Forest Fire Detection And Spatial Location System Based On Multi-task Learning

Forest resources are the valuable wealth of human beings,providing an important material basis for human production and life.In recent years,the frequent occurrence of forest fires,once the fires spread out of control,will cause great damage to the natural environment and threaten the safety of human life and property.In order to realize the automatic inspection of forest fires,as well as the early detection and early warning of the fire situation,this paper designs a forest fire detection system equipped with a UAV cruise.The system adopts a forest fire detection algorithm based on multi-task learning and a fire point target spatial localization algorithm based on binocular vision to solve the problems of low accuracy,high false detection rate,poor flexibility and limited detection information of the traditional forest fire visual detection scheme,which provides great help to the daily inspection of forest firefighting and information rescue.This paper revolves around three parts:forest fire detection algorithm based on multi-task learning,fire point spatial localization algorithm based on binocular vision,and forest fire monitoring system based on UAV cruise,and the main research work and research contents are as follows:（1）To address the lack of datasets in the field of forest fire visual detection,a high quality,broad category,and well-labeled forest fire detection dataset is prepared,containing data from major public flame datasets,actual flame scenes taken by relevant institutions,and laboratory simulated scenes.（2）A multi-task learning based forest fire detection algorithm model is proposed,which is divided into three tasks:forest fire target detection,forest fire semantic segmentation,and forest fire image classification,among which forest fire target detection is the main task.The algorithm model uses multi-task shared backbone to enhance the network model’s ability to extract forest fire image features.The shared feature extraction network enables multiple tasks to learn from each other to improve the performance and detection accuracy of their respective tasks and to achieve an implicit data augmentation effect.To further enhance the detection capability of the model,joint multi-task NMS processing is proposed,which enables the main task forest fire target detection to be influenced by subtasks and reduces the missed false detection rate,and the algorithm improves the m AP index by 1.6%and the m AR index by 0.8%.In addition,the forest fire detection algorithm network adopts a lightweight structure design to reduce the model inference time while improving the recognition accuracy,which is easy to deploy in the forest fire detection system to achieve real-time and efficient forest fire target detection.（3）To address the problem that the network model is weak in detecting small target flames that are semi-obscured and at the edge of the image in the actual detection,a data enhancement strategy based on random point image diagonal block exchange is proposed to enhance the detection capability of small targets in the actual scene,improving the AP＿s index by 5.4%and the AR＿s index by at 3.1%.The method allows small targets at the edge of the camera’s field of view,small targets obscured by branches and leaves,and small flames at the early stage to be effectively detected,reducing the missed detection rate and achieving early detection and early warning of forest fires.（4）For the lack of spatial information in the result of image target detection,binocular stereo vision technology is used to obtain the depth of the forest fire target and combine IMU and GNSS sensors to convert the forest fire target under the camera coordinate system into spatial coordinate latitude and longitude with an error range of 10^-5 magnitude to realize the spatial localization of the fire spot area and provide digital guidance and help for the rescue mission of forest firefighting.（5）A forest fire detection system based on UAV cruise is designed,which consists of UAV,Raspberry Pi main control,OAK-D camera and related sensor modules.The system transplants a forest fire detection algorithm model based on multi-task learning and a fire point spatial localization model with binocular stereo vision to an embedded high-calculus device to achieve real-time cruise monitoring.In the simulated real scenario test,a speed of nearly 20 FPS with89.34%frame accuracy was achieved.Compared with traditional forest fire human patrol,sensor monitoring and watchtower monitoring,it reflects the flexibility and automation of forest fire detection in the new era.The forest fire detection system designed in this paper is characterized by high flexibility,high recognition accuracy,fast recognition speed and good positioning accuracy,and the proposed multi-task forest fire detection algorithm and fire point spatial localization algorithm are easy to transplant and deploy.The system is suitable for the UAV inspection task of forest fires,has high reliability and mobility,and has good application prospects.