A 952Kb PDF of this article as it appeared in the magazine complete with images is available by clicking HERE
As I have mentioned in a number of past columns, we are addressing the small UAS mapping industry via our subsidiary company, AirGon LLC. While our primary business is developing and/or integrating technology to provide turnkey mapping systems, we also provide a small amount of mapping services. I have two goals for this services business; help customers transition from a services model to an owner-operator model and inform the development of our point cloud analytic software, TopolystTM.
When we first engage with a customer, we are usually put to a test. This test is "how do your results compare to those I have been obtaining through my existing data acquisition methods?" This is where a confusion in terminology exists between what is being collected via conventional photogrammetric stereo and new techniques. The use of inconsistent terminology in the press and various white papers only adds to the confusion.
Many of our mining customers (mainly aggregate quarry sites and associated stockpile yards) are using conventional manned aerial photogrammetry as their current method of mapping (volumetrics and topographic contours). In fact, some of these data are still being acquired with film-based aerial cameras with data collected on analytic stereo plotters!
Conventional photogrammetry uses two images that overlap the same area of the ground to form a stereo model. I offer an illustration from an 1893 US patent to validate the use of the term "conventional!" The concepts presented in this patent are fundamentally the same technique used in modern stereo pair photogrammetry.
For small sites such as mines, 3D models are constructed by extracting points and lines, typically in an interactive editing session on a stereo photogrammetric workstation. In Figure 2 is illustrated an example of a model constructed by stereo profiling. In this technique, a pattern of posts (usually in matrix layout) is driven through the reference image (say the `left’ image). An automatic correlation algorithm called "Dot on Ground" (DOG) is used to find the match point in the right image (the conjugate point). The 3D location is then derived via photogrammetric triangulation. While this is a time tested method of collecting three dimensional information from 2 dimensional images, it is not well suited to full 3D model reconstruction. In general, the more the earth undulates, the less accurate this modeling technique becomes. This is particularly problematic in very high spatial frequency areas such as mine site high walls.
Constructing 3D models by generating point clouds from multiple overlapping images is an entirely different process than conventional stereo photogrammetry. An illustration of a three image extraction is depicted in Figure 3 (note that it is typical to have five or more images covering each point in object space). I am not sure what the correct generic terminology is for this general class of algorithm but it seems the popular general term introduced in the early 1990’s is "multi-view stereo photogrammetry." From the generic approach of using more than two views of the same object point came specific algorithms, each with its own set of terminology. The most common general class of algorithm in use today is called "Structure from Motion" (SfM). The motion part can be a bit confusing but it basically means that the camera has moved from location to location (called "pose" in the computer vision world) while imaging the same area of the ground (or whatever object is being modeled). Popular software applications that implement SfM are Pix4D Mapper and PhotoScan.
This dense image matching can produce life-like, very detailed point cloud models of the area being imaged. An example of a stockpile at a sand mine is shown in Figure 4. These models have varying degrees of conformance (how well the point cloud actually matches the real world), depending on a number of factors such as camera quality, lighting, surface texture, curvature of the object and so forth.
So finally, the crux of this article. We, as an industry, need to distinguish conventional stereo photogrammetry from multi-view stereo. I am seeing articles, particularly related to the drone mapping industry, compare LIDAR to Photogrammetry, weighing the pros and cons of each technique. The authors of these articles are always referring to SfM when they discuss "photogrammetry." Unfortunately, a mine site operator reading the same article will interpret the term "photogrammetry" to mean the conventional two image stereo pair collection. The output products are, of course, entirely different.
I have been advocating that we refer to conventional stereo pair photogrammetry as "photogrammetry" and SfM (in its multitude of incarnations) as Dense Image Matching (DIM). So far, I am not gaining a lot of traction! Am I being pedantic or does this really matter? It actually is hugely important because the achievable accuracies as well as products are considerably different between the two modeling approaches. In a future article, I will present these three common modeling techniques (the third being LIDAR, of course) and suggest where each excels. In the meantime, start saying "Dense Image Matching" ten times per day!
Lewis Graham is the President and CTO of GeoCue Corporation. GeoCue is North America’s largest supplier of LIDAR production and workflow tools and consulting services for airborne and mobile laser scanning.
A 952Kb PDF of this article as it appeared in the magazine complete with images is available by clicking HERE