Research On The UAV Remote Sensing Images Matching Based On Local Invariant Feature

Unmanned aerial vehicle remote sensing system has the advantages of flexibility,fast response and small area detection capability.Now it is one of the important ways to obtain high-resolution remote sensing images quickly.The technology has been widely used in the surveying and mapping geographic information field,especially in emergency surveying and mapping provision.However,compared with the traditional satellite and aerial images,geometric transformation in the Unmanned aerial vehicle(UAV)images are more complicated,which makes the image matching processing more difficult,mainly in the constructive description of the robustness and real-time compatibility is relatively poor and so on.Meanwhile,the small and narrow coverage of UAV images and various types of image features cause texture similarity or texture lackage problem,which adds more difficulties to image matching technology.Focusing on these matching difficulties of UAV images,this thesis mainly studies the fast and robust descriptor construction method and the matching strategy suitable for texture similarity or lack region.The main work is listed as follows.1.According to the classification of existing image matching descriptors,the representative feature descriptors are selected to analyze the characteristics and application advantages of various descriptors and validation tests are carried out.The results show that the existing feature descriptors are difficult to balance the robustness and real-time needs of UAV image matching.2.Aiming at the limitation of existing feature descriptors,a method based on convolutional neural network(CNN)learning feature descriptor is proposed for UAV image matching.Considering the complicated transformation of UAV images,this thesis presents a triplet network training method using multi resolution samples.3.Aiming at the problem that CNN-based feature descriptor learning method mining "hard samples" leads to increased computational cost,the method of exchanging the reference sample among three samples is proposed.The experimental results show that the descriptor obtained by this method has good robustness and real-time performance,and it is better in UAV image matching application.4.For UAV images with similar textures or short of textures,the image feature is hard to distinguish,which increase the number of mismatching,Fourier-Mellin phase correlation method based on frequency domain is introduced to calculate the geometric parameters between the two images,which provides the geometric constraint for the image matching.The experimental results show that the matching strategy improves the number and success rate of image matching.