Albert Baumgartner
In addition, the geometric accuracy of the extraction is assessed. It
is expressed as the RMS difference between the matched extracted
and the matched reference data.
The evaluation figures given in Tab. 1 show that the sequential com-
bination of module I and module II, i.e., of local and global group-
ing criteria, increases the completeness, and in the integrated com-
bination also the good correctness and RMS-values of the results of
the local level can be kept. The fact that the integration of global
grouping criteria enforces the extraction of a connected road net-
work is not expressed in the figures.
It should be noted that the networks resulting from the combined
approaches are inhomogeneous with respect to the geometric ac-
curacy since most of the resulting network originates from the hy-
potheses for road axes from module I and parts of the network orig-
inate from line extraction at 2 m resolution.
On principle, the evaluation results depend on the buffer width
which is chosen for the matching between reference roads and ex-
tracted roads. The larger we set the buffer, the more likely extracted
roads will be matched with roads in the reference. I.e., enlarging the
buffer can raise the correctness and completeness figures at the cost
of worser RMS-values. For the results given in Tab. 1 the buffer
width was set to 3 m, i.e., about half of the road width of an average road in the given image. In our case the influence of
the buffer width on the evaluation figures showed to be marginal. Only for very narrow buffers of less than 1 m the figures
change significantly.
Figure 9: Manually plotted reference network
Module I: Local | Module II: Global | I-II: Sequential | I-II: Integrated
(Sect. 3.1) (Sect. 3.2) (Sect. 4.1) (Sect. 4.2)
Completeness [%] 81 71 86 87
Correctness [%] 94 88 90 93
RMS [m] 0.42 0.83 0.52 0.45
Table 1: Evaluation results for different combinations of local and global module
6 CONCLUSIONS AND OUTLOOK
By means of global grouping criteria, the knowledge about the topological properties of roads is incorporated, and we
are able to overcome some deficiencies of purely local grouping. We showed that a noticeable improvement especially
concerning the connectivity of the resulting road network is possible with an integration of global grouping criteria. One
point which still should be improved is the weak model for junctions. By now we can handle quite simple junctions only.
In contrast to semi-automatic approaches, where the human operator can decide about the use of automation tools in each
case separately and, what is even more important, he can accept or reject the results immediately, fully automatic systems
must be able to decide on their own, where to start their search for specific objects and what to do next. To be an useful
semi-automatic tool it is more important to optimize the interaction with the operator, than having a sophisticated self
diagnosis algorithms, which are necessary for fully automatic systems. Fully automatic systems can base their decision
only on knowledge about the object, which is available in the system or on the information which they can derive from
the provided data. Basically, many parts the knowledge of the system can be reduced to a set of parameters, which might
be hard-coded or selected by the user. The most important requirement for a fully automatic system is to be able to cope
with a wide variety of data sets without major parameter modifications. The number of user selectable parameters should
be reduced to a minimum and the influence of these parameters on the result of the system should be easily predictable
and self-evident also for an unexperienced user. In our case the combination of the two complementary approaches for
road extraction seems to be one important step towards this goal. The objection that by combining the two modules we
perhaps replaced some parameters in one module with at least as many needed in the other module is not really true. Here,
we were able to reduce the influence of some quite sensitive parameters, and to improve the quality of the results, too.
Even if the overall number of parameters increases, for us it seems to be more important that we could eliminate some
crucial parameters. Something which is not expressed by external evaluation criteria is the lower sensitivity to predefined
parameter settings of the integrated combination compared to the local grouping as stand alone system. Of course these
findings have to be validated on different data sets.
64 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
Howe
scene
roads
and e
shadc
ACK
The v
the D
Deut:
Airau
meth.
Baun
Auto
Baun
Scale
Bord
Inter|
Verla
Fisch
imag
Harv
Meet
Heip
Phot:
Helle
Dem
USA
Hinz
Sock
Sens
May
of PI
McK
Visic
Rusk
of ca
Rusk
Netv
Steg
Autc
Steg
Man
Trin
Phot
Vası
Com
Wie
mue
Wie
Inte