on deep recurrent neural network is proposed for complex scenes.Smoke detection based on multi-feature fusion is also studied.Due to the lack of large-scale smoke video database,this paper builds large-scale indoor and outdoor smoke video database.The algorithms are experimented on different database.Furthermore,we build a real-time smoke detection system.The main work of the paper is as'follows:(1)A large-scale indoor and outdoor smoke video database is built.The public smoke detection data is not sufficient,the paper collects a variety of real-life smoke videos under complex scenes,including campus,shopping malls,forests,roads and others with different lighting,weather and occlusion conditions.(2)A smoke detection method based on multi-feature fusion is proposed.The method detects motion region by three-frame difference method.After that,extracts HSV,energy,contrast,entropy,correlation and LBP features.Based on these features,we train the detection model using SVM and get the detection results.(3)A smoke detection approach based on deep recurrent neural network is presented.This method constructs a multilevel convolutional neural network based on cumulative time and space features.The multilevel convolutional model is used to extract convolution activation features.To significantly improve smoke detection accuracy,the extracted features are input into the RNNs(Recurrent neural networks)to effectively blend motion information between consecutive video frames.To put the method into application,we build a real-time smoke detection system using C++language and OpenCV library.Both of the algorithms have achieved smoke detection.Experiments on different database show that smoke detection accuracy is significantly improved,and smoke detection based on deep recurrent neural network performs better than other algorithms in the paper.