imultaneous
n is made in
eodesy and
using the
ninimum of
resampling
> textures, it
ars or trees.
are Gimp 2.
riate model
Sketchup 8.
1.
model first
| have to be
S:
el of detail
ification of
1 (modelled
prrection.
uracy of the
centimetres
er than the
yoint clouds
nal accuracy
as based on
ls pragmatic
f models. In
weas of the
irborne and
resolutions.
lels internal
joint clouds
ited from
lata only.
In order to determine the internal accuracy, the 3D models were
created to the extent to which the data is available, separately
from the ALS and TLS data. Composition of both models is
shown in Figure 4.
This figure shows that both models intersect. Further,
a comparison between the lower edge of the roof shows that the
model created from airborne laser scanning data is placed
higher than the model created from terrestrial laser scanning
data. To quantify the level of mismatch between the two models
there were compared to each other: heights, horizontal distances
and spatial distances of 39 ("Object 1”) and 89 ("Object 2")
homologous points in both models. The results of this
comparison are shown in the Table 1. As the RMSE was taken:
(I)
Es es
| 2
24;
— 1 i=l :
n
where | c- RMSE
d; = difference (residue) between heights, planar
distance or three-dimensional distance of homologous
points
n = number of homologous points
“Object 1” “Object 2°
Horizontal
min. residue 0.001m 0.001m
max. residue 0.288m 0.365m
mean residue 0.082m 0.142m
RMSE 0.118m 0.166m
Vertical
min. residue 0.000m -0.231m
max. residue 0.277m 0.260m
mean residue 0.165m -0.010m
RMSE 0.188m 0.090m
Three-dimensional
min. residue 0.004m 0.007m
max. residue 0.353m 0.366m
mean residue 0.200m 0.166m
RMSE 0.222m 0.189m
Table 1. Results of internal accuracy assessment.
Presented in Table 1 values show that the three-dimensional
internal accuracy is about 0.2m for both investigated objects.
The horizontal accuracy for “Object 1” is over 0.1m but better
than for “Object 2”. Analysing the vertical accuracy for “Object
1” it can be found that model created from ALS data is placed
over model created from TLS data (positive mean value of the
residue). Completely different case there is for “Object 2” —
model created from ALS data is slightly below model created
from TLS data. Also the vertical accuracy in case of “Object 1”
is more than twice times larger than for “Object 2”. It proves
that matching of ALS and TLS point clouds should be executed
separately for object or even building.
Internal accuracy of the model can also be linked to the
generalization and detail of modelling (4.2 Error sources; point
5). Examples of generalized surfaces and visualization of
residues between scanning points and modelled (as planes)
surfaces are shown in Figure 5 and Figure 6. The maximum
differences reach even the value of 0.3m.
ogo
im]
Figure 5. Left: Distance differences between modelled plane
of roof and scanning points. Right: part of modelled roof.
0.30 -0.25 -0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 [m]
Figure 6. Upper: part of modelled facade. Lower: Distance
differences between modelled plane of facade and scanning
points.
4.3 Reference data
In the following discussion of the accuracy assessment, all of
listed in points 1 to 6 errors (4.1 Error sources) are considered
together. The final evaluation of the model's accuracy was made
by comparing the model with the results of an independent on
site measurement. The measurement was performed using Leica
TCR407Power reflectorless total station from warp points that
coordinates were determined with use of GNSS technique
supported by ASG-EUPOS system. Based on tachymetric
measurements there were determined coordinates of 354
referenced points: 75 for “Object 1” and 279 (on several
buildings) for “Object 2”. There were measured two types of
referenced points: for vector elements (corners of vector model)
and for texture elements (corners not present in vector model
but on textures). Example of referenced points location is
shown in Figure 6.
