Boehm, Jan
AUTOMATED EXTRACTION OF FEATURES FROM CAD MODELS FOR 3D OBJECT
RECOGNITION
Jan BÖHM, Claus BRENNER, Jens GÜHRING, Dieter FRITSCH
Institute for Photogrammetry (ifp), Stuttgart University
Geschwister-Scholl-Strasse 24D, D-70174 Stuttgart
Jan.Boehm@ifp.uni-stuttgart.de
KEY WORDS: Object recognition, CAD, Modeling, Feature extraction
ABSTRACT
In this paper we report on our work on a CAD model-based object recognition system for industrial parts. We present
à novel approach which uses information derived from the CAD model in the early process of range image segmenta-
tion. This approach gives an exact classification of the range image enabling the development of a CAD based object
recognition system. We describe the feature extraction from CAD data and its use in the curvature based range im-
age classification. We carried out experiments with data from multiple sources. The results obtained are presented and
discussed.
1 INTRODUCTION
Our work is part of a large scale research project on optical measurement using sensor actor coupling and active ex-
ploration. This project is a collaboration of researchers from seven institutes of the University of Stuttgart including
mechanical engineers, electrical engineers and computer scientists. The goal of our work is to implement a measurement
system flexible enough to handle a large variety of objects from the industrial world. Using different types of sensors, e. g.
Mono cameras, stereo cameras and stripe projection systems, the measurement process is automatically tailored towards
the object presented to the system and measurement tasks specific to the object are executed. A CAD model for each
object forms the basis for measurement planning and assessment. Since we allow the objects to be presented to the system
in arbitrary position, one of the first steps of the measurement process has to be object recognition determining the pose
of the object. For our work we use 3D data, so called range images, as sensor input.
The task of object recognition can be formalized as follows: Finding the correspondence of features f; from the scene
S = (fi, fo, fn} With features F; from the model M = UA F5, ..., F4). The collection of features [5 D5 Fa}
form the description of the model. Model building by hand is a tedious and time consuming task which is also missing
repeatability and is therefore unacceptable in an industrial environment. In our framework a CAD model is a
vailable for
each measured object. Therefore it is evident that the model used for object recognition should be automatically derived
from CAD data.
In the past, several model-based object recognition systems for range ima
review can be found in (Arman and Aggarwal, 1993). Many of the systems
example, some systems were able to detect cylindrical objects,
edges which restricted them to polyhedral objects. In
shape, including rounded edges and free-form surfaces.
the fact that explicit three-dimensional CAD models are
not be used directly for object recognition, since it is o
ges have been reported. A comprehensive
rely on the presence of specific features. For
others were designed to handle planar surfaces with sharp
our context we are required to handle industrial parts of complex
Object recognition for industrial parts is a pretentious task, despite
available. One reason for this is that CAD models may in general
ften impossible to extract CAD specific high-level features from
eeded which forms the interface between CAD representation and
features which can be extracted from sensor data. In the work we present here, the CAD model is broken down into single
Segmentation or feature extraction of range images has been a popular research topic for the
have been developed including clustering, region growing and split and merge, see also (
problems became evident concerning reliability and curved surfaces.
with the segmentation of polyhedral objects using region growing (B
the edges of planar surfaces to terminate the growing of a region. Th
is also not extendible to arbitrary shaped objects.
past years. Several techniques
Hoover et al., 1996). But also
In our previous work we have obtained good results
óhm et al., 1999), However, this approach relies on
76 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.
Segment:
mum nur
of the rai
need a fe
surfaces
shadowir
is bound
from infc
We prope
has the p
the fund:
point acc
model sp
we alreac
Flynn an
planar, s
eithera p
Surface «
most wol
a methoc
that adva
use an ex
the CAD
shown be
While th
previous]
can be cc
2 DER
In this we
system h
system re
While th
shows th
A CAD 1
impleme
type and
Mean an
footprint
Le. the n
For exan
curvature
During it
paramete
