Research On Convolution Neural Network For UAV Detection In Low Altitude Airspace

The use of UAV aircraft in various fields has become more and more common.However,due to the unsatisfactory UAV regulatory system,countries have frequently threatened public safety.It is particularly important to develop an automatic UAV detection system.Radar,infrared,acoustic wave and other detection methods are applied to the UAV detection system,but there are detection defects under certain conditions.To this end,this paper proposes the use of image recognition,based on the deep learning method to real-time detection of UAV.This paper combines the detection of UAV from the following aspects:1.In order to better obtain the design idea of UAV detection algorithm,this paper combines the Deformable Part Model algorithm to implement the UAV detection algorithm research.The traditional sliding window method is used to realize UAV recognition,and the tracking algorithm is used to realize real-time tracking.Then,the research on the nominated neural network model is carried out.We improve the network architecture and improve the overall recognition performance and the small target UAV recognition effect,and finally realize the UAV detection.Provides better research ideas for subsequent high-precision and high-efficiency UAV recognition algorithms.2.This paper adopts deep learning method to realize the detection of UAVs by using deep convolutional neural network,which solves the problems of weak generalization ability,poor robustness,poor real-time performance and low accuracy of traditional algorithms.Since the standard data set of the open rotor UAV was not found,the camera was used to capture the UAV video sequences with various flight attitudes,different flight altitudes and different environmental backgrounds,and then a data set containing 11797valid UAV samples was produced.At the same time,the rotor UAV data set and source code created in this paper are disclosed.3.This paper combines the open source Darknet deep learning network framework,and improve the YOLOv3 model,including sample enhancement,bounding box design,loss function design,multi-scale research,prediction frame filtering,etc.,so that the UAV detection meets the real-time requirements and Detection accuracy.At the same time,it is proved that the convolutional neural network has the characteristics of good real-time and high accuracy in the detection of UAV.Compared with the most advanced detection algorithms,the improved method satisfies the real-time requirements while exhibiting excellent accuracy.In this paper,the test verification and comparison experiments are carried out on the UAV detection.The improved convolutional neural network UAV detection algorithm shows the best results compared with the current best algorithms.After evaluation,the₅₀ and₇₅ values were 1.000 and 0.850.Respectively,the inference time is 0.030seconds.Finally,the research results of this paper are applied to the engineering project,and the software platform of the UAV detection system is built.