Towards Fully Automated 3D City Model Generation

Recent developments of efficient software systems have enabled an area covering generation of 3D city models. Still, there is a continuing need for an improvement of these systems to enable a further reduction of human interaction. Nevertheless, the development of new applications including the provision of effective tools for visualization and management of the collected 3D city models now becomes of growing importance. Within the paper our work is exemplary used to give an overview on the current state-of-the-art of 3D city model generation. Considering the reconstruction of building geometry, the paper focuses especially on automatic approaches which integrate 2D ground plan information with digital surface models. A heuristic subdivision of ground plans into rectangles as well as a rule-based reconstruction relying on discrete relaxation are discussed. In the second part further processing tools are discussed based on some application scenarios of the collected data. itization of the road network for navigation purposes. Thus, it becomes clear that any economically viable solution must rely on automatic or at least partially automatic methods. 2 RECONSTRUCTION OF THE BUILDING GEOMETRY 2.1 Reconstruction Systems Photogrammetric methods are well suited for the economic acquisition of 3D city models (Forstner 1999), making it possible to recover the structure as well as the dimensions. On the other hand, classical photogrammetric measurement is mostly point-based, which does not exploit the inherent structure of buildings and thus cannot be optimal economically. Standard automatic image matching techniques, developed originally for the measurement of terrain points, proofed to be not as effective for built-up areas due to the large discontinuities which arise at building borders and roofs. This contradiction has led to substantial research efforts (see e.g. Grun et al. 1995, 1997), and many reconstruction approaches have been proposed. Some coarse classification can be made with regard to the used data sources and the distinction between semiautomatic and fully automatic operation. Automatic reconstruction from aerial images (e.g. Haala 1994, Henricsson & Baltsavias 1997, Fischer et al. 1998, Baillard & Zisserman 1999) has shown promising results, however one has to note that often special image material has been used which is not available in general, for example large scale, multiple overlap, or colour images, or additional height models. Even then, the reliable extraction of buildings in densely build-up areas has not been demonstrated yet. Automatic systems working solely on the basis of digital surface models (DSMs) acquired by laser scanning have been reported (Brunn & Weidner 1997, Maas 1999). Since DSMs represent the geometry of the surface directly, they have advantages with regard to automated interpretation. They used to suffer from the low density of measured points, however this is no technical restriction anymore, since a density of several measured points per square meter is attainable. Also, the recording of multiple return pulses and intensity values is feasible nowadays, opening up new possibilities for automated segmentation techniques. Semiautomatic approaches have been reported both for imageand DSM-based systems. They can be divided into approaches which model buildings from a fixed set of volumetric primitives which are combined (e.g. Gulch et al. 1999, Brenner 1999) and approaches which build the topology of the surface directly (e.g. Grun & Wang 1998). 2.2 Reconstruction using ground plan segmentation A high percentage of buildings can be modeled using a small number of building primitives like flat boxes, boxes with saddleback and hipped roofs and other geometric primitives like cylinders and cones. Even if one considers only primitives based on rectangular ground shapes, still the majority of buildings can be modeled. Thus, one approach to reconstruct buildings when 2D ground plans are available can be sketched as follows: − Try to infer (in 2D) from the ground plan how the building can be subdivided into primitives. − Select each primitive based on additional information, e.g. an aerial image or a DSM. Determine the dimensions by a measurement process. − Assemble all primitives to obtain a single body representing the building. For example, for a simple L-shaped building the first step would select two rectangles to cover the 2D ground plan, the second step would select hipped roofs on each of those rectangles and estimate the eaves and ridge heights and the third step would finally merge both volumetric primitives. This approach has been used by the authors in several projects, combining 2D ground plans with DSMs from laser scanning (Brenner 1999). Its main advantage is that by the use of digitized ground plans, interpreted information is “injected” into the reconstruction process which makes it relatively simple and reliable. On the other hand, since the selection of primitives is guided by analyzing the ground plan, the final reconstruction is strongly coupled to the ground plan shape. For example, roof structures like dormer windows will not appear in the result as long as there is no corresponding hint in the ground plan. This situation can be improved by interactive modeling tools, however. A second disadvantage is, that while most simple buildings lend themselves well to a subdivision into primitives, there are always buildings present which cannot be modeled properly or for which a subdivision into primitives is not very natural and thus modeling becomes involved. In those cases, other approaches which allow to specify the roof topology directly would be more desirable. 2.3 Reconstruction using tree search and DSM segmentation 2.3.1 Ground plans and simple roofs We consider a ground plan given as a closed polygon P consisting of n linear segments pi and vertices v12, v23, ..., vn-1,n, vn,1 (where P does not contain any inner polygon), see Figure 1. Based on P, a simple roof can be constructed which consists of planar roof faces Πi which intersect with the building walls in the eaves Pi and eaves points Vi,i+1. The intersections of the roof faces yields ridges (E12, E24) and grooves. Projected down into the plane the corresponding points and lines (indicated by lowercase letters) are obtained. If the original ground plan is omitted, the remaining points and segments form a planar graph G. It consists of leaves vij, inner nodes vijk and edges eij. If the indices are viewed as sets, then the edge eA connects nodes vB and vC iff A=B∩C. The graph G is connected and without cycles which means it is a tree. Simple roofs such as the one in Figure 1 can be recovered from the corresponding planar Graph G by adding a ground and eaves height. The graph G itself can be obtained from the ground plan P by computing the straight skeleton S(P). For convex ground plans, the straight skeleton S(P) is identical to the medial axis transform M(P). It is different for ground plans containing concave nodes, in which case M(P) contains parabolic segments. The construction of the straight skeleton is a well-known process in geometry but has received attention recently (Aichholzer et al. 1995). Interestingly, no time-optimal algorithm has been found so far (Eppstein & Erickson 1999).