Cross-domain Aerial-road Registration And Matching Based On Deep Learning

The registration and matching between aerial images and road landmarks are the key technologies to realize UAV autonomous positioning in urban areas when GPS is not available.The UAV positioning technology based on road landmarks is essentially the application of image registration and matching technology in the field of UAV positioning and navigation.Its purpose is to achieve real-time high-precision positioning of UAVs under GPS-denied environments.In view of the shortcomings of the traditional multi-stage image registration algorithm based on manual feature extraction,such as low accuracy,poor robustness and poor timeliness,this thesis proposes novel algorithms based on end-to-end learning for road navigation mark registration and matching,and adopts deep learning technology to directly establish the mapping relationship from input to output.This thesis makes an in-depth study on the registration and matching between aerial images and road landmarks and solves some problems of traditional algorithms by designing and improving the algorithm model.The main contributions of this article are as follows:1.In view of the shortage of currently available aerial image data sets,a large-scale aligned aerial image and road landmark data set was created.Since the landmark is usually defined as a region with distinctive features,a road with a relatively complex shape is selected as the landmark,an aerial image with the same position as the road landmark is intercepted as the aerial data set,and the aerial image is randomly rotated and transformed and carried out mark.2.With the help of deep learning technology,a novel cross-domain road navigation registration model based on attention mechanism is proposed to achieve high-precision registration of color aerial images and binary vector road maps.The model uses a two-branch asymmetric neural network structure with partial parameter sharing to map the input image to the same feature space to achieve cross-domain feature expression.Considering the sparse road features,the model introduces a multi-branch attention module on the basis of deep feature matching to filter out wrong feature point matching pairs,thereby improving the accuracy of navigation mark registration.3.Considering the limited storage space of the UAV onboard computer,a lightweight cross-domain road beacon matching algorithm is designed,which can achieve end-to-end fast and high-precision road beacon matching.The matching model is based on the above registration model.By performing lightweight processing on the registration model,the calculation complexity of matching is reduced,and the matching efficiency is improved.