Smoke Monitoring Algorithm Based On Machine Vision

The smoke monitoring algorithm has been widely applied in the fire security field owing to the continuous progress of the video monitoring technology and the machine vision technology.The current fire monitoring algorithm provides early warning of fire and obtains smoke range information mainly by segmenting thick smoke in images.Though it is of practical significance and of reference value,there are still problems:failure to provide early warning and an accurate thin smoke range at an early stage of smoke,failure to realize both the lightweight performance and accuracy,and single assistance in fire rescue.For this reason,this paper presents research on characteristics of thin smoke,lightweight algorithm and smoke density as well as a new smoke monitoring algorithm,aiming to improve the accurate and lightweight thin smoke segmentation and provide smoke density information for rescue workers besides smoke range.This can help rescue workers prepare the best fire aid program based on accurate smoke range,density and other information with details as follows:(1)A thin smoke high-accuracy segmentation algorithm based on the dynamic and static characteristics was proposed,which was subject to a new dynamic and static combined structure and included a static monitoring module and a dynamic monitoring module.The static monitoring module detects suspected smoke areas based on YOLOv5 s to eliminate static interference;the dynamic monitoring module realizes preliminary segmentation of thin smoke by taking advantage of the characteristic of vibe characteristic of low missing inspection of thin smoke and further filters optical flow interference not in compliance with smoke based on the Farneback algorithm,to obtain the final thin smoke segmentation range.YOLOv5 s and vibe were improved for smoke to improve the segmentation accuracy,including the addition of the attention mechanism and improvement of the loss function,elimination of the “ghost image”defect of vibe and statistical analysis of the optical flow vector of smoke.This concludes the optical flow characteristic of accurate smoke representation,so as to improve the interference filtering capability.Ultimately,the thin smoke segmentation accuracy of the algorithm in this paper has reached 73.7% and the average processing delay per frame was 143.8ms after repeated comparative tests,so the effectiveness of the algorithm was proved.(2)A lightweighted design was proposed for the static monitoring module of the thin smoke segmentation algorithm,aiming to reduce the overall algorithm complexity.First of all,based on the YOLOv5 s with improved accuracy and relying on the“Ghost Net” concept,a lightweighted C3＿ghost module was put forward in this paper.The C3＿2 in Neck of the original algorithm was replaced with C3＿ghost to reduce redundant computation.Secondly,the main network in this paper was replaced with Mobile Net V2,to improve the lightweight algorithm and the network feature extraction speed.Ultimately,knowledge distillation was designed based on the algorithm reality in this paper,which included a teacher model,the distillation method,the form of knowledge transfer,and the determination of the distillation loss function.This is to make up the accuracy of the YOLOv5 s with the network reconstructed and minimize the impact of the lightweight improvement of the static monitoring module.It was proved that the lightweighted static monitoring module has reduced the average time per frame of the thin smoke high-accuracy segmentation algorithm by 21.5ms.But,the segmentation accuracy decreased slightly,by only 3.5%.(3)A method of smoke density estimation based on a twin network was proposed.First of all,a smoke equation was established based on the optical model of smoke and a smoke density dataset was constructed on such basis.Then,an estimated smoke range density network based on the dual-channel structure of the twin network was designed with resnet50 as a feature collection network.A spatial pyramid pooling structure was added to improve estimate accuracy and the overall network loss function was designed.The method realized end-to-end mapping relation between smoke image and density and allowed quick estimation of smoke density in real-life scenarios with a mean absolute error(MAE)of 0.097 and average time consumed per frame of 28 ms.As a supplement to the smoke monitoring algorithm,the method provides more smoke information and addresses the problem that traditional smoke estimation methods cannot be applied in real-life scenarios.