Study On Video Fire Smoke Detection Based On Static Feature Enhancement And Spatio-temporal Feature Fusion

Fire has caused great harm and hidden danger to the development of human society.Accurate and reliable fire detection and alarm is an important technical means to eliminate the potential fire threat in time and reduce casualties and property losses to the greatest extent.Video fire smoke detection technology has the advantages of various application scenarios,fast response speed and strong anti-interference,which has attracted the focus of relevant fields and is widely used in indoor large space and outdoor space(such as industrial plant,forest grassland and other places)fire safety monitoring.The introduction of deep learning method accelerates the development of video fire smoke detection technology and significantly improves its application level.However,the lack of large-scale fire smoke video image data set in relevant research and application makes it difficult to support the training based on depth detection model,the fire image smoke detection algorithm is not sensitive to early smoke,which reduces the generalization of application scenarios,and the dynamic features of fire smoke are not fully recognized and utilized,which affects the more accurate judgment of fire occurrence.This paper aims to establish a large-scale and diversified fire smoke image database,study the early smoke depth feature enhancement method and video smoke detection algorithm based on spatio-temporal feature fusion,develop a video smoke detection method based on static feature enhancement and spatio-temporal feature fusion,and develop video image fire detection system.It provides theoretical and data support for the development of video fire smoke detection technology.Specific research work is as follows:(1)A smoke image generation method based on multi-scale extended fusion generation adversarial network is proposed and developed.Four kinds of convolution with different expansion rates are used to form a multi-scale extended fusion module.By extracting image features under different field of perception,the rationality of information filling in large void area is ensured by the encoder,and the generation of translucent smoke texture under complex background is realized.The local feature matching branch is added to the discriminant network to judge the authenticity of the smoke detail texture.The pre-training VGG model is introduced to provide effective feature matching loss in the initial stage of training to accelerate training.To ensure the authenticity of the smoke form,the numerical simulation of early fire under the specified scene was carried out by using FDS method,which provided accurate early smoke profile information for the generation of model.Real data,generated data and mixed data are used to train and test three advanced object detection networks.The results show that the smoke images generated by this method and less real images mixed training can achieve complementary effect,and effectively improve the robustness of the detection model to the new scene.(2)To develop enhancement methods for deep features of early smoke images.In the aspect of data preprocessing,Mosaic and Focus modules are used to increase the number of small object in the training set and ensure the complete retention of key information in the downsampling process.In terms of detecting the network structure,the multi-layer CSP residual module(MCSPRes),fusion subsampling module(FSM),and self-attention pyramid pool module(SASPP)were designed respectively,the feature extraction capability of the detection network is enhanced,so that the early smoke features are strengthened in the forward propagation of the network.At the same time,deep separation convolution is used to reduce network parameters.In the aspect of feature decoding,the classification and location tasks are decoupled to enhance the expression ability of each feature parameter.The experimental results show that the smoke image detection algorithm constructed by the above method is more sensitive to early smoke,with an accuracy of 95.3%,which is superior to the existing image object detection algorithms.(3)To study and develop early smoke video detection methods based on deep spatiotemporal feature fusion.3D convolution and ConvLSTM were used as the basic units to construct the spatial-temporal feature extraction network,and the deep spatialtemporal feature evolution network(DSTEN)and multi-scale spatial-temporal feature fusion network(MSTFN)were constructed respectively according to the fusion modes of deep spatial-temporal features and static features.By setting a control experiment,the network design of the algorithm and the time parameters of the input sequence are analyzed.The results show that:In terms of spatio-temporal feature extraction,3D convolution is more suitable for spatio-temporal feature extraction of smoke in video images than ConvLSTM;In terms of the fusion of spatio-temporal features and static features,the parallel mode of multi-scale fusion can ensure the effective transmission of spatio-temporal features.In terms of the input sequence time parameters,the larger the time span and the smaller the time step,the more sufficient the dynamic features can be extracted by the network,but the accumulation of dynamic features presents the law of diminishing marginal utility.(4)A video image fire detection system model is developed.The video image fire detection system is developed according to the back-end deployment,and is applied and tested based on the existing monitoring equipment on campus.The system can rapidly expand the coverage of intelligent fire detection service by using the existing monitoring network.