Research On Fire Detection Algorithm In Complex Scenes Based On 3D Convolutional Neural Networks

Fire disaster is of losing control in time or space of the combustion phenomena.Hundreds of thousands of fire accidents are reported in China every year,which seriously endangers people's life development and ecological resources on the earth.Therefore,prevention and surveillance of fire disaster have become a research topic of great concern in the field of security around the world.The fire recognition and early warning technologies are mainly divided into two categories,one is the sensor-based warning system,and the other is the fire recognition algorithm based on image processing and neural network.Generally,sensor-based fire identification methods monitor the fire by monitoring CO concentration,smoke size and temperature in the air.Based on the fire recognition algorithms of image processing mainly for the identification of fire and smoke,the general method is to extract the features of fire and smoke,the fusion of each feature after neural network identification.These two traditional algorithms can't achieve good recognition rate in poor exposure conditions or broad scenes.So we propose a fire recognition algorithm with good recognition rate in all kinds of complex scenarios.Both flame and smoke have abundant dynamic characteristics in fire,making full use of their dynamic characteristics can greatly improve the accuracy of fire recognition.Complex scene fire recognition algorithm is proposed in this paper using convolutional neural network 3D to stack continuous frames into a cube convolution in time division and fuse the dynamic information of flame and smoke,this method has excellent recognition accuracy with innovation and practical significance.This paper mainly from the following three aspects of the research content.First,the background model is built by using sub-area exposure algorithm.The collected image in fire scene has the characteristics of high brightness and high contrast,so the image will be exposed inaccurately.This paper uses exposure algorithm to adjust the collected image to a suitable exposure before fire recognition,so as to ensure the quality of the image and be more conducive to the next step.Second,it will implement target detection of fire,which combines continuous inter-frame difference method and background difference method to detect moving targets.By analyzing the underlying features of the video image,the background model is constructed to segment the moving foreground,and the position,size and shape of the moving foreground are given,and the target model is updated with time.Third,the fire Recognition algorithm Based on 3D Convolutional Neural Network is different from traditional neural network by adding time dimension(continuous frame)to the input of the neural network.It extracts the dynamic characteristics of fire more directly,endows the neural network with the ability of dynamic identification,and improves the accuracy of fire recognition.Finally,through the experiments,the background modeling and target extraction in fire disaster are completed.A 3D convolution neural network is designed.The feasibility of the algorithm is improved by optimization algorithm,and the robustness of the algorithm is increased.The result will further verify the feasibility of the research.