CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
74
vd>4. The third parameter (z) is the tree height. Using the tree
mask from multispectral classification, we calculated the
minimum tree height as 3m. The fourth parameter (d) is the
point density. The minimum point density has been calculated
for the tree masked areas as 20points/ 25m 2 . By applying these
four parameters to the raw DSM Lidar data, the tree points have
been extracted and eliminated from all off-terrain points to
extract the buildings. The workflow can be seen in Figure 4.
Figure 4. Workflow of detection of buildings in method 4
The density of point cloud directly affects the quality of the
result. In addition, some tree areas could not be extracted
because of the low point density of the Lidar data. The accuracy
analysis shows that 84% of buildings area are correctly
extracted, while 100 of 109 buildings have been detected but
not fully extracted, the omission error is 17% .(Figure 5).
Figure 5. Building detection result from method 4. (Left: airport
buildings. Right: residential area).
5. ANALYSIS OF RESULTS
Each method shows similar performance with differences in
completeness. The reasons of the failures for correctness and
completeness of each method can be seen in Table 2. The
improvement of the results is performed by taking into account
the advantages and disadvantages of the methods.
Correctness Failure Reasons
Completeness Failure
Reasons
Ml
Airplanes/Other moving objects
/shadow on
vegetation/construction process
Vegetation on roofs, lack of
some parts of buildings
which are being constructed.
M2
Airplanes/Other moving
objects/construction process
Vegetation on roofs, shadow
on roofs, lack of some parts
of buildings being
constructed.
M3
Moving objects (esp. car series
in parking lots)/ other man
made structures (highways etc.)
Vegetation on roofs,
temporal difference with
reference data
M4
Tree groups which could not be
extracted and eliminated
Non-detection of small
buildings (problem related
to low point density),
detection of walls as
vegetation, temporal
difference with reference
data
Table 2. The reasons of the failures regarding correctness and
completeness for each method (M: Method).
Regarding completeness, the reference data has been generated
using aerial images, and some buildings are in construction
process. Reference data has been provided from Unique
Company and they have produced it using aerial images. But, in
the construction areas, these buildings were measured as fully
completed, although they were only partly constructed in
reality. This increases the omission error especially for the
results of the methods 1 and 2 which use aerial images. On the
other hand, due to the temporal difference between the
reference vector and Lidar data, the completeness of Lidar-
based methods (methods 3 and 4) has also been negatively
affected.
5.1. Combination of the methods
The results from each method have been combined according to
their failures for different types of objects. Intersection of all
methods gives the best correctness, while the union of the
methods gives the best completeness. The combination of the
results has been performed for achieving the best correctness
with the best completeness.
(1D2): While method 2 does not include the errors resulted by
the shadow on vegetation, the intersection of these two methods
eliminates the problem of shadow on-vegetation (in Figure 12,
Rl). The correctness of extracted buildings from this
combination is 86%, and the omission error is 12%.
(1D2) D4: This combination eliminates the airplane objects
from the detection result (Figure 6). Consequently, another
advantage of this combination is that it reduces the omission
errors which arise from the construction process on some
buildings, i.e. multitemporal differences. The correctness of
extracted buildings from this result is 96%, and the omission
error is 20% (in Figure 12, R2).