Automatic Generation of 3D Building Models from Point Clouds - Geoinformatics for Intelligent Transportation - page 112

4.1 Input Data Quality

Software testing was performed on three datasets with different densities of point
clouds. The first and second dataset covered the identical area (part of the city

Rokytnice nad Jizerou) and so their outputs can be compared to each other. Sample
dataset attached to the software was a very dense point cloud and thus represents an
interesting source suitable for precise modelling.

4.2 Classification of Point Cloud

Figure 3 (top) shows point clouds in the colour hypsometry in relation to the height
of the terrain. The colour range is not created from absolute height values, but from
relative heights of points above the ground. This type of visualization eliminates the
height disparity of the terrain, which gives a better idea about objects on the ground
(buildings and vegetation). Classi cation results are in Fig. 3 (bottom), where each

class is displayed in a different colour.
Input datasets were classi

ed into four classes (terrain, buildings, other and
unprocessed). Figure 3 shows that the terrain was identi ed correctly in both

datasets, however, it does not apply to the buildings. In the first dataset, almost no

buildings were found. This is due to the too sparse point cloud from this dataset.
Points representing most of the buildings were classi

ed incorrectly into the class

other. The result of the classi cation of buildings in the second dataset is signifi- -

cantly better. All buildings have been identi ed. The class other contains only those

points that represent vegetation (trees and shrubs) and small objects with low height
(outhouses, greenhouses, fences, cars etc.).
The result of the classi

first dataset is
not satisfying. Changes of the recommended setting did not lead to improved
results. For this reason, the first dataset is not suitable for building modelling. The
next part of the analysis considers only the remaining two datasets in Table 1 (#2
and #3).
The following Fig. 4 shows only a part of the test area. It is approximately the
central part of the second dataset. The classi

cation of the sparse point cloud from the

cation result can be evaluated easily
by a visual comparison of the orthoimage [Fig. 4 (left)] and the classi ed point

cloud [Fig. 4 (right)]. A good classi cation is a prerequisite for the subsequent

automatic generation of 3D building models.

4.3 Creation of 3D Building Models

The automatic generation of 3D building models is based on the RANSAC
(RANdom SAmple Consensus) method [ 14 ]. The principle of this method is to nd