| With the rapid development of my country’s economy and the continuous expansion of computer rooms and warehouses,automatic fire alarm systems are of great significance to the normal operation of the industry.Fires often occur in places where materials are concentrated and have a tendency to spread rapidly.Traditional fire detection sensors have the disadvantage of delaying early warning in the early stage of a fire.In addition,the computer room and warehouse contain high-tech and expensive storage and operation equipment,which will cause major losses if burned.Therefore,how to combine video surveillance technology on the basis of existing fire early warning equipment to reduce the time of fire early warning has become an urgent problem to be solved.In response to the above problems,this paper proposes a fire detection algorithm and a set of fire monitoring and early warning systems.This article first introduces the research status of domestic fire detection technology,and briefly explains the difficulties of current fire detection technology.Then the relevant theoretical knowledge of convolutional neural network,deep learning framework,and target detection technology is explained.For the current research flame detect no accepted body mass,the whole issue of the type of flame datasets,this paper describes the structure of the data set fires,and how to use data enhancement methods to enhance the image data set fire information.The evaluation index adopted by the target detection algorithm in this paper is given.The detection effect of traditional flame detection methods cannot meet the application requirements.In this paper,an improved YOLOv4 flame detection algorithm is proposed and verified by experiments on a self-made fire data set.Improved YOLOv4 flame detection algorithm According to the size distribution of fire targets in the data set,K-means++ clustering algorithm is used to obtain target suggestion boxes suitable for flame detection tasks.Through the use of target suggestion box clustering optimization,network structure improvement and multi-scale training strategies,the bounding box regression loss function and attention mechanism are selected,and the convolutional layer optimization design is carried out to improve the real-time detection and accuracy of flame detection tasks.Experimental results show that when the input image resolution is 1024×1024,the m AP of the improved flame detection algorithm can reach 95.35%,and the detection speed can reach 27 FPS,which meets the requirements of video surveillance flame detection tasks.In response to the actual needs of the video fire monitoring system,the improved YOLOv4 flame detection algorithm was applied to the existing computer room and warehouse monitoring cameras,and a video fire monitoring system was built.The video monitoring system using fire B/S frame design,combined fire sensing data intelligence platform,provides more information to the fire detection determination.The upper computer part of the software includes the database relationship of the video fire monitoring system and the corresponding functional interface design.The software is mainly composed of functional modules such as login interface,fire detection interface,duty personnel interface,status monitoring interface,alarm information interface and maintenance work order interface.Experiments on the self-made flame data set in this article,the improved YOLOv4 flame detection algorithm is 15.04%,13.48%,8.71% higher than the commonly used SSD,YOLOv5,Fast RCNN and Faster R-CNN target detection algorithms,respectively.And 5.89%,verifying the obvious advantages of this algorithm in video fire monitoring,and in different distances and environments,the detection algorithm and the monitoring system built in this paper have good flame detection effects. |
