RATED 
ucl.ac.uk 
ied as a reliable and 
iscreet target point 
ts of interest dictate 
Ss paper is concerned 
face information by 
ion community. The 
texture. Processing 
: dynamic Delaunay 
xcessing. This paper 
] from point cloud 
jares the developed 
evaluation presented 
on and reliability of 
east squares bundle 
with assessing the 
reas of high surface 
parison of the point 
es constructed from 
[TON 
sation deals with the 
in a multiple image 
ed in a digital Close 
Metrology System 
> the retro-reflective 
ometry is recovered. 
o produce an initial 
2000), which is 
sed by applying an 
er, Gulch, 1987), in 
multiple image point 
ed for a number of 
a selected from a 
c properties. Precise 
with a least squares 
en, Baltsavias, 1988) 
meter starting values 
sponding image and 
conjugate points is 
e matching routine is 
onjugate point search 
is optimised through the use of epipolar geometry constrained 
by the bounding triangle in image and object space. This 
process generates a conjugate point list, which can then be 
refined through photogrammetric intersection to produce points 
with three-dimensional coordinates at a specified level of 
precision. Photogrammetric backdriving can be also used to 
compute candidate matching point locations once reliable 3D 
coordinates have been established by intersection. The results 
are imported into a bundle adjustment process to provide a 
rigorous evaluation of the network. 
Once the 3D point data have been densified, the Delaunay 
triangles are updated (Figure 1). The dynamic use of the 
Delaunay triangles provides a strong method for surface 
densification since it can be used to optimise the multiphoto 
matching as new point measurement data are introduced in a 
consistent triangulation between image and object space 
through the collinearity model (Papadaki et al, 2001b). 
The final product of the presented technique is a set of points 
and associated topological information. Edges extracted from 
the images (Deriche, 1993) can be processed in a similar 
manner. This is currently achieved by matching the beginning 
and end points of short edge segments. 
The triangulation network derived from the point cloud can still 
be lacking information in certain areas and be over defined in 
others. Accordingly the success of the approach is highly 
dependant on the existence of sufficient image texture 
information and the similarity of the radiometric properties in 
conjugate triangles. 
  
Figure 1: Triangulated dense point clouds 
3. PRECISION AND ACCURACY EVALUATION 
The data produced by the densification process is a 3D point 
cloud with known quality estimates. A first evaluation of the 
data precision and reliability is undertaken after 
photogrammetric intersection, which serves as an outlier 
rejection process and refines the point cloud. The 
implementation of the self-calibrating bundle adjustment allows 
for further evaluation of the ‘quality’ of the image 
measurements at all stages of the process. An evaluation of the 
data accuracy is made by comparing the photogrammetrically 
derived data with data of the same object surface provided by 
the CMM. 
Comparison of photogrammetric data to CMM data requires the 
definition of a common datum. The key limitation of not being 
able to directly measure retro reflective targets with the CMM 
was overcome by physically establishing a datum using four 
ball bearings of 12.6 mm diameter attached to the object. In the 
case of the CMM the ball centres are defined by multiple 
measurements using the CMM probe tip followed by a spherical 
best fit to the data. In the case of photogrammetric analysis, the 
reflective surface of the balls also acts as a retro reflector given 
axial illumination from the photogrammetric system. 
Intersecting lines of sight from multiple camera stations directly 
provides the coordinates of the sphere centres. Once the datum 
was established a small set of retro-reflective targets were then 
measured photogrammetrically to provide coordinates in the 
same system as that of the CMM. The estimated standard 
deviation for the target coordinates was 30 um. 
To evaluate the precision and stability of the different types of 
image measurements made, some thought needs to be given to 
assigning appropriate initial measurement weights. The measu- 
rements are not homogeneous and need to be processed accor- 
ding to an appropriate statistical measure, defined by the quality 
indication at each measurement stage. As a starting point, the 
weight attributed to the target image measurements has been set 
according to prior knowledge of retro target image measure- 
ment precisions for the camera systems used (Robson et al, 
1999). In the data sets presented in this paper a default weight 
for retro reflective target measurements was set at 0.5 um. 
The image measurements made by the multiphoto patch 
matching process developed in this research were given weights 
derived from the texture matching process. The matching 
solution provides standard deviations in x and y that vary from 
image to image and are used as weights for the measurements. 
The standard deviations reflect the internal stability of the least 
squares matching and are typically in the range of 1-1.5 um. 
The ‘natural’ points in the seed image, extracted by the feature 
operator, are assigned a mean weight from their matched 
corresponding points. 
4. EXPERIMENTAL DATA 
The data set used for this evaluation is an automotive gearbox 
housing, which is characterised by complex surfaces and depth 
discontinuities. The gearbox is metal casting and whilst smooth 
has some natural surface texture. One side of the gearbox has 
been imaged with two Kodak MegaPlus ES 1.0 digital cameras, 
which have a resolution of over one million pixels (1008x1018). 
This camera model outputs eight bit digital images with 256 
levels of gray resulting in high quality contrast detail ([1]). 
The retro-reflective targets defining the common CMM datum 
were processed in an image network acquired with a Kodak 
DCS460 (3060x2036). The measurement of retro reflective 
targets has been undertaken in vision metrology software 
(VMS) developed by (Robson & Shortis). These target 
coordinates are used in the datasets presented in this paper to 
ensure a common datum. 
The CMM data has been provided for certain areas of the 
gearbox (courtesy of NPL [2]). The analysis presented in this 
paper concentrates in point and surface comparisons ($4.2) for 
two particularly complex areas. The estimated accuracy of the 
CMM point data is 5-10 um. 
4.1 Data description 
Images of the gearbox have been acquired under controlled 
conditions in three different circumstances. The first involved 
the projection of laser dots on the gearbox surface (Figure 2). 
The laser dots were produced by a focusable laser diode emitter 
as a 19x19 grid. The image of the laser dots is similar to that of 
retro-reflective targets, although it is affected by speckle 
(Clarke, 1994). This data set was acquired under illumination 
conditions that yielded high contrast for the retro reflective 
targets and laser dots whilst eliminating virtually all 
—69—