Pixel-wise Dense Matching Of Aerial Images With Multi-Conditional Constraints

Stereo matching is a core critical issues in both computer vision and photogrammetry, the former whose study object is usually small indoor images generally results in a good matching by employing more extensive research ideas as well as building a complete mathematical model; the latter whose study object is the large-scale aerial remote sensing images usually results in a relatively poor matching by implementing a correlation based on the definition of matching. The matching accuracy of current domestic DSM production can reach about5-10times the resolution of the image, which cannot meet the needs of applications such as urban three-dimensional modeling, therefore the research of pixel by pixel matching should be carried out. Stereo matching according to different mathematical models used to be divided into local matching, global matching and other matching, in which, the best one is global matching that builds a complete energy function and implement the combinatorial optimization, the one of highest efficiency is the local matching which directly takes the local optimal solution of correlation, but the common idea which is shared by other matching algorithms is to introduce more constraints (such as seed points, reliable point triangulation, etc.) as a priori information, combined with the mathematical model explicitly (such as a dimensional energy function Bayesian theory) or specific algorithms implicitly (such as regional growth) to take advantage of this information to estimate the disparity. Compared to the local algorithms, these algorithms can make better use of prior information, so better matching results can be achieved; compared to the global algorithms, these algorithms’calculation methods are simpler, thus have a more practical computational efficiency. Therefore, pixel by pixel matching of aerial remote sensing images should be the most worthy to use of other matching algorithms. The main contents of this paper include:1. A systematic summary with a deep analysis for the theoretical basis and algorithm principle and characteristics of stereo matching is carried out. Besides, by summing up both the common problems faced by stereo matching and general resolutions or measures respectively, we summarize that the stereo matching algorithm should follow the two basic principles, namely on the one hand, as much priori information as possible should be incorporated to improve the matching effect, on the other hand images content itself should be taken into consideration to allocate computing tasks rationally in order to improve the efficiency of stereo matching.2. This paper analyzes the nature of classification methods for stereo matching similarity measure. Abandoning the traditional classification of gray matching and feature matching, we hold that stereo matching similarity measure is basically used to measure the similarity between two matching primitive metrics. So according to different matching primitives, stereo matching is divided into the dense matching and feature matching. Dense matching similarity measure including parametric ones, the non-parametric ones and those based on mutual information have been studied and characteristics of each matching similarity measure are summarized, in which way the basis of the specific implementation for dense stereo matching is laid.3. Based on the introduction section for the analysis of stereo matching problems and existing algorithms, we study three matching algorithms in computer vision which use a variety of different ways to incorporate prior information and result in good performance effectively. In particular, Firstly we research and improve PRPGM algorithm and make the calculation speed further accelerated. And then two algorithms based on reliable point triangulation matching have been also studied. At last we study the one-dimensional piecewise energy function model and semi-global stereo matching algorithm.4. For the post-processing of disparity estimation, research consists of the mismatch detection, holes fill on disparity map, sub-pixel interpolation of disparity map is carried out. We use Left-Right-Check, peaks removal and median filtering to detect mismatch. Specifically we proposed a method to fill the holes on the disparity map based on the region growing matching. We also analyze the necessity of the sub-pixel interpolation for disparity map and contrast the interpolation methods of linear function and quadratic function.5. A pixel by pixel matching method for aerial images based on the rough DSM is proposed. Firstly we convert the rough DSM into disparity map, and then with the disparity map as the constraint an image block matching will be carried out. Matching algorithm for blocks matching is chosen based on the information entropy of initial disparity map. With a comprehensive utilization of PRPGM, ELAS and SGM algorithm we have achieved a good pixel by pixel matching for aerial images consequently.