Research On Stereo Matching Algorithm For Aerial Images

Stereo matching, which is a process of finding and establishing correspondence between homologous points in stereo pairs, is an important problem in stereo vision. As principle of human stereopsis reveals itself, humans can sense and recognise this world via computer by simulating stereopsis. When homologous points correspond after stereo matching, people can reconstruct the scene they captured using principles of epipolar geometry and triangulation in photogrammetry. Stereo matching has been put into practice after development of several decades. People apply stereo matching in mass production of 3D scenes, orthographic images and precise topographic maps. Stereo matching is also applied in fields like entertainment, automatic driving and security.The history of researching stereo matching is roughly as long as the history of computer, and the period it booms corresponds to the period computer technology prospers. However, stereo matching is still a challenging problem to be solved. Currently, researches on stereo matching have been quite wide and thorough. New theories arise, and new algorithms keep emerging. To process aerial images with large scale and complicate structures is a problem that needs more concern. This article will try to solve this problem, consider the features of aerial images, and design an algorithm framework capable of processing large scale images with less computation and more expandability.The algorithm framework in this article will start with feature matching, which is quite developed by now, and generate a disparity map covering most pixels by a few steps. A sparse but precise homologous point collection can be obtained from feature matching, and a triangulated irregular network (TIN) will be constructed from the point collection as an initial estimation of disparity map. Finally, the renowned semi-global matching (SGM) will be used, to compute the pixel-level final result.Given the large scale of aerial images, this article will perform semi-global matching in two separate passes, and temporary tiling is used in the process. Images are processed tile by tile to save memory. In cost computation, this article uses a weighted sum of CENSUS and mutual information (MI), as they are insensitive to radiometric distortion.In the end, this article implements the algorithm using C++ and OpenCV library, and performs qualitative and quantitative evaluation on this algorithm framework using both the close range data from Middlebury test sets and actual aerial images from drones, including evaluation on accuracy, time and resource consumption.