I just had a riveting conversation about point clouds with Elaine Ball, the newest member of the TopoDOT team--and we actually do find these discussions "riveting".  What are point clouds really? Basically just millions of points each represented by three numbers. Look at a them as a list of numbers and they make absolutely no sense. Plot them and they represent the 3D world. Millions of points representing trees, signs, roads, houses, whatever was scanned or photographed  whirl around your computer in a perspective view that seems to invite you in for a walk. But to really take that walk and extract value from the data, you have to understand what the point cloud is--and isn't.

I've always found it instructive to view traditional survey coordinates as point clouds also. Why not? They are also essentially just three spatial coordinates--just far fewer of them. But each of these traditional coordinates hold far more information than any individual coordinates of a huge point cloud. Why? Because traditional survey points are "intelligently" sampled in the field. A surveyor walked to the point, aimed at or touched it, recorded it, categorized it and gave it a name. That's a tremendous amount of information associated with each traditional point. Contrast these surveyed points with those of a point cloud data. Each of these points has no intelligence associated with it. In fact they are simply sampled in a pattern across the scene. The vast amount of information contained in the point cloud is owed primarily to the overall density of points. Individually they hold little information, but collectively they can describe a scene in great detail.  While this may seem obvious, failure to understand this concept led many early process development efforts down unproductive and unprofitable paths.

The traditional land surveyor literally touches the real world point in some way to sample it. The millions of coordinates comprising a point cloud form a 3D image representing the scene. But these are just samples of the 3D world. So while the field surveyor selects the precise location of a coordinate to describe a feature location, a point cloud processing technician must either select a point close to the feature or "infer" the feature location from surrounding points.

Point selection 

Selecting the closest point to a feature is the method employed by many software solutions. This technique relies on point clouds being sufficiently dense such that an individual point can be identified and selected within the required tolerance of the extracted feature location. Such high density point clouds are generally only possible in contained areas such as inside a room, facility or similar.

Move outdoors and the effectiveness of this approach breaks down quickly. Traditional topography survey applications require long distance measurements at relatively shallow oblique angles to horizontal surfaces such as roads, ground, and distant buildings. The result is spacing between points typically far exceeding the required tolerance of feature location. Simply stated, if point spacing on a feature such as a curb exceeds for example 1 inch (25.4mm), how can any selected point yield a guaranteed spatial accuracy less than 1 inch? It can not.

Denser point spacing would be the immediate answer, but the price paid will include dramatically more field time and larger file sizes. These inefficiencies quickly result in acquisition times often exceeding those of traditional survey, especially for tripod based acquisition. The laser scanning solution can quickly become more costly than traditional survey--not a great feature of an expensive technology.

Point Inference

Since the beginning of TopoDOT development, we recognized that feature extraction could be much more productive by exploiting surrounding points to infer the location of a feature and extract a virtual point. Thus TopoDOT tools are primarily designed using digital signal, statistical and image processing techniques adapted to point cloud data. The ultimate consequence of this approach is that entire topography and/or building 3D models can be accurately extracted without ever clicking on a single point in the point cloud! Moreover, the inferred location of every feature is dependent on the location of many surrounding points. Thus employing appropriate extraction techniques greatly diminishes the effect of uncertainties due to statistically random noise in the data.

For a simple example, see the curb line extraction below. Here the point density spacing is roughly 2 inches (50mm). It would be practically impossible to select any one point to accurately represent the flow line, top of curb or back of curb much less all three features. But select enough data and look at it as a profile, an accurate selection of all three features is possible in the same plane--minus the random noise. This is but a simple example as much more sophisticated signal processing techniques are built into TopoDOT's ever increasing number of highly automated tools.

So much of TopoDOT's extraction productivity is owed to a fundamental understanding of point cloud data. The inference of a feature location based on surrounding points allows very accurate extraction to be accomplished with much lower requirements on point cloud density. Lower point density requirements results in faster field acquisition times and lower amounts of required data. As field acquisition time is expensive, TopoDOT actually saves field time resulting in more cost effective applications of laser scanning and other point cloud producing technologies.