Aerial Vehicle Detection Based On Depth Features

With the rapid development of China's economy and urbanization in recent years,urban vehicles are increasing and transportation networks are becoming more and more complex.At the same time,the concept of intelligent transportation came into being.With the advancement of unmanned aerial vehicle(UAV)technology,there are more and more application scenarios for aerial vehicle detection combined with UAV.However,there are many problems in the current aerial vehicle detection.The vehicle in the aerial image is very small,and the traditional detection method is worrisome.Most of deep learning detection methods have a general effect on the recall rate and the accuracy rate.And a large amount of label data is also a problem.This article has done the following work on the above issues:1.This paper firstly discusses the existing methods and problems of aerial vehicle detection,and establishes a Nanjing aerial vehicle dataset through UAV shooting.The dataset includes a total of 376 aerial photographs of Nanjing Shiyang Road and 101 aerial photography images of Nanjing Donglin Road,with 3732 vehicles and 1930 vehicles respectively.2.Based on the Faster R-CNN target detection framework,this paper designs a super-feature layer for the detection of small target vehicles in aerial images.Based on the VGG16 network,the Concat feature fusion model and the Eltwise feature fusion model are proposed by combining shallow features and deep feature.The final super feature layer is obtained by combining these two models.At the same time,based on the size characteristics of aerial vehicles,the anchor box generation in the RPN network is modified.Experiments show that our method has better detection results on aerial vehicle data.3.Based on pix2 pixGAN,this paper proposes a multi-condition constrained generative adversarial networks to generate vehicles with positional annotation information in real aerial scene images.The noise region preset in the image is perfectly converted into a vehicle image by constraining the fitting of the background and the generation of the vehicle in the image by respectively setting up a multi-discriminator in the generative adversarial networks.Comparative experiments show that our method has better vehicle generation effect and can improve the training results of aerial vehicle detection model.4.Based on aerial vehicle detection,an automatic traffic statistics system was designed and implemented.In the UAV aerial video,the system detects the positioning of the aerial vehicle through the vehicle detection model,tracks and counts the traffic flow information on each lane,and displays it on the system interface.