5. Istanbul 2004
lel of the object
uügh number of
' network taken
predictable by
servations and
ness to existing
sibility of each
not depend on
as the visibility
oint. The first
by defining a
| principal is a
'ctions by using
isibility can be
ne effects of the
nd hidden area
| the target and
fuzzy visibility
erfectly visible)
sation is shown
ion on VPS.
visible rays
ight) and
the positions of
| closer to these
. proper way to
nction like as
juency filter in
on 1. D» is the
ra stations. 7 is
width of fuzzy
transit areas.
stations are far
on object and
of the average
(1)
ow obstructions
per. Notably, a
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BS. Istanbul 2004
bad setting of these parameters leads to have a very unreal large
or small accessible area definition. In the next section we
briefly describe a fuzzy inference system for camera placement
which aims at satisfying the accuracy of a weak points as well
as vision constraints in an optimal way.
3. FUZZY INFERENCE SYSTEM (FIS) FOR
CAMERA PLACEMENT
Since there are several interrelated vision constraints and most
of them are highly uncertain, the camera placement is a subject
that fits well to the ideas of fuzzy inference systems (FIS). It is
noted that the efficiency of the FIS method depends
significantly on properly formulated rules and their parameters
(MATLAB, 2001).
Figure 5 shows the flowchart of the developed FIS for camera
placement. Input and output of the FIS are normalized values.
Input is a vector of constraint indicators and of distance and the
output includes distance variation, rotation variation, a
descriptor for enforcing camera position towards closest
existing camera stations to satisfy accessibility constraints and
another descriptor for enforcing camera situation toward the
most effective direction for improving precision. Both
descriptors are applied to the camera position and situation
before the next iteration us carried out. To select an optimum
estimate, the iteratively found estimates are evaluated with a
criterion based on accuracy and constraints violation. The
process leads to an optimal and stable state for camera
placement after a significant number of iteration.
Camera station Fitness Updating
(rotations & distance Optimum function
to the weak point) Oma calculation - lr
I tio and save Calculation of
Constraint indicators optimum eonection
state parameters
4 1
* Visibility * Distance var.
: eur Expert knowledge (rules) : Psion var
* No. of img pnt. E * Roll variation
* Point II e Closeness to
distribution 3 | | | closest camera
* Max distance Fuzzi | | Fuzzy. | |Defuzz| | station
(FOV &scale) ficationEJ erence ds e Direction var.
® Min distance H Engine Pification] © fpr accuracy
(DOF) improvement
* Distance
Figure 5: Camera placement flowchart by fuzzy inference
system
4. EVALUATION OF PROPOSED METHOD
The presented concepts and algorithms are implemented in
MATLAB code. For the experiments an ancient church
illustrated in Figure 6 is selected. Excavations and restoration
for this church are almost finished as can be seen in Figure 6.
The object has a fairly complex shape with hidden areas. It rises
{rom a deep pit and photographs had to be taken either from
outside or the bottom of the pit. These conditions restrict
modelling of the target visibility and camera accessibility. The
photographs are taken with a DCS420 digital camera under
incident solar radiation. The Vision Metrology software
Australis (Fraser, 1999) was used for point measurement, but
due to low quality of images most measurement had to eb
carried out manually. Figure 6 shows a sample of the network
comprising 215 object points and 57 images. The accuracy
fulfilment for the above network through camera placement
using 3D CAD model was not prosecuted. Instead out FIS
based method for placing optimal camera stations without 3D
CAD model requirement was tested.
Figure 6: Object (ancient church) and photogrammetric network
with several points of view
4.1 System Building and Adjusting
Since the proposed method is based on two main stages
including constraint modelling and camera placement, at the
first step the 3D coordinates of object points and their precision
as well as exterior orientation of camera stations including
camera self calibration is carried out with the Australis
software. The second step was VPS determination for each
object point to predict the visibility of every direction toward
each object point (Figure 7.Up) To model the camera
accessibility, we considered the parameters Do=5m and no=0.5
(Figure 7.Down).
