Object-Oriented True Ortho Image Rectification Method

Digital orthophoto Map (DOM) is an important geographic information products by using DEM (Digital Elevation Model) to eliminate the projection distortion of camera tilt and terrain. True ortho-photo Map (TDOM) is a kind of real ortho image map, maintaining the map geometry and containing the pictorial information of aerial photographs. It has played an increasing important roles in urban applications, instead of traditional DOM with the imperfections such as tilt and occlusion caused by man-made and natural objects.After reviewing the existing TDOM generation methods, this thesis summarizes two key approaches which are different from the traditional DOM generation. Both object and terrain displacement distortion are corrected in differential rectification by using DSM (Digital Surface Model) instead of DEM. And the occlusion areas in perspective images are compensated to the greatest extent by using occlusion detection algorithms and resampling from overlap images. However, those image rectification and texture resampling process under the pixel-oriented strategy, causing inaccurate geometric feature and imperfect texture structure of man-made object, occlusion recovery and shadow compensation are difficult to automatically process, requires a lot of manual intervention.Aiming at above mentioned problems, this paper proposes a novel object-oriented true-ortho rectification method, which implements a high-level object-oriented strategy instead of a low-level pixel-oriented strategy. There are three main steps:(1) Definition of physical objects and image object containing geometric and semantic information. In3-D physical space, all the objects in the horizontal plane of projection form a seamless surface comprising of a series of contiguous triangular facets, in this paper this surface is represented by Semantics constrained Triangulated Irregular Network (S-TIN). The physical objects are characterized by different aspects: property, outline, geometry and topological relation, which can be extracted from3-D city models and points cloud. The image objects are made up of several pixels with semantic descriptions of spectral information,2-D geometry and topological relations. They can be extracted by image segmentation, edge detection and texture clustering.(2) Establishment of a global hierarchical spatial index of image objects. After deriving visual correspondence relationship between3-D physical objects and2-D image objects, we use the index to organize all the information of physical and image objects efficiently and provide a foundation for high performance true-ortho rectification and optimized image sampling.(3) object-oriented true-ortho rectification and optimal image resampling. The true-ortho rectification is carried out based on the physical objects as S-TIN. And based on the semantic links in the global index, the image objects with the semantic information of visibility, integrality and radiometric features can be chosen for optimized sampling in order to adaptive handling the occlusion and shadows.The proposed object-oriented true ortho-image rectification algorithm is implemented using C++. We chose multi-view urban images of Yangjiang with corresponding DSM for the experiment. The experimental results revealed that the proposed method is able to generate true ortho-image maintaining typical geometry and radiation features, remove tiny blunders caused by complex building structure, and improve the degree of automaticity of the whole process.