ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision", Graz, 2002
vertical (building) surfaces, windows, road surfaces and other
surfaces, while surrounding surfaces are generated to enclose
volume-structured objects, i.e. trees.
In the followings, we discuss each procedure in details. An
experiment and discussion is followed, where a real urban out-
door environment is reconstructed, and the efficiency of the
method is proved.
3. CLASSIFICATION OF RANGE POINTS
As the vehicle moves ahead, LD-As take the cross-sections of
urban objects by scan lines (1200 range points per scan line).
Since LD-A has a circular scanning resolution of 0.25 , spatial
resolution of range points in one scan line (vertical spatial
resolution) depends on the distance from LD-A to target object.
If the target object is 20m (r) far from LD-A, range points are
sampled at a vertical spatial resolution of about 0.087m
(r*tan0.25). On the other hand, when the vehicle moves straight
at a speed of 10-20km/h, scanning planes are almost parallel
with an interval of about 0.1m. Subsequently, horizontal spatial
resolution of range points at the same sequential number is
about 0.1m. However, it alters as the vehicle's moving direction
changes. Classification is conducted by examining the local
connectivity between the range points of the same and
neighbouring scan lines.
3.1 Segmentation of scan lines
Scan lines are first segmented into linear patches, where
successive range points are extracted, which are linearly
distributed with a variance lower than a given threshold. Linear
patches are then compared with the extraction of neighbouring
scan lines. Isolated linear patches are discarded, which cannot
find a linear patch of nearby sequential number and of similar
direction in the extraction of neighbouring scan lines. It means
that the range points have only vertical but no horizontal linear
continuity, so that they are not the measurement of surface
object. Finally, range points in one scan line are divided into
four groups as follows.
1) Range points belonging to vertical linear patches are
the measurement of vertical building surface.
2) Range points belonging to horizontal linear patches
and at ground elevation are the measurement of road
surface. Relative elevation from the origin of LD-A
to the nearest ground surface is almost constant. It
can be measured previously in calibration stage.
3) Range points belonging to other linear patches are
the measurement of other surface.
4) Range points do not belong to any of the above
groups are scatter points.
3.2 Correction of window area
In this research, we assume that the building surface between
windows of different floor is vertical as a whole, is range
sampled enough that can be extracted, and its material is neither
penetrative nor mirror reflective to laser beam. In each scan line,
penetrative measurement on window area has the following
characteristics.
1) They are behind the building surface from the
viewpoints of LD-A.
A - 414
2) Their sequential numbers are between two pieces of
vertical linear patches, which are on a common
vertical line.
Discrimination of window area is conducted using all other
range points besides those belong to vertical building and road
surface. To the range points that satisfying both conditions,
their range values are corrected to fit the common vertical line.
To the false points (no range value) that satisfying the second
condition, their range values are interpolated.
i | | Vertical
| Erroneous line | building surface
| extraction |
| No range data | Tree
: i £n Ground surface
2, ~~ +}
eat 4 4
T a uade t
(a) Linear patches that extracted from a scan line, (b) erroneous
extractions are discarded by comparing the horizontal linear
continuity with the extractions of neighbouring scan lines
suas Windows m Corrected
x^ | window data
T of
| ~ ^
Indoor + Y i 2
objects
(c) Some window glasses are penetrative to laser beam, so that
indoor objects are measured, (d) Range points at window area
are corrected using the data of surrounding vertical building
surface.
Figure 2. Classification and correction of range points
3.3 Trees and others
Range measurement of volume-structured objects, e.g. trees, has
the following characteristics.
1) They are in front of the building surface from the
viewpoint of LD-A.
2) They have higher elevation values than the ground
surface.
3) The cloud of range points implies not a surface but a
volume.
NS Y
Vertical building 3 /
surface on *
d +
7
»
Cloud of
Tree points wa
™N vf 7
^ UU Road Pd
9 i AS
P 4j e o Vertical building
=
surface
Figure 3. Extraction of tree and other volume-structured objects
Volume-structured objects are discriminated as followed. All
scatter points that satisfying the first two conditions are
projected onto a regularly tessellated horizontal plane. For each
grid pixel i, three values are recorded. They are the maximal
elevation value 2 the minimal elevation value _ 2 the