Research On Sound Localization And Image Recognition Of Multi-rotor UAV

Multi-rotor drones have developed rapidly in recent years and are widely used in many fields,such as agriculture,transportation,traffic monitoring,and aerial photography.However,due to the widespread use of amateur multi-rotor drones,more and more public Security threats and social issues.For this reason,anti unmanned aerial vehicle(anti-UAV)technology is very important.This article studies voice positioning technology to locate UAVs and image recognition technology to identify UAVs.The positioning technology of multi-rotor UAV plays an important role in the antiUAV system.Using machine learning-based wireless sensor network sound source localization method:(1)Establish a sound attenuation model to simulate the attenuation and distortion caused by environmental noise and changing environment.(2)Use the maximum likelihood method to solve the sound attenuation factor.The attenuation factor of DJI spark UAV calculated in this paper is 0.13.(3)Five machine learning algorithms are used to estimate the coordinates of a single UAV.These five algorithms are artificial neural network(ANN),naive Bayes,decision tree(DT),K nearest neighbor(KNN)and random Forest(RF).Both the received signal strength(RSS)based on sound energy and the difference between RSS(RSS＿D)are used as the input of the machine learning algorithm.Experiments show that in addition to ANN,the positioning performance of other machine learning algorithms is good.In the case of environmental noise,using RSS＿D as an input has better positioning accuracy than using RSS noise only.For the recognition of multi-rotor UAVs,this article adopts object detection related algorithms in the field of image processing.In this article,in order to improve the detection speed,a small drone detection method based on cropped deep convolutional neural networks is adopted:(1)Because deep convolutional neural networks have powerful feature extraction capabilities,the deep convolutional neural network with higher detection accuracy is obtained,and four latest deep volumes are tested.Product neural network detection algorithms: Retina Net,FCOS,YOLOv3 and YOLOv4.These four methods obtained 90.3%,90.5%,89.1% and 93.6% of m AP respectively.Choose YOLOv4 as the benchmark model,with a capacity of 245.8 MB and an FPS of 43.(2)Use different parameters to crop the convolution channel and shortcut layer of YOLOv4 to obtain a shallower model.Among these models,the cropped YOLOv4 model is pruned-YOLOv4,which has a 0.8-channel cropping rate and 24-layer cropping,achieving 90.5% m AP,69 FPS and 15.1 MB.pruned-YOLOv4 can increase the processing speed by 60% with less accuracy loss.The performance of tiny-YOLOv4 and pruned-YOLOv4 is compared through experiments.Considering the trade-off between speed and accuracy,since the m AP of tiny-YOLOv4 is relatively low by25.9%,prouned-YOLOv4 is chosen as the detector.(3)Enhance small objects to enhance the detection ability of small drones and compensate for the loss of accuracy.YOLOv4 has not made major improvements,but both tiny-YOLOv4 and prunedYOLOv4 have made major improvements,of which pruned-YOLOv4 The accuracy and recall rate were increased by 22.8% and 12.7%,respectively.Under the balance of accuracy and speed,the pruned-YOLOv4 enhanced by small targets has better performance in detecting small UAVs.Finally,it is deployed to embedded devices through Baidu Cloud,and the drones are identified and detected on the embedded side.