Research On Tie Point Extraction Of Multi-View High-Resolution Satellite Images

Multi-view and high-resolution satellite images have become a hot issue in photogrammetry product generation and 3D reconstruction processing,and the quality of the product mainly depends on the processing of Aerial Triangulation.Tie point extraction and Bundle Adjustment are the two most important links in Aerial Triangulation.The accuracy,distribution and overlap of the tie point will directly affect the quality of the Bundle Adjustment and thus the accuracy of direct image positioning.However,there are large differences in resolution,radiation amount and viewing angle between high-resolution satellite images.These differences make the automatic tie point extraction of satellite image still a big challenge.Aiming at the problems of low point accuracy,mismatched tie point,and low overlap of tie points in satellite imagery,the image data is based on a large-scale benchmark dataset of multi-view stereo satellite images provided by the Johns Hopkins University Applied Physics Laboratory and panchromatic images of Resource-3 satellite,combined with the influence of the Bundle Adjustment,the automatic extraction of high-precision and high-overlap tie points for high-resolution multi-view satellites is studied as follows:(1)A method for optimizing the accuracy of tie points based on LAD solution is proposed.The initial corresponding relationship of feature matching is obtained by SIFT matching,and the initial value of the affine model of image grayscale information matching is accurately initialized from the position and direction information in the feature descriptor,so as to match the absolute difference of pixel grayscale in the window.The minimum absolute value(LAD)is used as the criterion to iteratively solve the affine transformation parameters and radiation correction values,which weakens the influence of noise in the window on the solution,and finally obtains the refined points of successful convergence.The experimental results show that there is little difference in LAD matching and LSM performance between satellite remote sensing images with less significant differences in deformation and radiation,and the matching accuracy is improved by 0.2 pixels compared with SIFT.Among images with significant differences in deformation and radiation,the LAD matching accuracy is better than LSM by 0.1 pixel,but the convergence success rate is slightly lower than that of LSM,and the accuracy is 0.2 pixel higher than SIFT matching.(2)A threshold parameter-adaptive method of matching gross error elimination with local constraint is proposed.According to the accuracy of image matching feature points,the uncertainty of distance local constraints and topological local constraints are deduced,so as to achieve a more accurate effect of detecting gross errors.At the same time,the threshold parameter can be adaptive with the accuracy of local constraints between feature points.Finally,a good gross error detection and removal effect is achieved.The experimental results of test data show that in the case of sparse initial matching and different false matching rates,the method in this paper can achieve 100% accuracy and recall rate,and completely eliminate gross errors.Experiments on real multi-view and high-resolution image data show that,in the case that the false matching rate of the initial matching is high and the density of points is large,the method in this paper can accurately eliminate the gross errors and matching points with low matching accuracy,the correct rate is higher than 99.5%,and the RMSE is within 0.5 pixels.(3)A method for automatic extraction of tie points with high overlapping degree is proposed.Firstly,the traditional pairwise matching strategy is used to complete the initial feature matching of multiple images,and based on the results of the reference image,an index table is established to track the features.In the process of feature tracking,the incorrect tie points are eliminated and the tie points with high overlapping degree are screened,and finally the tie points with high overlapping degree are obtained.Tie point extraction of 47 scene images experiments show that the average overlap degree of extracted tie points can reach about 10,while the average overlap degree of image connection points is about 21.When the average overlap degree of connection points drops to between 3and 4,the average RSME by reducing 0.2 pixels,the RMSE of images with high average overlap of junction points is also reduced by about 0.1 pixels.