^ for locating
tly advantages,
vith parametric
by GEMS data
ribe primitives,
natical models
2D to 2D, but
1 digital form.
tems. Thus this
teract between
algorithm, two
angle (see Fig,
es is much less
geometry to
| Scientists are
wide role in
nputer Science
ontributions to
a). The SIVE
line segments
on. [EEE Tran.
C Approach to
Line Features,
bot location by
451.
otogrammetry,
era parameters
vature, Pattern
ct motion in a
Automation, 3
| body motion
Graphics, and
lerpretation of
e, in Int. C. of
ity Press.
motion using
171-183.
tric Modeling,
ique for High-
-the-Shelf TV
Vol. 3, No. 4,
and structure
iqueness, and
6.
rs for Space
puter Sci. and
or Automatic
)bject. Ph.D.
nd Mapping.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 5. Hakodate 1998
CAD-BASED OBJECT RECOGNITION FOR A SENSOR/ACTOR MEASUREMENT ROBOT
Claus Brenner, Jan Böhm, Jens Gühring
Institute for Photogrammetry (ifp)
University of Stuttgart
Geschwister-Scholl-Straße 24, D-70174 Stuttgart, Germany
Ph.: +49-711-121-4097, Fax: +49-711-121-3297
e-mail: Claus.Brenner Q ifp.uni-stuttgart.de
KEY WORDS: industiral inspection, range images, object recognition, CAD
ABSTRACT
Seven institutes of the University of Stuttgart have applied successfully for the funding of a special research field to set
up and investigate a multi-sensor measurement robot for industrial close range inspection. The application was preceded
by the work of a research group which has shown the feasibility of the approach by setting up a measurement cell which
uses optical sensors and actors to identify and gauge industrial objects located in the measurement volume. This paper
describes the results that have been obtained so far and were demonstrated during a test run in 1997. It then focuses on
our latest developments concerning 3D data acquisition, registration, segmentation, model generation from CAD data and
object recognition.
1 INTRODUCTION
In the last 20 years we have seen dramatic changes in pro-
duction methods for industrial goods. The advent of general
purpose industrial robots made it feasible to have the same
robot for different tasks or across different products. It be-
came evident that, whereever practical, itis more econom-
ical to have non-specialized production units. This way, the
cost of change can be kept small after a product has been
redesigned. Nowadays, the situation is characterized by
two contrary developments: part complexity increases and
production lot size decreases. Time-to-market is more im-
portant than ever before. For an example, just take a look at
the ever-decreasing development cycles in the automobile
industry.
Apart from the changes in production, all other steps of the
product cycle are affected as well. Considering product de-
sign, parts are modeled using feature based and paramet-
ric CAD systems, which allow rapid changes. Part design is
evaluated at early stages for technical soundness by sim-
ulation methods. Aesthetic quality is judged by the early
fabrication of models that look almost like the final product,
which has become feasible using rapid prototyping tech-
niques such as stereolithography and vacuum moulding.
However, looking at quality assurance, we find that changes
have not been as dramatic as in other areas. Still, in many
cases part geometry is checked against the specification by
individually prepared gauges or specialized measuring sys-
tems. Sometimes, random samples are drawn and mea-
sured by coordinate measurement machines (CMM). Fac-
ing the trend towards a 100% quality control, it is obvious
that those techniques are too expensive and provide not
enough flexibility.
Optical measurement techniques, on the other hand, have
several properties which make them ideally suited for flexi-
ble gauging and inspection tasks: they are able to measure
thousands of points in a matter of seconds; they are appli-
cable to a wide range of materials, including deformable ob-
jects; and they can yield very accurate results when used in
conjunction with proper calibration techniques. Moreover,
optical techniques can capture other important object fea-
tures like transparency, color and surface gloss. And, since
the object is captured anyways, some simple vision tasks
like reading a barcode on the object or checking for com-
pleteness can be done without the expense of additional
sensors.
Despite their advantages, optical measurement techniques
are not very well accepted in industry (Grün, 1994). One
reason for this is that traditional measurement techniques,
like CMM's, are well established whilst optical systems
with comparable performance have not been commercially
available until recently. Also, properties like surface rough-
ness today are defined in terms of CMM measurement re-
sults, which necessitates the definition of an “optical equiv-
alent” before market acceptance can be expected. An-
other drawback is that optical techniques are considered
to be too complicated to operate under factory conditions.
Furthermore, optical measurements often give accuracies
which depend on the specific object. In unfavourable cases,
for example if an object's surface is soiled, measurement
may become impossible using fixed sensor and lighting po-
sitions. But changing these conditions (e.g. by changing the
sensor, lighting or object positions) usually requires some
skilled person familiar with that particular measuring sys-
tem. Therefore, the measurement systems that have made
their way into industrial applications usually use very sta-
ble features such as retroreflective targets (Beyer, 1995) or
very well controlled environments (Bósemann and Sinnre-
ich, 1994).
However, in our opinion, the slow industry acceptance of
optical 3D measurement techniques is not a vote against
those techniques but rather reflects the standard learning
process in industry. Unfortunately, heavy competition and
outsourcing of product development to supplier companies
often limit the research horizon to one or two years. On
the other hand, we see a parallel to vision systems used in
2D inspection such as number recognition and complete-
ness tests. Due to technical progress in the fields of cam-
era (CCD and CMOS high dynamic range technology, "in-
telligent" cameras with integrated processing) and software
technology (standardized image processing modules, re-
209
