Research On Forest Fire Detection Algorithm Based On Computer Vision

Forest fires have become a serious natural hazard threatening ecosystems,infrastructure and human life.In the work of preventing and extinguishing forest fires,countries around the world have invested a lot of manpower and material resources.In order to prevent the rapid spread of forest fires and prevent their formation of large-scale disasters,it is necessary to monitor and timely warn the occurrence of forest fires.Based on computer vision technology,this thesis studied two forest fire detection algorithms,analyzed the advantages and disadvantages of these two algorithms.Improved their performance,and a real-time forest fire detection system using unmanned aerial vehicle(UAV)is implemented:(1)Traditional image recognition algorithms based on artificially set features are used as the theoretical basis.The advantage of this kind of algorithm is low cost.However,due to its need to rely on experience for feature setting,the anti-jamming performance is poor.Based on the difference of H,S,V components among the pixels of the image to be detected,the K-Means clustering algorithm is used to separate the suspected fire area from the background.According to the color characteristics of the fire image,the suspected fire areas are extracted to reduce background environmental interference,the accuracy rate of fire detection is 92.22%.In addition,setting sample entropy thresholds to re-identify suspected fire areas is used to eliminate some fire-like disturbance sources.It improves the antijamming performance of the algorithm and finally reaches the purpose of forest fire detection,the accuracy rate of fire detection was improved to 95.03%.(2)Object detection algorithm based on convolution neural network is used as the theoretical basis.This kind of algorithm has high accuracy and strong anti-jamming performance,but it has the disadvantage of large parameter and calculation.In this paper,the YOLOv4 object detection model is used as the main framework for forest fire detection,and two parts are improved.The first part: the algorithm uses lightweight Mobile Net V3 to reduce the complexity of the conventional YOLOv4 model.Second part: the pruning algorithm eliminates redundant channels and further reduces the amount of parameters and computation.The detection accuracy of the model is improved by using the knowledge distillation algorithm.(3)The accuracy,running speed and memory occupation of the two forest fire detection algorithms are compared,and the implementation process of two forest fire detection algorithms is designed.After comprehensive consideration,the Pruned + KD model is chosen to be deployed on the embedded development board.Compared with the YOLOv4 model,the number of parameters is reduced by 95.87%,the model inference time is reduced by 74.36%,the detection accuracy is only reduced by 5.80% and the accuracy is 99.35%.Finally,the algorithm is deployed on the NVIDIA Jetson Xavier NX development board,which can detect 26.74 images in real time per second.The forest fire detection algorithm based on computer vision proposed in this paper has the advantages of small number of parameters and fast running speed,so it is suitable to be deployed on the miniaturized embedded system based on UAV.