Research On Target Recognition Technology Of Small UAV Based On Deep Learning

With the widespread use of small UAV in the military and civil fields,the battlefield threats and negative effects in the civilian field brought by small UAV are gradually emerging.An anti-UAV system for small UAV and a target detection as well as recognition technology for the low-altitude area defense system have become one of the research hot spot in small target recognition field.Supported by the Xi’an Science and Technology Project,a small target UAV recognition technology based on deep learning algorithm and an algorithm integration technology based on embedded platform were studied to deal with the low-altitude identification and early warning of small UAV target application requirements in this paper.The main research contents are as follows:1)An analysis of deep learning network structure was carried out and a data set for target recognition was built.An analysis of deep learning network structure was carried out and a data set for target recognition was built.Based on the research of common deep learning frameworks,the selection of deep learning framework was completed by combining the characteristics of target recognition algorithm and the needs of evaluation indicators.By analyzing the movement scenes of small UAV,the relationship between UAV image imaging quality,image diversity and training model was studied.Combined with the technical problems of low,small and slow target recognition,the training,testing and verification data set of deep learning network for small UAV was built.2)A target recognition method of small UAV based on YOLOv3 network was studied.By comparing and analyzing the one-stage deep learning algorithm model,the method was proposed to solve the difficult problems of UAV target recognition and positioning,such as the diversity of UAV movement background and small size of flying target.Aiming at the problem of data imbalance caused by complex background and small number of samples in multi-target recognition,SMOTE algorithm was adopted to optimize the data set and improve the performance of YOLOv3 model.The experiment results show that the precision,recall,and mAP of the YOLOv3 model have been greatly improved after optimization.3)A deep learning target recognition technology based on embedded platform was studied.The NVIDIA Jetson TX2 hardware platform was used to complete the algorithm transplantation and solve the problem of poor real-time monitoring of UAV targets in real life.Based on the software architecture of Nvidia Jetson TX2 embedded platform,the operating environment of the small UAV target recognition system was built.According to the algorithm configuration process,the TensorFlow framework was configured in turn,and the YOLOv3 algorithm was compiled to analyze TensorRT technology to improve the system reasoning speed.Through experimental verification,the system miniaturization design of the small UAV target identification and positioning system constructed in this paper has been completed,which achieves the expected goal of improving the target positioning accuracy as well as identification accuracy under complex background.The system is of great significance to the real-time monitoring and management of anti-UAV system.