Research On UAV Object Recognition Based On Deep Learning

With the development of multi-rotor UAV technology and the further increase in the marketization rate of UAVs,a large number of convenient and functional multi-rotor UAVs are used in daily life.Due to users’ lack of aviation safety awareness and loopholes in the supervision of relevant departments,many airports exist the phenomenon of "black flying" and "overflying" of UAVs,resulting in grounded flights and flight accidents happen frequently.Airport safety is an important prerequisite for ensuring aviation flight.Among the low,small and slow objects near the airport,the impact caused by drones is particularly serious.Although there are technical means such as radar to detect foreign objects,they cannot accurately and effectively identify and drive away illegally invaded UAVs.How to quickly and effectively identify UAVs in airport airspace has become one of the research focuses.In response to the above problems,this article adopted the current mainstream algorithms in the object recognition field,and developed 3 Res structures based on Retinanet and 4 models of YOLOv5 to identify and locate UAV samples.Through training,testing and inferrence,the two algorithms’ detection effect on the human-machine data set were compared,so as to realize the fast and accurate identification and positioning of the UAV object in the actual project.The experimental results show that in the Res34,Res50,and Res101 structures of the Retinanet algorithm,the Retinanet algorithm of the Res34 structure has the highest MAP value,reaching 92%,the Res50 structure has the highest MAP value of 85%,and the Res101 structure has the highest MAP value of 87%.For large-view and small-target UAVs 200 m away,the recognition accuracy can reach up to 47%.At the video inference speed,the single-frame time was 1321.54 ms,1433.76 ms,and 1883.42 ms respectively.Therefore,the comprehensive performance with Res34 structure is the highest.The MAP@0.5 values of the four network models of the YOLOv5 algorithm YOLOv5 s,YOLOv5m,YOLOv5 l,and YOLOv5 x on the UAV data sets were 99.5%,98.5%,98.2%,and 99.5% respectively.And the accuracy of longterm target recognition was 88%,89%,87%,and 87% respectively,and the video inference speed was 1.4ms,3ms,4.9ms,9.4ms.So,the YOLOv5 m model has the best performance in UAV recognition.In terms of the fastest inferrence speed,the FPS of YOLOv5 is 1000 times that of Retinanet,and for small distant targets,the recognition accuracy of YOLOv5 is twice that of Retinanet.In terms of algorithm transplantation,it can be used for airport equipment to quickly locate and detect long-range UAV targets,so as to take countermeasures effectively,which is of great significance to ensure the safety of airport operations.