Aerial Perspective: Photogrammetry Versus Lidar
Professional Surveyor Magazine - August 2009
Clearing the Air
Multi-cone frame cameras extend the range of photogrammetric applications; through digital frame cameras, photogrammetry has been reinvented and will continue to evolve. Meanwhile, we have not even scratched the surface of what digital photogrammetry can do. Future opportunities in enhanced processing software will extend digital-camera sensor capabilities to more and more applications. There is huge potential in what can be called "software leveraged hardware."
One visible and important application is DSM production from digital imagery. All major software vendors are implementing algorithms to automatically extract surface models from digital imagery. Vexcel Imaging, for example, has developed fully automated ortho-rectification software and automated 3D model generation algorithms for Microsoft's Bing Maps (formerly Virtual Earth). These software developments will process thousands of UltraCam images for generating 3D city models for Bing Maps.
Multi-ray photogrammetry has created a significant change in photogrammetry with the advent of the digital camera and a fully digital work flow. This allowed for significantly increased forward overlap of images as well as the ability to collect more images virtually and without increasing acquisition costs, because only hard disk storage and computation time is required to store and process the additional imagery, and both are low-cost. It's a significant improvement compared to film-based cameras where each image affects direct costs such as film, development, and scanning.
Multi-ray photogrammetry is not exactly a new technology, rather a specific flight pattern with a very high forward overlap (80 percent, even 90 percent) and an increased sidelap (up to 60 percent). The result is considerable redundancy, critical for robust automated matching. One pixel on the ground is visible in up to 15 images. Such a dataset allows automated "dense matching" to extract surface models from the imagery.
Once the DSM has been processed from the imagery, a filtering and classification process - comparable to the filtering and classification of elevation models acquired by lidar scanning - can be applied to achieve the terrain model (DTM).
Figure 3 shows a greyscaled relief of the DSM (left) and the DTM (right). The DSM has been processed using UltraMap software and a set of UltraCam images. The DTM has then been processed out of the DSM using a "Winston-Salem" algorithm developed by Microsoft.
High-resolution DSM and DTM production is no longer a domain of lidar scanning and can be accomplished also through photogrammetry and digital frame cameras. In most applications, multi-ray photogrammetry achieves a significantly higher point density with superior collection efficiency. The achievable height accuracy is better than the GSD; thus a 10cm imagery leads to <10cm height accuracy of the DSM. The quality of the automated dense matching process depends on the camera and the structure of the terrain. Geometric stability and radiometric dynamic of the camera directly affect the matching results.
The work Vexcel does for Bing Maps shows that roughly 50 percent of all pixels in the image match automatically. In "boring" terrain, this ratio drops below 50 percent, and in well-structured terrain such as cities, this ratio increases drastically. Due to the substantial amount of pixels collected by a digital camera (a 10cm GSD image has 10x10 = 100 pixels per square meter), an average dense matching ratio of 50 percent results in a DSM with a point density of 50 points per square meter. This is significantly higher than the point density of a typical lidar project.
In Figure 4 the left image shows a part of the city of Gleisdorf near Graz, captured at a ground sample distance of 24 centimeters. The city has been surveyed with an 80 percent forward overlap and a 60 percent sidelap. This allowed automatic processing software to acquire a surface model from the imagery with an average point density of more than 8 points per square meter. The DSM is shown on the right.
Another survey was performed with 10cm imagery (Figure 5). The raw imagery was then processed and a very dense DSM automatically created employing the multi-ray photogrammetry approach and through dense matching of the imagery. The average point density of the DSM is higher than 50 points per square meter and the height accuracy is better than 10cm. The DSM was then used to process a true-ortho image from the raw imagery.
In Figure 6 the image on the left shows a grey shaded relief of the dense DSM that was processed by automatically matching the digital imagery from the UltraCamL flight. The grey values represent height information. The right image shows the true-ortho image that was then processed by overlaying the raw imagery with the dense DSM.
Noticeable is the extremely sharp edge representation in the DSM due to the high point density achieved through the automated matching of the high-resolution imagery. The result is that the true-ortho imagery shows almost no artifacts in spite of the 10cm high-resolution GSD.
The table below compares a lidar project with an image project. The goal of the scenario project is the creation of a DSM with a density of 8 points per square meter. This translates into a 25cm GSD image collection project for the cameras. A generic high-performance lidar sensor system has been chosen for the comparison with the UltraCamXp and UltraCamLp.
The comparison shows that with a high-performance lidar sensor system, an effective strip width of 322 meter can be achieved, compared to 1,731 meters when using the UltraCamXp and 1,170 meter through the UltraCamLp. Setting the UltraCamXp strip width to 100 percent, the collection efficiency of the UltraCamLp is 68 percent of that and the lidar sensor system is 17 percent.
The multi-ray photogrammetry approach extends the capability of photogrammetry for applications currently being served through lidar scanning. In combination with multi cone frame-based digital cameras, it allows an automated point-cloud and DSM generation from digital images through robust dense matching. Due to the superior collection capacity of digital frame cameras, the frame-image based DSM generation leads to higher point densities at lower collection costs when compared to lidar scanning. A DTM can be processed by applying filtering and classification comparable to the filtering of lidar data for the DSM.
A future improvement of the camera-based DTM production will be achieved by combining traditional filtering and classification with image-based classification. Image-based classification allows automatic extraction of more information about objects. This information can be used to improve the filtering and the classification of the DTM data.
An additional and significant benefit is that users will be able to process DSM and DTM from digital frame images without an additional sensor or an additional workflow. They'll be able to do so with pre-collected digital frame images and the standard photogrammetric workflow already in place.
Alexander Wiechert is general manager at Vexcel Imaging GmbH, a Microsoft company in Austria. He holds degrees in Aerospace and Aeronautics and in Business Administration.
Dr. Michael Gruber is chief scientist at Vexcel Imaging GmbH.
» Back to our August 2009 Issue