Figure 1: View of the measuring cell with actor modules (background) and general purpose computing hardware used for 
sensor data processing and system control (foreground). 
in order to obtain favourable capturing conditions. Thus, in 
connection with sensor fusion, active exploration promises 
to increase the operability of optical measuring systems un- 
der factory conditions. 
2 THE EXPERIMENTAL MEASURING 
SYSTEM 
2.1 The currently used sensors 
To investigate concepts on active exploration and multisen- 
sor fusion, an experimental measuring system is under de- 
velopment (see Fig. 1). The system currently uses 
e a laser projector which can be used with the coded 
light approach for range image measurement and as 
a point triangulation sensor. With the ability to gener- 
ate variable light patterns the laser can also be used 
for texture projection to support stereo image analy- 
sis. 
e a multi-parameter color CCD-camera which has 
the ability to change parameters like focus, focal 
length and aperture and several electronic parame- 
ters based on the interpretation of the current image. 
e a stereo camera. 
e several light source arrays consisting of regularly 
spaced individually controllable light spots located at 
the edge of the measuring volume. 
The sensors are mounted on three actor modules with a 
total of 13 axes. The size of the measuring volume is about 
1000x 1000 x 700 mm?. 
60 
2.2 On the role of recognition and pose estima- 
tion 
Object recognition and location is an important part of the 
overall system concept since it increases the flexibility of 
the measuring system. Of course, there are inspection 
tasks in which the object is known a priori and there is no 
need for recognition or identification of the specific object. 
In this case the task is to determine the object’s pose which 
allows to avoid object specific conveyors needed to position 
the object accurately in the measuring volume. But may be 
already in the near future the ability to recognize objects will 
be the conspicious feature of such systems. Object recog- 
nition aims at finding out the object type among. a library 
of possible types, but can also be used to identify parts of 
objects. 
In the next paragraph we discuss how sensor fusion and ac- 
tive exploration can be exploited for the purposes of object 
recognition and point out our basic recognition concept. 
3 OBJECT RECOGNITION 
3.1 General aspects of object recognition 
Object recognition can be defined as the problem of as- 
signing the correct label to an object present in a scene. 
For that, an object has to be detected and may be loca- 
lized and then identified by comparison with a given model 
of the object. In simple cases, global approaches can be 
used. Usually, global properties like volume, roundness or 
higher order moments are computed, forming a vector of 
parameters. Matching objects to models then reduces to a 
comparison of parameter vectors, i.e. matching is done in 
parameter space. In general, however, global methods are 
not considered to be robust enough, particularly in the pres- 
ence of occlusion and clutter (Grimson 1990). Thus, many 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
Figure 2: (: 
scheme. 
object reco: 
ric primitive 
matching tl 
a number o 
object and i 
Fig. 2(a) st 
recognizing 
Jain 1991). 
ultimately r 
this domain 
specific or 
representat 
Fig. 2(a) ar 
essary in tl 
to sample tl 
quiring sen: 
fact that the 
ber and are 
compute (E 
ule has to c 
thus each | 
ing module 
model-spec 
ing a usefu 
so far not p 
matching, | 
CAD object 
Since it is 
and many 
CAD/CAM 
CAD descr 
Horaud & F 
& Jain 19€ 
cal feature 
lists. Exam 
tion of line 
given in the 
which a cir 
between ge 
tion of prin