Research On Forest Fire Detection Method Based On UAV Dual-Light Image

Forest fire detection is one of the important links in the task of forest protection.Unmanned aerial vehicle(UAV)has the characteristics of flexiibility,convenience and speed.Therefore,using UAV to detect forest fire is one of the hotspots of current research.Visible images contain limited information,which limits the ability to detect forest fires.In this paper,using DJI’s M300 RTK UAV platform and H20T sensor to collect visible and infrared images,forest fire detection method based on UAV dual-light image is studied.The main contents include.(1)A fast registration method for visible and infrared images is presented.H20T is equipped with visible and infrared cameras with different positions and focal lengths,and the image size,target size and location collected are not the same.Therefore,the two images are first transformed to the same size.Then,the difference of target size in the two images is eliminated by using the scaling factors in the X and Y directions.Finally,the target location difference is eliminated by using the offset pixels in the X and Y directions of the two images.In this way,the registered dual-light images can be obtained quickly,which lays a foundation for the study of fire detection methods based on dual-light images.(2)A dual-light forest fire detection method based on multicolor spatial rules is proposed.First,the color threshold segmentation rules for flame and smoke are designed by using the gray scale channels in infrared images and the Cr,Cb,H,S and V channels in visible images.Then,the mask binary map of flame and smoke is obtained by bits and operations to detect the flame and smoke.Finally,the dual-light video data collected in this paper at different flight altitudes,different scenes and different times of flame and smoke are used to verify the experimental results.The accuracy of flame detection is more than 92%,and the accuracy of smoke detection is about 80%.In addition,the detection speed at ground and airborne ends is tested,reaching 200 and 100 FPS respectively.The experimental results show that the detection speed of this method can meet the real-time requirements at both ground and on-board.(3)A dual-light forest fire detection method based on multi-color feature and uniform rotation LBP is proposed.Firstly,the registered double-light image is divided and extracted.The average image block values of the gray-scale channels in the infrared image and the Cr,Cb,H,S and V channels in the visible image are taken as color features,and the image blocks of S,Cr-Cb and gray-scale channels are rotated uniformly and LBP is taken as texture features.Then,the artificial neural network is used to classify the dual-light image blocks to detect the flame and smoke.Finally,the experimental validation is carried out on the data set established in this paper.The results show that the F1 score of flame detection can reach 97%,the F1 score of smoke detection can reach 98%,and the detection speed at ground and on-board is about 10.8FPS and 5FPS,respectively.The computational complexity of the texture features makes the method slower but more accurate and more suitable for use on the ground.(4)A dual-light forest fire detection method based on YOLOX is proposed.Firstly,for the registered infrared image and visible image,different weighting coefficients are set to overlay the two images to get a dual-light fusion image.Then,the fused image is input to the YOLOX-1 network for flame and smoke detection.Finally,experiments are carried out on the ground and on-board ends,and the detection accuracy of mAP reaches 97.04%,the inference speed of single image reaches about 60FPS,and the inference speed of video stream reaches about 36FPS.On-board end iChrest 2-s,with model pruning and TensorRT acceleration,the single inference speed of the compressed model is about 34.6 FPS,the video stream inference speed is about 16FPS,and the detection accuracy mAP reaches 93.76%.The experimental results show that the detection accuracy of the models on the ground and on-board is high,and the detection speed can meet the real-time requirements.