, 9-77 Nov. 1999
ful for examining laser
mple around buildings.
> the surface within the
into the interaction of
Waveform analysis as a
uch as material, rough-
)enefit greatly from this
ilyzing surfaces around
s. This information is
rocessed by NASA Wal-
interesting modification
tting a human measure
ints and surface, we use
tiques. If the laser point
en its backprojected po-
.. We can check by com-
nd the conjugate points
« (1999b)). The match-
ked in this fashion area
er points agree with the
uilding
a set over the building
aspects. First we briefly
aser points with the data
1etry. We then analyze
] present results.
rammetry data sets de-
e no identical points and
yrmed with interpolated
ctive view of the model
Since the building was
breaklines, it is no sur-
Fig. 5(a) looks more re-
ted from the irregularly
)).
uilding model. The left im-
photogrammetrically mea-
N model of the laser points.
rior because twice as many
ng outline (breakline) was
two data sets was per-
cal differences between
nodel to the correspond-
cludes interpolation and
rpolation errors. These
ly near breaklines. Not
dard deviation of +1.08
jer than what one would
nstrates the inability of
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
this popular comparison method to express a meaning-
ful point accuracy, except for smooth surfaces.
The second test with the same data was concerned
with assessing the accuracy of derived features. Sur-
face properties such as breaklines and smooth surface
patches are obtained by segmentation (Csathó et al.
(1999). Segmenting the surface points in the build-
ing area should result in planar surfaces and break-
lines. Our simple segmentation method proceeds in two
steps; first, the points are grouped into potential surface
patches, postulating a surface hypothesis. The second
step is concerned with verifying the hypothesis by fitting
a plane through the points. The deviations of the points
to the plane serve as a validity measure for accepting or
rejecting the plane hypothesis. The validity is not a fixed
threshold value; domain knowledge about expected sur-
face roughness (e.g. man-made objects vs. vegetated
areas) and a priori error estimates of the surface points
influence the acceptance criteria. The results of surface
segmentation, together with other information, are fur-
ther analyzed in an object recognition system.
The final result of the grouping process in the building
area divides the points into potential roof points and
non-roof points. This is achieved by analyzing the el-
evations of points, similarly to histogram thresholding,
except that the spatial distribution of roof point candi-
dates is taken into account. This reduces the chance that
points on trees or other objects of similar height may
accidentally be labeled as roof points. Fig. 6(a) shows
a histogram of the laser point elevations. All the points
clustered within -35 m to -37 m satisfied the spatial ex-
tent criteria and subsequently entered the second phase.
Fig. 6(b) depicts the spatial distribution of the labeled
points; large dots symbolize roof points.
The roof points enter a planar surface adjustment. The
three parameters a, b,c of the equation
Z-a x+b-y+c (2)
are determined in a least-squares adjustment which is
based on the simplified error model that random errors
occur only in z while x and y are considered as con-
stants (Schenk (2000a)). The standard deviation of the
adjustment is a good measure of how well the points lie
on a plane. One out of 94 laser points was identified as a
blunder and removed from the adjustment. An analysis
of the blunder revealed that the point was on a chim-
ney and not on the roof surface. The resulting standard
deviation of the plane adjustment was of = +5.7cm.
Considering the large redundancy, the error measure is
quite reliable. It confirms the high (internal) accuracy of
laser ranging. One may even argue that part of the stan-
dard deviation is due to the non-flatness of the roof.
We have repeated the same experiment with the man-
ually measured DEM. Out of 265 roof points, the ad-
justment procedure eliminated six points as blunders.
These points were measured on small objects on the
roof, such as chimneys and vents. The resulting stan-
dard deviation of = «6.2 cm is nearly identical to the
one obtained from the laser points. However, it is higher
than the expected value of 3cm, obtained with Eq. 1.
150 4
>
e
& 1007
mj
c
o
em 50 pt
0 I 1 1
-40 -30 -20
z-coordinate [m]
(a)
ce s o .
"e KA Dan .
ove * oe C
: $1°°% te, ons.
Se A “ee
ee
Figure 6: The laser point elevations are represented as a
histogram (a). The two peaks are related to roof points and
ground points. An extended histogram thresholding (see text)
leads to the grouping depicted in (b). Solid dots label roof
points.
Again, the larger value may well be caused by the non-
flatness of the roof.
3.4 Segmentation Experiments in Residential Area
The residential area, highlighted in the right part of
Fig. 3, poses new problems for the segmentation. The
procedure described in the previous section must be
modified. Not only are the buildings much smaller but
the roofs have a more complicated structure, consist-
ing of several roof planes with surface normals pointing
in different directions. Objects near buildings, such as
shrubs and trees may have similar heights, challenging
the separation of roof points by histogram thresholding.
To cope with this situation, we have modified the seg-
mentation approach that consists now of the following
four major tasks:
hump detection is a rough analysis of the entire project
area with the purpose of identifying local areas
that contain objects of certain vertical dimension.
We skip the details here and refer the interested
reader to Wang (1999).
grouping generates hypotheses of roof points belong-
ing to one roof plane. Grouping is a local process,
confined to the regions identified by humps de-
tection. Planes are found by a Hough transform
technique.
plane fitting is performed by a robust adjustment, tak-
