Point cloud cleverly modelled using 3D skeleton
Geographical point clouds are a useful tool for modelling our built and natural environment. In practice, however, this demands a lot of manual work. Ravi Peters completed his PhD project on the use of a 3D skeleton, enabling to divide point clouds into objects that then became easier to process automatically.
Point clouds are three dimensional, but in practice, 2.5-dimensional methods are often used to process them. These are based on what is known as ‘boundary representation’, whereby only the contours of the objects are modelled. As a result, information is lost and raw point clouds become difficult to structure. This makes automatic object recognition impossible. “The method I am introducing shows the skeleton of an object,” explains Peters. “You see the pivotal points, which makes it easier to understand the various components that make up an object.” This is useful not only on the scale of buildings, but also on the scale of entire regions.
Peters is not changing the method of gathering points; this is still done using a 3D laser scanner. But he now structures the data collected using the 3D Medial Axis Transform (MAT). This method is already in use for computer graphics, but hasn’t previously been applied to geographical point clouds. The MAT uses an algorithm that calculates the skeleton of a shape by projecting three-dimensional spheres into it. In addition, the method has features that allow the geometric properties of a shape to be defined in a compact way.
The advantage of the method Peters has developed is that the MAT removes a lot of 'interference' from the point cloud. Much of the data for geographical point clouds is gathered from aeroplanes. Mistakes caused by passing clouds or small jumps in the GPS position are unavoidable. When the mistakes are removed, a lot of the data is lost. The MAT is able to thin out the number of points without losing the details. This makes it much easier to work with enormous sets of data. To illustrate: a geographical point cloud covering an area the size of the Rijnmond region comprises around 12 billion points. “Thanks to the simplification made possible by the new method, this number is reduced by about 90 percent without affecting a visualisation of the point cloud,” says Peters.
Another important advantage is the fact that the point cloud can be divided into objects. The way things are currently done makes this impossible and that is particularly problematic when using point clouds for Geographical Information Systems (GIS). At present, identifying a single object in a sea of data is a task that must be done manually. Peters' thesis 'Geographical point cloud with the 3D medial axis transform' demonstrates that the MAT makes it possible to pinpoint houses, offices, dykes, and waterways automatically.
The MAT also simplifies the use of data from point clouds for applications such as 3D printers or VR goggles. Thanks to the new method, the separation between the object and the air can be more clearly distinguished. It prevents a 3D printer from printing in a vacuum.
The Canadian software company Safe Software recently included the (open source) method in its FME package.