| With integrated control and development of related technologies,such as artificial intelligence as well as a decline in the cost of the product,in the field of consumption level,the volume of the unmanned aerial vehicle(uav)is more and more small,more and more high performance and control more and more simple,gets gradually the favour of consumer,the market lower threshold also enables ordinary people to anywhere at any time to enjoy the joy and convenience of unmanned aerial vehicle.However,if UAVs are wantonly released and illegally used,they will also bring about major security risks to aviation and other fields.Therefore,how to quickly and accurately identify UAVs in the air is particularly important.This thesis is devoted to the research and implementation of UAV detection based on deep learning,mainly to solve the problems of backlit object detection and small object detection.Factors influencing the effect of deep learning-based algorithms mainly include data,network structure and loss function,etc.Previous studies mainly focused on the optimization of network structure and loss function of algorithms,while this thesis focuses on data.Aiming at the difficulty in the identification of backlit objects and the low accuracy of the identification of small objects in the process of UAV identification,this thesis gives the corresponding solutions.The main innovations and contributions of this thesis are as follows:1)A Bi-CycleGAN algorithm with binarization was proposed to improve the generalization ability of the detection model to the backlit UAV.For complex data set background image binarization,cannot be directly used for conversion of real images,the introduction of binarization Bi-CycleGAN algorithm,reduce the degree of dependence on data sets original CycleGAN algorithm,to be able to live images of unmanned aerial vehicle into black,realize the fitting of the object of backlit unmanned aerial vehicle,change the data set backlit uav object distribution is not balanced,and enhance object detection model of backlit uav identify generalization ability.2)A data enhancement method based on image pyramid is proposed to improve the accuracy of detection model for small object detection.This method can combine images of different sizes generated by a single image pyramid into a new image,and the new combined image can increase the number of small and medium-sized objects in each image without changing the object information in the original image,which overcomes the complexity of the original method.By generating a new combined image from a single image,the problem of unbalanced distribution of small objects in the training set can be changed,and the detection accuracy and generalization ability of the object detection model for small objects can be improved.3)The object detection algorithm is transplanted to the edge processor,so that the edge server can realize the detection of the UAV.The detection effect of the UAV is verified by experiments. |
