CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
After solving the linear programming problem, gap edges
whose corresponding unknowns were determined to be 0 are
discarded. This results in consistent road subgraphs that
correspond to roads, which are considered to be the results of
road extraction. However, all road parts are kept at this stage,
so the falsely extracted road parts must be removed during the
road network formation, which is still under development.
2.4 Including the Digital Surface Model
The road part extraction can produce false positives among
segments having properties that are similar to road segments.
False positives can disturb the later steps of linking the road
parts and forming a road network. To avoid this, they have to be
sorted out later when the network is formed or they have to be
prevented from being extracted. The majority of the falsely
extracted road parts are buildings, which can lead to subgraphs
consisting only of false positives because buildings are often
arranged in rows. As the most distinctive property of buildings
that distinguishes them from roads is their height, a DSM can
provide valuable additional information.
The DSM is employed in the grouping and road part extraction
steps. It is not used for the initial segmentation which operates
at pixel level because DSM inaccuracies in shadows and
alignment errors caused by the fact that orthophotos are
generated using a Digital Terrain Model (DTM) would affect
the results adversely. In the grouping step the DSM is used to
prevent segments with different heights from being merged. For
this purpose, the differences of the mean heights are added to
the grouping criteria. If the difference is larger than a threshold,
the segments are not merged. The threshold is empirically
determined; in our examples it is set to 1.5 m. This prevents
building segments from being merged with road segments but
allows for smaller height variations in the background.
For the road part extraction the DSM is used to prevent high
objects from being extracted as roads. For this purpose, a
normalised DSM (nDSM) representing objects above ground is
determined. A coarse Digital Terrain Model (DTM) is
generated from the DSM by morphological grey value opening.
The DTM is then subtracted from the DSM, which yields the
nDSM (Weidner and Forstner, 1995). The mean heights of the
segments obtained from the nDSM are compared to a threshold.
It was found that a threshold of about 1 m reliably distinguishes
building parts and road parts. This threshold is used as
additional criterion in the road part extraction.
3. RESULTS
The approach was tested on three subsets of an image showing
a suburban scene from Grangemouth, Scotland. The image is a
CIR orthoimage with a resolution of 10 cm. The data set also
contains a DSM that was generated by image matching at a
resolution of 20 cm in position and 10 cm in elevation. Elevated
objects are represented well in the DSM, though unfortunately
it is not known which method was used for its generation. For
the three subsets, results of the road part extraction and road
subgraph generation are presented, first obtained from the
image data alone, and then from additionally using the DSM.
3.1 Results without DSM
Segmentation, grouping and road part extraction were carried
out as described in Section 2.2 and (Grote et al., 2007; Grote
and Heipke, 2008). Figures 3, 4 and 5 show the results of the
road part extraction for the image subsets 1, 2 and 3,
respectively. Whereas in subsets 1 and 2 most parts of the road
network were extracted, significant parts of the road network
are missed in subset 3. Each subset contains false positives,
which are mainly found on buildings because buildings have
similar radiometric and geometric properties to road parts. The
results of the road part extraction were compared to manually
extracted road regions. The manually extracted regions include
areas occluded by shadows or trees, but exclude pavements.
The completeness and correctness of the road parts computed
according to (Heipke et al., 1997) are displayed in Table 3.
They were determined on a per-pixel level and thus refer to the
extracted areas. Table 3 shows that about two thirds of the road
area could be detected, but almost half of the area classified as
road area consists of false positives.
Completeness
Correctness
Subset 1
66%
57%
Subset 2
89%
59%
Subset 3
31 %
49%
Total
62%
55 %
Table 3. Evaluation of road part extraction without a DSM.
Figure 3. Road parts extracted in subset 1 (yellow).
Figure 4. Road parts extracted in subset 2 (yellow).
30