Research On Cross-view Geo-localization Method Based On Drone And Satellite Image

Image geo-localization is the study of estimating the geographic location of an image.Given a query image of an unknown location,the reference image most similar to the query image is retrieved from the database through image retrieval technology,and the GPS tag of the reference image is used to estimate the location of the query image.Drone visual positioning and navigation technology based on image geo-localization can assist traditional technologies such as GPS and INS,and also has the potential to independently complete positioning and navigation tasks,which has great application value and research significance.This thesis mainly discusses the inter-category similarity and intra-category difference brought about by the sharp change of viewing angle between drone images and satellite images,and analyzes and researches from multiple perspectives such as image preprocessing,feature extraction,and retrieval speed optimization.The main research work is as follows:(1)Image preprocessing and dataset reconstructionSince the heterogeneous images are collected from different platforms and have great differences,in order to enhance the unity of the cross-view images and improve the accuracy and efficiency of subsequent algorithms,this thesis preprocesses the cross-view images according to the difference in color style.At the same time,the Sf M algorithm is used to estimate the pose information of the drone image,and the dataset is reconstructed through the pose information to lay the data foundation for subsequent experiments.(2)Image geo-localization algorithm based on multi-scale ring partitionAiming at the problem that the existing feature extraction methods tend to ignore the relationship between features and the potential salient features are easily covered up,this thesis proposes an image geo-localization algorithm based on multi-scale ring division,using Res Net-50 as the backbone network to construct multi-view and multi-branch siamese network,through the multi-scale ring partition strategy to obtain coarse-grained,medium-grained and fine-grained features,and through Ghost Vlad to obtain the relative distribution of features,retain the basic structure of different types of geographical scenes,strengthen the correlation between features to complete the expression Overall properties of the image.Experimental results on the University-1652 dataset show that the algorithm is superior to most mainstream methods,and can better complete cross-view image geolocalization tasks.(3)Fast image geo-localization algorithm based on vision perception and deep hashingIn order to bridge the intra-class differences between drone images,reduce feature redundancy in the feature extraction process,and solve the problem that it is difficult to ensure accuracy and real-time performance at the same time,this thesis proposes a fast image geo-localization based on field of view perception and deep hashing algorithm.The algorithm first uses the Mobile Vi T network to build a lightweight visual field perception model to judge the visual range of the drone image;then,it uses the visual field information of the image to input to a specific branch,and executes a ring division strategy with targeted settings in the specific branch.Improve the discrimination of image features;finally,improve the image retrieval speed through the deep hash algorithm.The experimental results on the University-1652 dataset show that the method in this thesis significantly improves the retrieval speed with less loss of accuracy,so as to better balance real-time and accuracy.