Automatic Generation of 3D Building Models from Point Clouds - Geoinformatics for Intelligent Transportation - page 110

In one approach, which is explained in detail [ 9 ], a robust algorithm for the
autonomous reconstruction of buildings from sparse LiDAR data was used. No prior
knowledge and supplementary sources were needed in this paper. The classi

cation

of the point cloud into terrain and off-terrain points was done by a filtering process

that uses global functions in the form of orthogonal polynomials. The iteration
process was based on a fitted function that passes between laser points. In the sub-

sequent iterations, the degree of the polynomial was decreased. The reduction of the
polygon degree was done in accordance with an evaluation of the residuals. The
assumption was that terrain points have negative residuals and off-terrain points have
positive residuals. The

final result was reached when the iteration did not change the
shape of the terrain, so only the terrain points in fl uenced the polynomial [ 10 ].

Separated areas of off-terrain points were then filtered by size and height above the

ground. Filtered points that created planes as roof parts were further grouped into
segments and classi

ed in unique roof faces. Between roof faces, topological rela-

tions were identi ed and from this information, an adjacency graph was created

which was very useful for the determination of roof types. The crease edges between
the roof faces were computed by the plane intersection. Boundaries of buildings were
derived from their roof edges that were detected using the Hough Transformation. It
was an approximate solution with lines that were geometrically incorrect, they were
not parallel and rectangular. Extracted lines were also fixed using an adjustment with

specific weights for roof edges according to their classification into three classes:
horizontal crease edges, non-horizontal crease edges and border lines. The adjust-
ment was not a destructive operation so it had no in uence on topology. The gen-

eration of buildings was completed after the adjustment of their bounding lines.
This solution [ 9 ] is much more sophisticated than the one that is described in
Haala and Brenner [ 1 ], where the shapes of buildings were determined by a seg-
mentation of DSM in the raster form and a subsequent extraction of planar regions.
The planar range image segmentation algorithm [ 11 ] was used because it was fast,
essentially simple and scored very well if compared to other algorithms [ 12 ]. The
algorithm is based on the segmentation of a regular DSM grid into straight 3D line
segments which are used as a starting position for the region growing process. The
main advantage of this algorithm is that it requires no a priori knowledge. How-
ever, the algorithm has a problem with the precise extraction of region boundaries.
The author solved this problem by using ground plans. The problem was bypassed
instead of being solved.
The following part of the text is dedicated to an analysis of commercial software,
which is designed for the automatic generation of 3D building models.

fl

3 Methodology and Data

The automatic generation of 3D building models will be tested in ENVI LiDAR
(more detailed in [ 13 ]). Other commercial products are INPHO Building Generator,
tridicon CityModeller and tridicon BuildingFinder. INPHO Building Generator and