Deep Learning-based Smoke And Fire Detection Algorithm And Monitoring System Research

Fire is a very destructive natural disaster that can cause a lot of damage in a very short period of time,seriously threatening people’s lives and property.In order to prevent further fire damage and reduce the losses caused by fires,it is practical to monitor fires in real-time,quickly,and accurately.By combining deep learning and edge computing technologies,the thesis proposes a real-time fire monitoring system architecture that improves the accuracy of the original flame and smoke detection algorithm to identify fires accurately and provide early warning.The main research works in the thesis are:(1)The original object detection algorithm YOLOX is poor in detecting flames and smoke,and two improvements are proposed in the thesis to improve the accuracy of the flame and smoke detection algorithm.Firstly,in order to achieve a better balance between detection speed and detection accuracy for the improved flame and smoke detection algorithm,a novel lightweight convolutional GSConv is cited in the thesis to improve the backbone and neck of the baseline.The improved algorithm reduces the number of parameters by 1.13 M,the floating-point operations by 2.72 GFLOPs,and improves the m AP by 0.84%.Secondly,the thesis also proposes a CSPLayer＿attention structure based on the attention mechanism,which solves the problem of difficulty in fusing feature information between different channels of the original algorithm by using the SE channel attention mechanism and the channel shuffle technique,and improves the ability of the algorithm model to fuse global feature information.The improved network model increases flame and smoke detection accuracy by 0.87% and 1.96%,respectively,and improves m AP by 1.42%.(2)The thesis proposes an architecture for a real-time fire monitoring system based on edge computing.In this system,the enhanced algorithm is deployed at the edge device side,and the video stream captured by the camera is sent to Deep Stream,which detects flames and smoke by invoking the enhanced object detection algorithm while using Tensor RT for inference acceleration.The processed data is sent via the MQTT protocol to the MQTT server in the cloud,and the server forwards the data to the web function,which calls the cloud database TD-SQL to store the data.At the same time,the MQTT server in the cloud also republishes the fire alarm information with a specific topic,and managers receive alarm messages by subscribing to that topic.After testing,the system has a millisecond delay and has some practical application value.