ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
will be places in the point-cloud where an operator will not be 
sure of the classification of points or even worse will not be able 
to make a classification. Therefore, three classifications are 
possible, "Object" and "Unclassified". The 
"Unclassified" points are those that an operator may not be able 
to classify with full certainty. This is in contrast with the current 
practice of filtering DEMs from laser scanner point-clouds, 
where only two classes are allowed for, "Terrain" or "Object". 
"Terrain", 
The difficulty in manually classifying a point-cloud lies partly in 
the inherent characteristics of a point-cloud. If we can identify 
those characteristics of a point-cloud that lead to a difficulty to 
classify then we would be able to devise an alternative means of 
checking if classification is possible. This paper is a look at 
some issues related to this problem with a view to developing 
some framework for classification and a means of identifying 
regions in a point-cloud that maybe difficult to classify. 
In the first section of the paper, several spatial characteristics 
that contribute to a difficulty in classifying a point-cloud are 
outlined. In the second section, the implication of the outline for 
future filter designs and strategies is discussed. In the final 
section, a trial procedure for detecting "unclassifiable" regions is 
described. 
Test point, p Neighbourhood, N 
  
à =f(p.N) | 
Type ll?  Typel? 
« bie +» 
  
  
Terrain 
4 Threshold $ 
Figure 1 Current approach to filtering 
2 DIFFICULTY TO CLASSIFY 
Considering the characteristics of a point-cloud, the difficulty in 
classifying originates from the characteristics of the landscape 
and the characteristics of the laser scanner data. 
2.1 Characteristics of Landscape 
“Unclassifiable” regions due to landscape characteristics occur 
because of the nature and arrangement of objects and the terrain 
- e.g., terrain, buildings, vegetation, etc. The difficulty to 
classify due to landscape characteristics will be a problem in 
every laser scanner point-cloud without exception. 
Complexity of the terrain - Here complexity is in reference to 
the form of the terrain (not roughness), 
discontinuities (e.g., embankments, raised platforms, terraces, 
etc.,). This characteristic especially comes into play if an 
operator is only able to see a portion of the point-cloud at a time, 
in particular 
or if the terrain in a small neighborhood changes drastically in 
relation to the terrain around it (discontinuity, platforms, etc.,) 
such that it no longer fits into the general form of the terrain. 
Usually an operator is able to see quite a large portion of the 
terrain so he/she is able to appreciate the form of the terrain. 
This problem is demonstrated in Figure 2 (a) and 2(b). 
Complexity of objects - While most buildings in urban areas 
are regular in shape (blocks, prisms, etc.,) there are some that 
are more complex (layered roofs, platforms, etc.,) and hence 
more difficult to classify. Another way in which objects are 
made complex is by the nature of objects and terrain around 
them. For example, a depression (e.g., a pool, excavation, etc.,) 
near a building makes any land between the depression and the 
building uncertain (Figure 2d). 
Proximity of Object and terrain - The closer an object is to the 
terrain the more difficult it becomes to distinguish it from the 
terrain. This separation can be lateral (in the case of sloped 
terrain), vertical or both. Expressed differently the separation of 
the surface of an object and the surface of the terrain becomes 
more difficult as the surfaces start to approach each other in 
whole or part, Figure 2(c). The ability to separate the terrain and 
objects is further complicated by the size of objects. The closer 
and larger an object is in relation to the terrain, the more 
difficult it is to separate it from the terrain. Therefore, the larger 
the area of the object surface, in relation to the area of the terrain 
surface, the greater the difficulty of separation. 
Zero ground returns (lack of terrain information) - In built- 
up and densely vegetated areas, the number of returns from the 
terrain surface can approach nil. All filtering involves the 
comparison of a point with its neighborhood. For filtering, this 
comparison is meaningful only if some points in the 
neighborhood belong to the terrain. Zero ground returns for a 
neighborhood may make the result from a filter meaningless. 
The ability to detect this problem depends on the extent of the 
area for which there is no ground return. The larger the area the 
greater the difficulty of separation. 
2.2 Characteristics of data 
“Unclassifiable” regions due to data characteristics occur 
because of filtering objectives and filtering assumptions. The 
difficulty to classify due to data characteristics will be a problem 
most of the time. However, unlike those due to landscape 
characteristics, the difficulty of classification due to data 
characteristics can to a certain extent be controlled. 
Resolution of the point cloud - The resolution of a point cloud 
has a direct influence on the spatial definition of objects. The 
less defined an object the more it starts to blend into the general 
character of the terrain. In Figure 2e is shown buildings on a 
hillside. Because of the low resolution, buildings (about 15m in 
length) to the bottom and right have started to loose definition 
and blend into the terrain. Here the proximity of object points 
also has an influence on the ability to discriminate between the 
object and the terrain. The higher the object, the easier it is to 
discriminate it from the terrain. As the resolution of laser 
A - 331