Filtering and Denoising

Some of the returns in a delivered point cloud measure no surface at all, and the returns that do measure a surface do not lie exactly on it. You can see both defects in the airborne section of figure 6.4. Inside one window of that section, 67 of 9 513 returns lie in empty space, with at most three neighbours within a metre against a median of 70 for the rest. The flat roof in the same section scatters about its own plane fit with a standard deviation of 2.7 cm. Now, those are two different defects, and one repair will not serve for both. We delete the returns that measure nothing and move the ones that measure a surface badly. We take these operations in that order in this chapter.

How much of either should you expect on your own data? Neither quantity is a constant of the instrument. What share of the returns measures nothing depends on the scene materials, on moving objects, on range and incidence angle, and on the rule used to label a return an outlier. This is why glass, polished stone, vegetation and passers-by all raise it. That thickening is set by the scanner, the target and the acquisition geometry together. So I read both off the station's own diagnostics, or measure them on a control patch of flat cooperative material, rather than take a figure from a specification sheet.

You arrive here with the georeferenced and downsampled point cloud of chapter 5 and the spatial indexes of chapter 4, against which every filter below issues one neighbour query per point. You still need a statement about which of your returns measured a surface. We add that statement in this chapter through the two operations in definition 6.1. Figure 6.4 carries an instance of each. Whatever you leave in the file propagates into the normal estimation of chapter 7, on which the bilateral filter of this chapter itself relies. It also propagates into the registration of chapter 8 and the surface reconstruction of chapter 15. All of these degrade visibly on unfiltered data.

Three questions run through the chapter. How do you tell a return that measured nothing from one that measured a surface badly? How hard may you clean before the cleaning becomes the larger error? And how do you show that the pass you ran was the right one, on a survey where no clean reference exists to compare it against? Let's start with the defects themselves. The filter that suits your point cloud is fixed by the one that dominates it.