ACCURACY OF DENSE SURFACE MEASUREMENTS IN AN INTEGRA TED
PHOTOGRAMMETRY AND MACHINE VISION FRAMEWORK
H. Papadaki*””
? UCL, Department of Geomatic Engineering, Gower street, WCIE 6BT London, UK - hpapadak@ge.ucl.ac.uk
Commission V, WG V/1
KEY WORDS: Dense surface representation, Delaunay, measurement quality, precision, reliability, texture
ABSTRACT:
The increasing use of digital close range photogrammetry in engineering and metrology applications is established as a reliable and
dependable method of high accuracy measurement. Photogrammetric processing traditionally relies on discreet target point
measurements to achieve very high accuracy 3D results. However, the shape and surface complexity of the objects of interest dictate
the need for more dense surface measurements than can be achieved using targets. The research described in this paper is concerned
with the development of an automated photogrammetric measurement method that can provide dense object surface information by
integrating target measurements with image processing algorithms, some of which eminate from the machine vision community. The
method can produce a dense cloud of object surface points provided there is sufficient surface and image texture. Processing
involves the automated densification of a sparse triangle network created by the targets to a surface model using dynamic Delaunay
triangulation and geometric image constraints. The quality of the data is monitored throughout the stages of processing. This paper
examines the accuracy of the produced point cloud by comparing the data with surface models derived from point cloud
measurements provided by a Coordinate Measuring Machine (CMM). Further performance evaluation compares the developed
method with pattern and laser dot projection under controlled conditions.
1. INTRODUCTION
Coordinate Measuring Machines (CMM) are currently the most
established and reliable measurement systems in industrial and
engineering metrology applications. Such systems are used for
very high accuracy point based measurements and are typically
applied to industrial quality control and mass production part
inspection (Boeseman et al, 2001). However, CMM systems are
expensive and inflexible in terms of their need for direct surface
contact and measurement volume limitations.
The automated measurement of engineering surfaces from
multi-station close range photogrammetry can provide dense
point clouds of known quality. The density of the point cloud
can be directly comparable with the ones provided by some
laser scanning solutions for close range applications (Beraldin
et al, 2000). The research described in this paper is aimed at
developing a flexible automated photogrammetric measurement
method, which can create dense point data of an engineering
standard. The method is concerned with the construction of a
set of tools to densify a sparse triangle mesh, created by retro
reflective targets in a photogrammetric network (Papadaki et al,
2001a). The data is required to be dense enough to produce a
valid representation of the object surface with a known degree
of accuracy, precision and reliability.
This paper is principally concerned with evaluating the
accuracy of the photogrammetrically derived dense point cloud
and surface model by comparing the results with measurements
provided by a CMM. Further evaluation of the measurement
quality is undertaken by integrating the method with pattern and
laser dot projection. It is shown that the developed method can
be successfully applied in cases of low image and object surface
texture by projecting artificial texture patterns. Additionally,
surface measurements from the densification method, using the
existing image texture, are compared with measurements from
laser dot projection as well as those from artificial texture
projection.
Consequently there are two aspects to the evaluation presented
in this paper. The first involves the precision and reliability of
the point clouds as derived from the least squares bundle
adjustment. The second is concerned with assessing the
accuracy of the point cloud data in local areas of high surface
complexity. This is achieved by direct comparison of the point
clouds to small reference surface patches constructed from
CMM point data.
2. METHOD DESCRIPTION
The developed approach to model densification deals with the
problem of image point correspondence in a multiple image
network. The network set-up is implemented in a digital Close
Range Photogrammetry System (Vision Metrology System
VMS (software), Robson, Shortis), where the retro-reflective
targets are measured and the imaging geometry is recovered.
The set of target points is triangulated to produce an initial
Delaunay triangle mesh (Woodhouse, 2000), which is
consistent in both image and object space.
The process of densification is initialised by applying an
interest operator, such as Forstner (Forstner, Gulch, 1987), in
one image to create a seed point cloud. A multiple image point
correspondence can be reliably established for a number of
these seed points by defining criteria selected from a
combination of radiometric and geometric properties. Precise
correspondence is subsequently achieved with a least squares
based image patch matching routine (Gruen, Baltsavias, 1988)
that has been modified to use affinity parameter starting values
derived from differences between corresponding image and
object triangle shapes. The number of conjugate points is
increased in a series of stages, in which the matching routine is
applied to all images in the network. The conjugate point search
—68-
is Of
by t
proc
refin
with
preci
com]
coori
are 1
rigor
Once
trian;
Dela
densi
matc|
consi
throu
The |
and :
the i
mann
ande
The t
be la:
other:
deper
inforr
conju
Figure
3.
The d
cloud
data
photos
rejecti
implet
for f
measu
data a
derive
the C^
Comp:
definit
able tc
Was Oo
ball be
case Q
measu
best fi
