site computer. The photogrammetric approach requires but one
camera and a flash. In terms of the time required to collect the
data the photogrammetric approach required but 20 minutes.
This represents an 18 fold saving on the estimated theodolite
triangulation case.
In this project it was fortunate that there was no requirement for
online coordinate determination. Had this been the case then
the digital photogrammetric method would have undoubtedly
proven to be unsuitable. Projects which require “real time”
coordinates will remain the domain of theodolite triangulation
systems until digital photogrammetry can match the theodolite
systems. The likelihood of this *real time" aspect developing is
remote given that by their very nature these systems are based
on single sensors and multiple exposures. What is perhaps
more plausible is the likelihood that these systems will become
streamlined to a point where the time lag between collection
and coordinate determination will be acceptable for the majority
of applications
IMPLEMENTATION
Having established a methodology it was necessary to verify
that the accuracy requirements could be met. This was
accomplished through the use of a network simulation (Fraser,
1984) to design and analyse the photogrammetric network.
Utilising approximate camera station positions and likely target
locations, plus the expected measurements and their respective
precision, it is possible to estimate the likely precision of all
coordinates. No actual measurements are required at this stage,
the results of the simulation are purely based on the geometry of
the networks and the types and precision of measurements.
The coordinate precisions obtained from the simulation are a
very useful diagnostic tool for the design, and re-design, of the
photogrammetric network. The specified tolerance for the
design dimension determinations was set at +0.5mm. To be
confident of meeting such a tolerance, the precision of
coordinate measurements should be of the order of £0.15mm at
the one sigma level.
Based on a network of 9 camera stations with 2 exposures at
each station the simulation yielded an object space accuracy of
the order of £0.1mm. Using the results of the simulation it was
possible to compute the minimum target diameter necessary to
satisfy the requirements for centroid determination. The size of
these targets is computed using the lens focal length, the
average object distance and the minimum desirable image
diameter. Based on a minimum target image diameter of four
pixels it was determined that 22mm diameter targets would be
required.
Once the object was available for data collection the targets
were applied to the pre-defined locations (see Figure 5). These
target locations were determined interactively during the
simulation process. The target placement is a mix of pre-
selected primary locations to define surfaces and circles, and
secondary locations selected on-site to strengthen the
photogrammetric network and fill the format of the still video
camera.
An unforseen complication was the location of the drive
components of the gantry crane. The location of these
components prevented an unobstructed view of the factory floor
on one side. This tended to bias the network of camera stations,
as the crane had to be driven further to one side than expected
to obtain adequate coverage of the hopper. For the first epoch
of photography this forced some rapid changes to the initial
network design. Additional camera stations were included to
compensate for what would otherwise be a weaker network,
leading to 11 stations with an average of 2 exposures per
station.
For the second epoch a new design was adapted to overcome
the physical limitations of the environment. The original
number of 9 camera stations was employed, but in this case
biased to compensate for the weakness detected in the first
network. In essence, more camera stations were used on the
more distant side of the network.
A small selection of frames were measured manually on site to
facilitate the computation of initial target and station
coordinates. These coordinates formed the basis of the
resection process utilised to measure the remaining frames
semi-automatically. The resection is almost entirely automatic
except for the initial target identification needed to orientate the
frame. All image locations were computed using an intensity
weighted centroid algorithm. Thresholds were computed in a
16 by 16 pixel window for each target using the pixel intensity
values (Shortis et al, 1994).
Following image mensuration, restitution via bundle
triangulation takes place. This least-squares estimation
operation essentially reconstructs 3D XYZ data from 2D image
measurements, while at the same time providing a self
calibration of the camera and the precisions of the target
coordinate data.
MEASUREMENT ANALYSIS
Due to changes in the design of the networks, changes in the
flash exposure intensity and changes in the quality of some
target images, the two networks gave markedly different results.
As can be seen from Table 1, the image space precision for the
second epoch is significantly improved.
Result Epoch 1 Epoch 2
Image space RMS error (pixels) 0.044 0.032
Number of digital images 22 18
Number of targets 63 62
Mean object space precision (mm) 0.16 0.17
Relative accuracy 1:58,000 / 1:50,000
Min. object space precision (mm) 0.07 0.07
Max. object space precision (mm) 0.56 1,01)
Table 1. Results of photogrammetric network computations
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
In an appare
for the sec
unforseen lo
the two epoc
oil. reducing
cleaning. ch:
some targets.
The precisio
ranged from
plate. which
object space
the accuracy
confidence.
specification
The results c
design dime
hopper. usin
some 6.3 mr
not an erro
network, spi
the separatic
excess separ
agreed with 1
The design c
tolerance, al
deliberately
of surface m
computed. 1
The flange :
milling macl
like measur
measuremen
photogramm
in the worst
within Imm
After the se
were verifie
flange had 1
height of the
critical flan
photogramm
reconciled, s
which result
this problem
within the +
Developmen
techniques v
metrology a
sensors and
