Design Of Forest Fire Monitoring System Based On YOLOv4 Algorithm

Forest fire not only causes great damage to ecological resources and natural environment,but also poses severe challenges to the safety of people’s life and property and the sustainable development of national economy.The forest resources are widely distributed in China,and the traditional manpower patrol has great limitations due to the terrain and human resource cost.With the development of computer vision and the price fall of embedded communication equipment,remote forest fire monitoring based on deep learning object detection has become a hot research topic.In order to realize the real-time monitoring of forest fire,a forest fire monitoring system based on YOLOv4 algorithm is designed in this study.The video stream is collected and monitored by remote monitoring terminal in real time and transmitted to the host computer which is deployed with the improved YOLOv4 model optimized for forest fire detection task to realize the identification of flame and smoke.It has the advantages of strong real-time performance,high detection accuracy and low installation cost.The main research achievements and contents of this thesis include:(1)In view of the shortcomings of the current public forest fire dataset like insufficient sample,poor sample quality and single sample scenario,this study proposes a self-defined forest fire dataset composed of high-resolution forest fire images of multiple scenes under different lighting conditions.After image size adjustment,online data augmentation and sample annotation,the forest fire dataset containing 1500 sample images is obtained for subsequent model training,validation and test.(2)Aiming at the requirement of forest fire object detection task for multi-feature fusion to improve detection accuracy and lightweight network to accelerate inference speed,the structure of the backbone network,feature fusion module and additional module were optimized on the basis of the original YOLOv4 model.In the backbone network part,MobileNetV3 is used to replace CSPDarknet53 to realize the lightweight of the model and improve the feature extraction capability of the model.In the feature fusion module,BiFPN was used to replace PANet to improve the multiscale feature fusion ability of the model.In addition,global average pooling is introduced in SPP module to improve the robustness of model output feature space.At the same time,the influence of different IoU loss functions on the prediction performance of model in forest fire detection task was evaluated based on five evaluation indexes: Precision,Recall,mAP@0.5,F1-score and Inference Time.The experimental results showed that the CIoU loss function model has the best performance parameters,and was selected as the final improved YOLOv4 model.The enhancement of positioning and classification performance of the model was preliminarily verified by training loss,classification loss and objectness loss.(3)A total of 5 object detection models including Cascade R-CNN,Faster R-CNN,SSD,YOLOv3 and YOLOv4 with outstanding performance were selected to conduct comparative experiments with the improved YOLOv4 model under unified experimental environments and training parameters.The experimental results showed that the improved YOLOv4 model was 3.3%,13.1% and 6.8% higher in Recall,mAP@0.5 and F1 indexes,respectively,and the P-R curve was more balanced.Combined with the Inference time of 44.1ms,the improved YOLOv4 model has potential application value in forest fire real-time detection tasks.At the same time,the robustness of the model prediction performance under the conditions of image size change,image noise interference,flame like object interference and so on was tested.The experimental results show that the improved YOLOv4 model has strong robustness and has practical application value in forest fire real-time detection task.(4)The hardware equipment composition and system function realization of forest fire remote monitoring terminal are designed.The terminal uses Raspberry PI 4B as the control core,two18650 lithium batteries form the uninterruptible power supply,EC20 module to achieve 4G communication and GPS positioning function,HBV-RPI 1509 B camera module to obtain monitoring images and videos,WS-SG90 servo and WS-MG90 S servo constitute two degrees of freedom servos to drive the camera module to realize automatic patrol function.The real-time transmission of video stream between terminal and host computer is realized by video push and pull,Intranet penetration and other technologies.The terminal has the advantages of low power consumption,low transmission delay,high positioning accuracy and wide monitoring range.(5)Based on PyQt5,the interface design of the remote forest fire monitoring system is carried out to realize the real-time detection and visualization function of forest fire targets in static images,remote monitoring video streams and video files.