Design And Implementation Of Forest Fire Detection Algorithm Based On Video Monitoring

Forest fires are one of the most frequent natural disasters around the world,burning large areas of forests and causing great damage to the global ecosystem.Video monitoring is a computer vision-based technology with the characteristics of real-time,high adaptability and high accuracy,which is very effective for timely and accurate monitoring of forest fires.At present,forest fire detection methods based on video monitoring can be divided into two categories: traditional image recognition algorithms and target detection algorithms.However,traditional image recognition algorithms have the defects of difficult physical feature selection and design and low detection accuracy.And the target detection algorithm has problems such as complex network and large number of parameters,and poor quality of existing forest fire datasets.In this paper,we build a highquality forest fire dataset by manually screening the open source forest fire dataset,field experiment collection data and forest fire video monitoring data provided by the Forestry and Grass Bureau,and optimizing the data annotation method.The research is carried out to address the problems of forest fire detection methods,and the main contents are as follows:(1)Design and implement a set of forest fire detection algorithms based on multifeature fusion.To address the image quality problems of existing forest fire video monitoring equipment,Gaussian filtering,mean filtering and saturation adjustment are selected as pre-processing methods.The color and shape of smoke and flame are used as static features for the design and extraction of the method.Vi Be algorithm is used as a dynamic feature extraction method for smoke and flame.The performance of the implemented multi-feature fusion algorithm is evaluated.The experimental results show that the color feature extraction method proposed in this thesis has better results(71%accuracy)than the methods proposed by other scholars;the introduction of the shape extraction method further improves the performance of the method(82.40% accuracy,11.40% improvement).(82.40%,an improvement of 11.40%);the implemented multifeature fusion forest fire detection algorithm has the advantage of high accuracy and can be real-time compared with the existing traditional methods(frame detection rate of84.63%,FPS of 39.51).(2)A set of forest fire detection algorithms based on the lightweight convolutional neural network YOLOv4-Light is designed and implemented.To address the lack of computing power of existing forest fire monitoring systems,this thesis introduces Mobile Net V3 in the YOLOv4 algorithm to replace its original backbone feature extraction network,and then introduces deep separable convolution into the enhanced feature network and prediction network to implement the YOLOv4-Light algorithm,and analyzes and compares the performance of the algorithm.The experimental results of algorithm performance show that the proposed YOLOv4-Light-based forest fire detection algorithm achieves better accuracy and realtime performance.Compared with YOLOv4,YOLOv4-Light has a smaller number of network parameters(12,615,535 parameters)and higher accuracy(86.03% m AP),and it can also meet the real-time detection in forest fire video detection.Compared with the conventional algorithm and YOLOv4,YOLOv4-Light has higher accuracy(frame detection rate of 90.30%)and target integrity.(3)A GXLD forest fire detection algorithm is designed and implemented based on YOLOX-L-Light and the dark channel defogging method.The experimental results show that the m AP of YOLOX-L-Light is higher than other models in terms of accuracy with all kinds of APs.It has less number of parameters than YOLOX-Tiny in terms of number of parameters(3,994,609).The results of the ablation experiments show that the proposed lightweight method can effectively reduce the network parameters and improve the network performance.The video detection performance experiments show that the forest fire detection performance of GXLD is better than the mainstream YOLO algorithm in terms of accuracy(m AP of 87.47% and frame detection rate of 92.17%),target integrity,and average confidence,and can achieve more accurate forest fire detection in foggy environment(frame detection rate of 86.38%for foggy video).