Research On Fire Smoke Detection Method Based On Video Image

Fires occur frequently all over the world,posing a serious threat to public safety.As smoke is most likely to accompany fires,and smoke is the signal of fire,timely detection of smoke is the first line of defense to prevent fires.Traditional physical sensors have limited detection distances,which are only suitable for indoors or small spaces and are not conducive to remote monitoring.With the continuous advancement of the construction of intelligent surveillance,the video smoke detection system based on video surveillance has a wider application space.Traditional smoke detection algorithms based on video images mostly manually extract the color,texture and other physical characteristics of the smoke for detection and recognition.It ignores the timing relationship between video frames,that is,the nature of the diffusion movement of the smoke essence.Therefore,in order to improve the accuracy of smoke detection,the smoke characteristics in the time domain need to be considered.Therefore,this paper uses 2D and 3D convolutional neural networks to design a video smoke detection algorithm,and integrates the spatial and temporal characteristics of multiple frames of video images for detection to improve the accuracy.Based on the above problems,the main work of this paper is as follows:1.The basic convolutional neural network and target detection algorithm used in this paper are introduced in detail.The EfficientDet detection algorithm is improved,and the original network’s Bi-FPN feature fusion algorithm is longitudinally cross-layered.It reduces the complexity of Bi-FPN and improves performance.It also incorporates a dual-channel attention module based on channel and spatial scale.CBAM makes it possible to recall smoke areas of long-distance or small targets.Using the clustering method for my smoke data set,the size ratio of the anchor frame was recalibrated to improve the detection accuracy.In order to evaluate the quality of the detection algorithm more reasonably,the use of dual evaluation indicators is proposed.The accuracy and recall rates based on the image level are more practical for reference.2.EfficientDet replaces the extraction of suspected smoke blocks in the traditional algorithm.According to the detection results of EfficientDet,detailed rules for generating smoke candidate regions are formulated based on the score and position of the prediction frame.According to the generated smoke candidate area,16 frames of continuous video are cropped according to its size and position as the input of the 3D convolutional neural network for second confirmation of smoke.By screening out candidate regions,most of the background regions are excluded to reduce the calculation amount of 3D convolution.3.In order to fuse the spatio-temporal features of the smoke video without adding too many calculation parameters,the 3D convolution is split into one-dimensional temporal convolution and two-dimensional spatial convolution.And according to the diffusion speed of smoke before and after the fire,adjust the two-dimensional space convolution and onedimensional time convolution arrangement structure.With the help of residual network and dense connection network ideas,the channel attention mechanism is embedded,and finally a lightweight 3D residual dense attention network is formed.4.Through online collection of different scenes,different concentrations,different diffusion speeds,and different colors of smoke videos,and balance the proportion of various types of smoke in the training set,manually label one by one to build a complete smoke data set,and use data enhancement methods to expand the smoke data set.Finally,the complete endto-end dynamic smoke detection strategy process is described,and the detection effect of the video smoke detection system on different types of smoke is shown.