CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
56
extraction of buildings (Figure 7). The area was calculated for
each building after conversion to vector format and used as a
threshold for separating main buildings from smaller structures
in front or behind the main buildings. These areas also include
trucks and vans on the roads or parked near to buildings and are
detected as buildings. These areas were retained and classified
later into buildings, vegetation or other objects.
Figure 7: Extracted Buildings
To assign each building an individual height, building centroids
were determined. For each centroid, heights were determined
from the DSM by bilinear interpolation and the same process
was used for determining vegetation height.
For the purpose of vegetation extraction, the final building data
was subtracted from the NDSM. This resulted in vegetation
present in the NDSM layer that is higher than 2.5m. However,
the filtered multiple echo data (Figure 5) was also processed
further. First, intermediate and last echoes represent reflections
from the edges of buildings and trees. Once buildings were
classified, building boundaries were used as an input to remove
all multiple echo points that belong to building edges. Multiple
reflections from large trees, together with compactness
(area/perimeter 2 ) were used to classify large single trees and
groves (Figure 8).
The remaining objects to be classified were roads which are
part of the generated DTM. NDVI data below a threshold of
0.1 and the previously classified objects were used. Using the
threshold eliminates most of the area having vegetation but
does not help much in the areas with barren land. Their spectral
signature value is also very close to the roads. Even for the
roads the reflection value is not constant. It varies with age and
type of material used in the road surface. Previously extracted
buildings and vegetation were subtracted from the NDVI to
extract road candidates.
Gaps are present in the extracted roads and this is because of
the trees and building shadows. This section of the road is not
visible in aerial photos and LiDAR and needs further research
for its successful extraction. Finally all the extracted objects
were combined together and integrated into a specialized GIS
system (Figures 10 and 11).
4. CLASSIFICATION ASSESSMENT
Results of the workflow were evaluated after the method of
Heipke et al. (1997). Using this method, three different states
for any feature can be identified (Hatger, 2006).
True positive (TP) - A phenomenon that is present within the
input data and that has successfully been identified within the
output data.
False positive (FP) - A phenomenon that is not present within
the input data but that has falsely been found to be a
phenomenon by the algorithm. Thus it is written to output data.
False negative (FN) - A phenomenon that is present within the
input data but that has not been identified by the algorithm and
therefore has been omitted from the output data.
Then we define Completeness, Correctness and Quality by
Completeness = TP / (TP + FN)
Correctness = TP/ (TP + FP)
Quality = TP / (TP + FP + FN)
Figure 8: Extracted Trees and Groves