International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
drawing already exists. Therefore, significant research has been
devoted to convert such drawings to CAD representation. This
can be difficult due to the complexity and variety of drawing
notations and the necessity to handle noise and unwanted
features. Kernighan and Van Wyk, 1996, developed a technique
that can extract lines and poly-lines from digitized floor plans
and remove any information not needed for 3-D modeling.
3.3 Limited Geometric Constraints
Many of the automated modeling techniques mentioned in 2.3
rely on assumptions made about surface shapes and relations
between surfaces such as plane perpendicularity and parallelism
and the availability of vanishing points from multiple sets of
parallel lines. However, in medieval castles these assumptions
do not apply to most of the structures (see figure 1). This limits
the applicability of many automated techniques. Therefore, we
have to acquire the 3-D coordinates based only on overlapped
images without assumptions. The exceptions may be arches,
columns, doors and windows whose shapes are mostly regular
and can be modeled semi-automatically [El-Hakim, 2002].
3.4 Model Assembly and Integration
Models created from different sets of data and floor plans must
be assembled to create one model suitable for visualization. In
addition to differences in coordinate systems and scale, the
models will also not perfectly match at joint primitives such as
surfaces, edges, and vertices. Some of those will overlap or
intersect and some will be disjointed, which is unacceptable for
visualization. Again, commercial CAD and rendering software
do not address these problems. Therefore, a procedure must be
developed to seamlessly put together different models into one.
4. DETAILS OF THE APPROACH
Our approach is inspired by the Level Of Detail (LOD) concept,
used in visualization of complex models. Some city modeling
techniques also use an adaptation of this idea [Lee and Nevatia,
2003]. Our procedure is hierarchical by the data source (figure
3). In the hierarchy, the details, accuracy and reliability increase
as we advance from one data level to the next. As a rule, data in
one level overrides the data in previous levels.
*
Floor Plans
i
Models from Aerial
Images S es :
i os we wi sw en nn mw
A
'
i
Outdoor
Calibration
I
Models from Close- |
Range Images | "indoor
Details, Accuracy, Reliability
Models from
Laser Scanning
v a Sidra bai aT donat. eg d
Figure 3: Hierarchy of model assembly (dotted means optional).
The approach can be summarized as follows, noting that the
first five steps can be done in any order or even simultaneously:
Acquire a floor plan in digital form and create rough model.
Calibrate the digital camera for its internal parameters.
urvey some points with total station.
is
2.C
3-5
4- Acquire aerial images and create the overall model.
5- Create detailed models from terrestrial images and, when
available, a high-accuracy laser scanner.
6- Register and integrate the models created from sensor data.
7- Parts without data exposure are completed from floor plans.
Missing sensor data can be intentional for uninteresting parts,
or inevitable due to lack of access or improper coverage.
We detail some of the steps in the following sub-sections.
4.1 Camera Calibration
Self-calibration is necessary if camera settings are unknown and
vary between images. But to achieve accurate self-calibration,
certain geometric configurations of images are needed. Since
this is not guaranteed at the project site, and makes imaging
more restrictive, it is sensible to decide on high-quality camera
and take the images at fixed known settings. Many modern
digital cameras can save a number of settings. We then calibrate
in the lab at those settings using surveyed points (figure 4).
Figure 4: Calibration targets placed on walls, floor and ceiling.
4.2 The Overall Model from Aerial Images
Traditionally aerial images are used to model terrain and roof
tops. Here we suggest images from a low-flying helicopter at
various viewing angles around the castle to model roofs, outside
of buildings, most of the courtyards, surrounding walls, and
close-by grounds. The images should have strong geometric
configurations - being convergent with large base-to-height
ratio. Bundle adjustment is applied, using some surveyed points
to define the reference coordinate system and the scale.
4.3 Three-Dimensional Information from Floor Plans
3-D information from floor plans has two main purposes:
|. To add 3-D sections missed by imaging or scanning.
. To support the modeling process by verifying surface shapes
and the relationships between surfaces. Those are often not
obvious from images or even from visiting the site.
bo
To convert the floor plans into a 3-D model we need to provide
semantic information such as room identities and connecting
openings followed by walls extrusion to given heights. The
heights will be known from the image-based models. Window
and door insertions are carried out where needed.
4.4 Detailed Models from Ground Images
Images from ground levels are used to semi-automatically
create detailed models of selected elements such as entrances,
sections occluded from the aerial views, and indoor spaces. We
start by manually creating a less-detailed model with selected
seed points using bundle adjustment. Then, we select two
Internationc
images witl
hierarchical
(similar ide
automatical
constrained
automaticall
regular sha
created sem
described in
Figure 5. T
pair with aui
Combining |
two models
that the final
Relative s
Joint prin
from adja
unlikely t
adjacent n
in a detail.
Sensor-ba:
than perf
plans are «
exactly pl
surfaces bi
e No gaps, r
Commercial
Special techn
can be regis
max” using «
unnecessary
the same coc
however wil
need interact