Minggu, 10 Agustus 2008

DigitalGlobe on NBC: A Closer Look at 3D Olympics

I've seen a couple of posts regarding DigitalGlobe's Bejing Oympic Games coverage. In my opinion this is a great consumer-grade marriage of geospatial technology and mainstream media. On the DG homepage "DigitalGlobe on NBC" is prominently displayed, which takes you to http://www.digitalglobe-aegistg.com/, their site with EAgis Technologies (a joint effort between DG, EAgis, and NBC).

The posts above comment on the technological underpinnings on the 3D scenes offered on the new site. One common theme is how a combination of different technologies (photogrammetry, 3D modeling, satellite imagery, 3D visualization) can be used together to provide a powerful and immersive viewing experience.

The http://www.digitalglobe-aegistg.com/ site allows visitors to download 3D PDFs, a KMZ and perspective views of various Olympic sites. 3D PDFs have been around for awhile, but this is the first time I've had a chance to examine one: very cool, although not the same experience as a KML file in a virtual world. Compare for yourself below:

KML in Google Earth:

Aside from the availability of 3D example data, the site also provides some insight into the creation of the dataset. The "How Can This Be Possible?" heading expands to provide a high-level introduction to the technology and workflow. The workflow is divided up into four parts with a graphic associated with each one: the "3D wireframe" generation, imagery capture, feature extraction and extrusion, and fully textured 3D model generation.

"3D Wireframe" Generation

The site mentions that the wireframe represents the earth's terrain, and that it was derived from two DG satellite images. Sounds like classical photogrammetry! Two images associated with sensor model can be viewed in stereo to extract (measure) 3D positions (e.g. points with an accurate XYZ location). For high-accuracy applications relying on the satellite sensor model may not be enough, and there would be a need to collect and measure ground control points and then run through the triangulation process. However, once that is done there are numerous applications that may be used for terrain extraction.

The screen capture of the wireframe is actually a Triangulated Irregular Network (TIN), which when compared to raster DEMs is a more efficient means of modeling terrain. These can be automatically correlated using point matching algorithms or manually compiled by hand - which can be a very time-consuming process.

Why is the TIN important? The terrain represents a fundamental part of an immersive 3D scene. If it isn't accurate then the scene will not look realistic... In inaccurate terrain model could also cause problems in the image processing (orthorectification) part of the workflow.

It is also important to note that terrain can come from a number of sources: manual compilation, automatic correlation, LIDAR, IFSAR, and other sources.

Imagery Capture

This screen capture shows imagery draped over the terrain. The imagery would have come from QuickBird or WorldView-1 satellites. For a good-looking scene high-resolution satellite imagery or aerial photography is important. Sometimes satellite imagery is useful, but often aerial photography is the best solution. Why? If imagery is captured from a sensor mounted on a plane, the data acquisition organization has full control over the scale/resolution of the photography. Flying low equals higher resolution...

Another important note is that the terrain model discussed in the previous step would likely be used to orthorectify the image. This will result in a geometrically accurate orthophoto with real-world coordinates. The accuracy of the terrain is important: if there are large errors the 3D features discussed in the next step may not appear in the correct position if they were extracted in a stereo feature extraction system (building footprints digitized off the orthos and then extruded would be ok though).

Feature Extraction and Extrusion

The text for this segment talks about "special tools" being used to determine "footprints" of buildings and then extruding them. This might work for some rectangular buildings with flat roofs, but it is clear that all the download-able content on the site was not derived from automatic extrusion. There are a couple of ways to generate 3D buildings. The quick an dirty way (extrusion) involves digitizing the building footprints in 2D from a digital orthophoto. Then you need to tag the building polygon with an attribute to represent height. This is a fairly straightforward procedure if a digital surface model (DSM) of the area is available. The drawback of extrusion is that, although quick, it may not be accurate. Extrusion assigns one elevation value for the entire building (roof) area, so buildings with pitched or complex roof structures will not be modeled accurately.

Photogrammetric feature extraction can model buildings with greater detail, since specific building detail can be modeled in stereo by viewing and measuring buildings in 3D. However, photogrammetric feature extraction is performed from a "top-down" perspective, so features like balconies may be difficult to model. This is where CAD or CAD-like 3D modeling packages and ground-based photography can help. One workflow for 3D city construction is to photogrammetrically extract the buildings and then import them into a CAD package to add more detail to the models. Ground photos can also provide photo-realistic image texture, as can aerial photography, but capturing all four sides of a building can be difficult without planning the acquisition flight with a very high degree of overlap - which can add to the project cost (more fuel, more data to process). In addition, aerial photography may not be able to capture street-level image texture or areas with dense skyscrapers.

At any rate, there are many ways to go about generating the 3D buildings - it all depends on the level of detail required and the project budget...

Textured 3D Model Generation

As I mentioned above, texture can be applied to buildings from both ground and aerial photography. There's a number of tools that can be used to texture the buildings, here is an example video of how this can be done in SketchUp. There's a number of 3D modeling applications out there to do this sort of work. Again, production costs rise an accordance with the level of detail applied to a building. A "perfect" building cannot be easily automated and can be laborious to produce in sophisticated packages such as Autodesk's 3ds Max.

Looking at the Beijing Institute of Technology model it is clear that a lot of effort went into building it. Not only does the texture look great, but there is a lot of 3D modeling that has been done in a professional 3D modeling system. The rounded rooftop would be very difficult to model in a photogrammetric feature extraction system, and the model contains detail of the roof overhangs - which would likely come from the use of ground photography.

At any rate it is nice to see this technology getting some mainstream media coverage. Photogrammetry and 3D mapping have been around for a long time, but the mass-market popularity of visualization packages such as Google Earth is exposing this technology to a much broader audience.


0 komentar: