Design And Implementation Of Lightweight Realtime Fire Detection System For End Device

Fire has always been a major disaster problem faced by human society,posing a serious threat to the safety of human life and property.Therefore,fire detection has always been a hot issue concerned by academia and industry.In recent years,with the development of computer science,image processing technology based on computer vision has developed rapidly,among which the target detection algorithm based on deep learning has the best performance in the field of fire detection.However,these algorithms are not designed for fire targets,and due to their high hardware requirements and the limitation of computing resources of end devices,it is difficult to apply them in production and life.To improve the detection rate of the target detection model on the fire detection task and overcome the difficulty of deploying the model on the end device side,this thesis proposes a lightweight real-time fire detection algorithm for end devices.Specifically,this thesis proposes a flame feature extraction module to extract features of different scales of flames;this thesis proposes a feature fusion module,which enhances the robustness of the algorithm and improves the recognition accuracy of the model for small flame targets by merging features in different forms;this thesis proposes a decoupled a decoupled prediction module,the detection tasks of fire detection are output through different networks,which enhances the detection ability;in order to realize the lightweight of the model and reduce the occupation of the hardware memory,this thesis adopts Depthwise separate convolution method in the specific implementation of each module,and uses model quantization and knowledge distillation method to achieve real-time detection at the end-side with less loss of accuracy.By comparing the ablation experimental results of each module of the model and the comparison experimental results with other fire detection algorithms,it is proved that the algorithm proposed in this thesis is superior to other algorithms in flame detection performance.To prove the practicability of the flame detection algorithm proposed in this thesis,this thesis constructs a flame detection system consisting of a central server and an end device.The central server includes a front-end module,a logic control module and a message processing module,and the end side is composed of a video stream module,a logic control module,a flame detection module and a message processing module.The system realizes data and instruction communication through the lightweight communication protocol MQTT,realizes file transmission through FTP protocol,and realizes video stream transmission through RTSP protocol.The flame detection model proposed in this thesis is deployed in the Docker container environment of end devices.The model performs realtime flame detection by receiving video streams and transmits the detection results to users.The system can meet the actual needs of production and life.