Current issues on 3D city models Marcel Lancelle,∗ Dieter W. Fellner Computer Graphics, University of Technology at Braunschweig 2003/2004 Abstract ible and modular. A Key challenge to enable further improvement of algorithms This research covers issues of automatic and tools is to establish methods to im- generation and visualization of a 3D city port and to fuse data streams adhering to model using existing digital data. Air various GIS standards. It is often the case photos, LIDAR point clouds and soil us- that own formats and conventions exist for age and cadastral maps are used as data each city. It is inefficient and costly to sources for the automated data fusion in develop powerful tools that can only be a sample implementation. Combining ex- used for one city. The process of such isting techniques and new ideas, the de- a standardization is long and expensive. veloped framework enables automatic in- The Open GIS Consortium (OGC)[6] in- formation enrichment and generation of troduces specifications such as the Geog- plausible details. raphy Markup Language (GML) [5] that seem to be more and more accepted. Data conversion, model generation and real time visualization are achieved using a To generate a model usually existing data standard PC for the whole area of 200km2. can be used. An easy access of all the Real world data was used including all the data sets is necessary for the creation practical problems involved. of tools that can generate automatically or semi automatically additional informa- Keywords: 3D city models, GIS, com- tion. This process might be supervised by puter graphics, real time visualization an operator. Many applications using 3D city models do exist. They require very different types of city data. Crucial for the current and later use is storing as much se- 1 Introduction mantic information as possible with each of the data sets. The needed data for each application can be extracted from the Various 3D city models exist with very dif- model and the result might become a part ferent features and intended uses. The au- of the city model itself. Theoretically ev- thors use the term 3D city model to refer ery object (or information) in a city can to a semantic description of objects in a be included in the model but usually only city including 3D information. big static objects like buildings are worth the effort. Building and maintaining such models for professional use requires great expenses GIS (Geographic Information Systems) for data acquisition and operators. To be functionality and visualization may then able to support as many applications as be used to get the desired information. possible, the system has to be very flex- Combining a number of state of the art ∗[email protected] techniques it is even possible to show the Acquisition By storing temporal information of each old GIS Company B once: object, these 4D city models can be used Company A Company C conversion of Ortho- existing data for even more applications. photos DSM, DTM Data bases Manual with city generated model data data Orthophotos Street names, Traffic networks house numbers GIS- DSM, DTM Power networks Data noise immis- Soil usage City objects sion map (buildings, trees, ...) Noise Transportation propagation company simulation Figure 2: A 3D city model of Philadelphia Power supplier Architecture focusing on visualization purpose com- office City object City pared to a photograph (source: [2]). extraction Webserver Editors Visualization tool Participants Data private users Exports Programs Figure 1: Proposed system and an exem- plary data flow. city data in real time in case of a moderate camera speed. Figure 3: Traffic noise map, generated The practical part of the project was a with CadnaA (source: [1]) feasibility study. Only by trying to imple- ment the ideas the most important prob- lems became apparent. With a running application members of the city council can have a much better idea of the possi- bilities and requirements for the long term planned professional change of their sys- 2 Data sources tem. The following list contains some of the Data sources are usually either in vecto- most important applications: rial information or in 2D or 2.5D raster maps. The choice of a common coordi- nate system facilitates the work flow a lot. • information and navigation (tourism, DSMs (digital surface models) are usu- navigation systems) ally generated by airborne photogrammet- • real estate and architecture ric or LIDAR acquisition. Maps (cadas- • user interface for GIS tral, city, soil), photos (aerial photos like orthophotos, terrestrial photos, satellite • city planning photos) and data bases containing any • physical and statistical analysises and location based information are the main simulations (radio wave propagation, data sources. Some data has to be in- flooding, fire, air flow, traffic noise, serted or added manually which usually ...) costs lots of money and the financing is • military and security (training) an important issue. One solution might • multimedia (computer games, enter- be a wiki for some data such as opening tainment, movies, advertisement) hours of shops. The terrain surface can be created by ap- plying morphologic filters to the LIDAR scans. Landmarks are usually manually modeled – as a matter of fact, it turns out that it is crucial that the import of manually added and augmented objects is provided by such a system – because their special properties are most important for a recognition but also most difficult to cap- ture with an automatic process. For visualization purposes moving and an- imated objects enhance the realism a lot. For real time rendering a suitable LoD (level of detail) technique for each type of Figure 4: Orthophoto, LIDAR scan, objects is required and an automatic gen- cadastral map, soil usage. eration of random plausible detail is very helpful. 3.1 Buildings Buildings are the most important part of a 3D city model for many applications. Manual modeling can lead to very good re- sults but is only feasible for very small ar- eas. Using building contours from cadas- tral maps, LIDAR data and orthophotos, polygonal models of the outside of build- ings can be generated automatically. The recognition of the correct roof shape (see Figure 6) requires complex algorithms. Facades can be textured in a low reso- Figure 5: Result of automatic classifica- lution using non-orthographic aerial pho- tion of soil usage type using LIDAR and tos. More accurate acquisition can be NIR data (source: [14]) done using terrestrial scans. The results are promising (see Figure 7) but could not be applied to a whole city yet. 3 City objects Important city objects are: 4 Visualization • terrain surface and sky box • buildings Drawing every data available is simple but orders of magnitudes too slow. Numer- • landmarks ous optimizations are necessary to render • vegetation only the most important visible parts of • street furniture the model. Besides culling of invisible or • moving objects barely visible geometry (see figure 8) ap- • traffic and transportation networks propriate LoD techniques for each object City model Contribution culling Occlusion culling Backface culling Frustum culling LoD choice Clipping Rasterization OpenGL Figure 8: Data reduction for rendering. 5 Results Many of the ideas were implemented and Figure 6: Result of automatic roof recon- tested with data provided by the city of struction (source: [16]). Braunschweig. The area of available in- formation covers about 200km2. Height data, orthophotos, soil usage and cadas- tral maps are the most important exist- ing digital informations that were used to generate the model. Both conversion and visualization are done on a standard PC (1 GHz processor). Automatic conversion and fusion of about 13 GB input data and generation of buildings took about four days on two of these PCs. OpenGL is used for rendering. One of the hardest parts was to obtain the Figure 7: Automatically combined air- data sets. It took five month to get the borne and terrestrial scans for accurate fa- data for one city. For a second city only cades (source: [11]). parts of the objects were available but it is still enough to prove that the software can be used for several cities. One problem or clusters of objects are essential. Sim- was exporting the information from the plification and reduction techniques, bill- database to tiles. It was difficult to find a boards, impostors and point based render- format were most of the information was ing also belong to this category of opti- maintained and updates of the GIS soft- mizations. ware during development caused changing of the file format of the exported cadas- Instead of rendering the city model on tral maps. Much of the existing informa- a globe it is usually sufficient and much tion in these maps was stored for the pur- faster to approximate the earth locally by pose of (2D) visualization without seman- a plane without the necessity of expensive tic information. For example, more or less transformations. unconnected lines representing street and walkway boundaries can be recognized by Depending on the application, different a human viewer but are almost useless for styles such as NPR (non photo realistic) automatic methods. For building contours rendering may be used. Overlay of GIS search algorithms had to be implemented data, current position or a coarse map of that are now able to extract the shapes of the city are possible. almost every building correctly. The or- thophotos available are not true orthopho- More features were implemented such as tos (see Figure 4) and thus not suited for support for stereo projection and a CAVE texturing of the roofs. (see Figure 11), offline rendering of videos and high resolution screenshots. The project duration of nine month didn’t City model Data sources Conversion Read Visualization allow by far to build a sophisticated set of data Orthophotos conversion programs, tools and databases Orthophotos (image tiles) as proposed and necessary for a profes- DSM, DTM DSM, DTM (raw data) sional use.