The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi XXXVII. Part B4. Beijing 2008
283
A thorough comparison of road tracking and path/network
optimization in terms of effectiveness and acquisition quality
has not been conducted up to now. Nonetheless, behavior and
acquisition time of the tracking algorithm of (Baumgartner et al.,
2002), for instance, has been extensively tested and evaluated
(Scholderle, 2005). Neglecting the time for data handling, geo
coding, and so forth, we experienced a reduction in plotting time
of up to 50% depending on the complexity of the scene. For
most rural scenes the time effort was reduced to 50%-70%. For
more complex scenes, i.e., for urban or suburban areas, the
performance of the tracking tool was too poor to compete with
snake algorithms. As expected, in urban areas the automatic
tracking failed very often, and putting the tracker back on the
road every few seconds is quite annoying and time consuming.
4. CONCLUSION AND OUTLOOK
Figure 7. User-assisted road tracking in aRADARSAT SAR image.
Yellow: tracked road sections; blue: user clicks.
Both strategies of semi-automatic road extraction have
complementary advantages and limitations. Ideally, it would be
possible to combine both in a unique framework. From a
methodological point of view, this would include that
appearance-based components of road tracking should be
incorporated into the optimization framework of network
snakes. This would help to utilize more knowledge about the
objects during optimization. Furthermore, it would allow to
evaluating the results obtained with snake algorithms - an issue
still not solved for the general case so far. Although first
attempts have been undertaken, a sound solution has not been
found for this so far. Nor it seems possible to estimate from the
data which strategy for user-assisted road extraction is
promising for a given scene and which not, so that the better one
could be provided to an operator. Therefore, our current concept
for road extraction in the context of disaster management is
designed to provide both modules to the operator and hand over
the choice of the appropriate procedure to him.
Especially under the light of today’s and tomorrow’s available
optical and SAR satellite systems, the development of integrated
approaches for object extraction from multi-sensorial images are
a substantial element to support fast and accurate information
extraction. To this end, models and extraction strategies need to
be developed that integrate the different geometric and
radiometric sensor characteristics attached with stochastic
models to accommodate for the inherent modeling and
measurement uncertainties. Despite of encouraging results, there
are still many fundamental questions to be solved for object
extraction, e.g.:
• Which type of modelling is appropriate to capture the vari
ability of the object classes, especially under the light of the
success of appearance-based approaches?
• Which relations between objects can support object extrac
tion, and which are more or less clutter?
• Is it possible to design a strategy that adapts itself to the
given extraction without loosing control over the computa
tional load? Or, is it better to start with a monolithic strategy
and incorporate dynamic elements or use generic search al
gorithms and apply heuristics to control the search space?
• Which decisions should be handed over to a human operator
and which can be done by the computer?
The challenges for research and development in this area are
laid out and well-known. Time will show whether they can be
successfully met in the long run.
Figure 8. Road network optimization in a SAR image. Top: Interactive
initialization; bottom: result after optimization.