Urban 3D Building Model Applied to True Orthoimage Generation

Three dimensional (3D) building models are increasingly necessary for urban planning, tourism, etc. How to effectively describe the architectural objects becomes a key point due to the fact that the urban buildings extremely vary in height, sizes, shapes, textures, etc. This paper presents a model for exactly describing urban 3D buildings for large-scale urban true orthoimage generation. This method is based on CSG (Constructive Solid Geometry), belonging to volumetric representation in computer graphics. This method is well suited to describing complex shapes, which can be composed by a set of primitives. Within this proposed approach the buildings are described by combining a set of basic primitives, such as box, wedge, and rectangular pyramid. This representation model is particularly useful for urban true orthoimage generation because a complex building in this model can be partitioned into many simple building parts, each of them corresponding to a basic building model. We implemented this work using the ground plan information according to digital surface model. The ground plan of a building is divided in rectangles, arcs, and circles each of the primitives representing the ground plane of a building part. The primitives are combined by means of the Boolean operations union, intersection, and difference. So, the buildings will be described as a CSG tree. Our experimental result in Downtown, Denver, Colorado demonstrated our method can effectively and exactly represent the complex buildings, and produce high accuracy when applied in urban true orthoimage generation. INTRODUCTION Theoretically, the digital orthoimages should be a spatially accurate image with ground features represented in their true planimetric positions. However, the algorithms and procedures in traditional digital orthoimage generation did not consider the spatial objects, such as buildings, resulting in the spatial objects of orthoimage in urban areas are distorted from their true positions (Zhou et al., 2005). In order to orthorectify a building to its correct, upright position, the building must be represented as part of the surface to be rectified. Therefore, an exact digital building model, which describes the building structure, threedimensional coordinates, topologic relationship, etc., is required. Aiming at this purpose, the researches about automatic or semiautomatic building extraction (Gűlch, 1996; Förstner, 1996; Vosselman, 1999; Heuvel, 2000; Gerke, 2001) become a key problem. Many different approaches and algorithms were addressed with different source image type, terrain complexity and field of application etc. In recent decade years, CSG model, as a kind of model-based building extraction, is commonly used for building extraction in the field of computer vision and photogrammetry (Braun et al., 1995; Englert and Gűlch,1996; Lang and Förstner, 1996; Tseng 2003), because of its flexibility for the representation of buildings and its diversity for containing object constraints and classification. CSG model is composed of a combination of volumetric primitives. It is possible to construct a complex model with a small set of primitives, depending on the required detail. A primitive is a simple solid model to determine the interior geometric properties of a building, and is associated with some transformation parameters. The combination of primitives can finish via Boolean set operations, such as union, intersection, and difference. 1st EARSeL Workshop of the SIG Urban Remote Sensing Humboldt-Universität zu Berlin, 2-3 March 2006 2 PARAMETERIZED CSG BUILDING MODEL CSG modeling is time-consuming process. Tseng presented a semi-automated building extraction method from aerial images (Tseng 2003). The fitting CSG model is selected interactively by the operator, and then optimal model-image fitting is performed automatically with a least-square algorithm. Gűlch addressed a semi-automatic method of fitting parametric models to multiple overlapping images and then combining them into the whole model. They can reach about 20 seconds for the modeling of a volumetric primitive (Gűlch 1999). However, for large urban area, it is impossible to construct CSG model with human interaction. a) Data structure The generation of a 3D urban building model is a rather challenging task, because different applications (e.g., city planning, communication design, tourism, pollution distribution, etc.) require different data types and manipulation functions (Breunig 1996, Graz 1999, Grun and Wang 1998, Zhou et al. 2000). For the purpose of true orthoimage generation, the data structure to be developed in this paper requires not only the fitness for generating the DBMbased high-quality orthoimage, but also easily creating, storing, designing, analyzing and querying city objects for orthoimage-based urban applications (Zhou et al. 2005). CSG model is composed of many CSG primitives. Any 3D primitives come from 2D basic graphs with different parameters of Heights. This paper presents the technique of three-level data structure. The first level is 2D basic graphs with parameters, such as rectangle, circle, triangle, polygon etc. Each graph has respectively parameters of its own, such as length, width, radius, position of point etc. The second level is 3D primitive. 2D basic graph with height parameter becomes 3D graph. For example, rectangle with height is box; circle with height is cylinder. Box, cylinder, cone, pyramid etc. belong to single primitives. Furthermore, we can obtain the multi primitives with different single primitives. The third level is building model. The combination of 3D primitives can generate the CSG building model. In this process, topological relationship should be considered for the convenience of data retrieve. Figure 1:Three-level data structure b) Automatic CSG model extraction From digital surface model, the edges of buildings can be detected. These line features are only the low-level depict, not the whole model depict. Feature grouping should be applied by grouping these isolate, partial vector features into the whole semantic coherent symbol structures. Recently, feature grouping based on similarity relationship is commonly adopted. In this method, the geometric relationship (such as position, direction) and area property (such as color, gray value) are taken into account. Taking the line feature for example, this paper quantitatively depicts the similarity relationships of line feature. Six parameters (i.e. distance of points, collinear difference, collinear distance, degree of overlapping, gray value and texture) are adopted to measure the combining probability of line features (Henricsson 1996). This paper only takes two of them for example to illustrate it. Length