Research On Dense Matching Methods Of Aerial Stereo Image Pairs

In the field of aerial photogrammetry,stereo matching based on aerial images has always been a hot topic.This paper presents a stereo matching method which is able to generate a relatively good result in a short time to deal with big aerial images.Many stereo matching methods were presented in the field of aerial photogrammetry and computer vision in recent years.However,these methods have not been systematically compared based on aerial images.This paper compares several methods by using a considerable number of aerial image pairs and reference DSM,which we wish to help the subsequent researches to a certain extent.This paper mainly focuses on following aspects:1)A stereo matching method based on fast guided filtering is proposed.This paper proposes a fast local matching method whose result is roughly equivalent to the global and semi-global matching methods and has certain advantages in performance.The method uses the sparse matching result and the result of reliable matching on a certain spatially spaced grid point to estimate the disparity distribution on the image and obtain the candidate disparity sequence.After calculating the Census cost for each candidate disparity pixel-wise,the guided filter is performed on cost aggregation step.Finally,the WTA method is used to select the optimal disparity and form a disparity map.2)Classification of stereo matching methodsStereo matching methods are classified into three categories based on how they use the smoothness assumptions.Local stereo matching uses the smoothness assumptions implicitly in the matching window;Global stereo matching uses the smoothness assumptions explicitly in the energy function;Methods based on deep learning no longer care smoothness assumptions.This paper mainly studies the cost aggregation of local stereo matching,and divides it into adaptive window methods,adaptive weight methods and methods with slanted support windows according to the way of solving the conflict between the reality and fronto-parallel assumption.This paper also studies a cross-scale cost aggregation method designed to solve problems in textureless regions.3)Comparison of stereo matching methodsBased on aerial images and the DSM of vahingen provided by ISPRS this paper designs and implements the stereo matching evaluation program.The program can evaluate the matching quality indicated by several indicators from the object space and image space.In this paper,we use the program to compare eight different types of methods on the dataset and several image pairs of the typical objects.These methods include local stereo matching methods,semi-global matching methods,global matching methods,and methods based on deep learning.The experimental results of this paper show that: 1)The improved semi-global matching method and the global matching method using the semi-global matching to obtain the initial value perform best in all tested methods.The precision of improved semi-global matching method in a flight strip,an image pair of city,an image pair of field and an image pair of forest is 1.76 m,1.33 m,0.48 m and 2.18 m,respectively.That of global matching method based on the initial value obtained by semi-global matching is 3.25 m,1.86 m,0.44 m and 3.86 m respectively.2)The modified fast stereo dense matching method based on guided filter proposed in this paper is the best local method of comparison,and its quality exceeds some global matching methods;3)Aerial images like the weak texture,no texture,repeated texture,depth discontinuity,depth change complex areas,are still challenges to stereo matching.4)GC-Net is no better than the traditional methods if it is not re-training for target dataset.