Video Smoke Detection Algorithm Based On Deep Learning

Fire is a frequent disaster,threaten human production,and destroy the natural environment,and effectively prevent fires in human development.Smoke is an early signal in the fire,and smog detection can provide more sufficient schedule for fire rescue,reduce property,and personally possible loss.The traditional sensor-based smoke detection algorithm has high environmental dependence,high error rate,and lag.The video-based smoke detection algorithm is strong,and the response speed is fast and more flexible,and it is increasingly concerned.The traditional smoke video detection method based on artificial extraction characteristics,there is a difficulty of calculating redundant and artificial characteristics during the detection process.At the same time,due to the irregularity of the smoke itself,the uncertainty of the motion state,the variability of the scene,etc.,resulting in a lower smoke detection classification accuracy.With the development of convolutional neural network,more advanced,abstract smoke features can be extracted by deepening the convolutional layer.Existing video smoke detecting methods based on convolutional neural networks,most of the category of smoke frames and the frame selection of smoke target areas.But in the fire rescue,the precision detection and segmentation of the smoke area is also very important.The precise smoke area can provide strong data support for the fire and spread trend of fire.However,research on deep learning-based smoke area detection is very small.This paper proposes a smoke area detection method based on depth time and space characteristics,divided into two phases.The first stage is a few frames,the second stage,the second phase of the first phase of the first phase,and the accurate segmentation of the smoke area in the first phase of the video frame is detected from massive monitoring video.The main research contents of this article are as follows:(1)Smoke detection algorithm based on motion target detection and characteristic fusion.The input video is used to use three inter-frame difference method combined with the Gaussian mixed model(GMM),and then extract smoke color characteristics,LBP characteristics,and HOG characteristics,fusion input support vector machine(SVM)training,to achieve smoke Accurate recognition of frames.(2)A dual-flow structure fume area detection convolution network based on time and space attention mechanism is proposed.First,in the study,the smoke video frame detected in the study content(1)is an input template frame,and then the spatial stream uses a semi-supervised sort model(RANet)to extract the spatial domain characteristics of the smoke object,the time flow is input,indicated.Dynamic characteristics such as diffusion and fluttering,in the end,accurate segmentation of the smoke area is achieved based on the space and space characteristics.(3)In order to eliminate smoke interference,more attention to the time characteristics of the smoke area,this paper uses the time and space attention to the fusion of time and space.The mechanism can predict the channel payment weight,thereby improving the response of the smoke moving portion corresponding to the attribute or channel,helping to divide the complete smoke area.At the same time,for the contradiction between a large number of fire monitoring data and the real-time requirements,the semi-supervised moving target sort model calculates the characteristic correlation between the current frame and the template frame,and selects a feature map with high similarity instead of all feature.Static feature of the smoke area to achieve a balance of detection speed and accuracy.(4)On the basis of the overall architecture and system deployment of the analysis system,the smoke automatic detection software system platform is developed,and the smoke image is automatically identified from the fire monitoring video to detect the smoke area,and the visual detection results are output through the human-machine interface.