ut 
0] 
es 
le 
IS 
ve 
SO 
ed 
ct 
nd 
ot 
is 
239 
3 RANGE DATA ACQUISITION AND ANALYSIS 
3.1 Data acquisition 
Range data used in this work were obtained with range sensors relying on the so-called coded structured-light 
approach. A sequence of stripe patterns is projected onto a scene and images thereof are captured by CCD- 
cameras and analyzed afterwards. The result of this processing is an array of 3-D points in a coordinate system 
which, in our case, is identical with the robot world coordinate system. Several sources of errors which reduce 
the attainable precision were uncovered and compensated for: intensity dependent localization of stripe edges, 
slightly differing widths of black and white stripes of the projector and multiple reflections in the scene. [Trobina 
and Leonardis 1995]. 
When doing object recognition with the object model based approach, visual data from typically one view are 
available. Features or structures are extracted and matched with corresponding features and structures in a 
model data base. When successful, more information (mostly geometrical) can be extracted from the object 
model and exploited. If we disregard object models like we do, geometrical properties must be extracted directly 
from the data. This is because we have virtually no a priori knowledge about the world the robot is working in. 
As we are using a two-fingered parallel gripper, it is mandatory for our system to see opposite surface patches. 
Consequently, the optimal use of the two range sensors is crucial for achieving our goals. Therefore it is 
important to determine the best data acquisition geometry. For obvious reasons the configuration of the range 
sensors about the vertical axis was chosen to be symmetric. Parameters to be optimized were the angle between 
a projector and its camera and the angle between the two range sensors as a whole. Criteria for the optimization 
of the first of these parameters have been the accuracy of the range measurements and the range of slopes for 
which data can be obtained. Based on experiments on the dependence of noise on the slope of a surface element 
and the sampling geometry, we chose an angle of 15 degrees. On the other hand, the angle between two range 
sensors depends on the desired degree of overlap of the two fields of view (which is important for the merging. 
process) and the capacity to see “below the equator” of objects. A good compromise was found in letting the 
sensor axes diverge by about 80 degrees. Figure 1 shows the configuration of the range sensors schematically 
together with a spherical object. 
Projector 1 Projector 2 
B6 Area reliably seen 
by at least one sensor 
Area unreliably seen by 
sensor 1 or sensor 2 
(O Area not seen by sensor 1 
nor sensor 2 
  
  
JZ7777777/71/f4f1f 
Figure 1: Data acquisition geometry with two range sen- 
sors: Above: Schematic drawing of the geometry. The shaded 
areas roughly show the acquisition error. Right: Experimental 
setup at our lab. 
  
3.2 Merging Several Range Views 
A single range view is an array of depth measures which is representable as a graph surface z — g(z, y). This is 
called a 21-dimensional description of the scene. Using two range sensors to obtain a more complete description 
of the scene gives rise to delicate problems: 
- The input data cannot be represented any more by a unique explicit function, i.e., z — g(x,y). Therefore 
the transition from a 21-D representation to a true 3-D description of the surface is necessary. The 
neighborhood relations between points acquired by different sensors are not at all obvious. 
- The surface descriptions of several sensors can overlap. In these regions a consistent representation of the 
surface must be achieved. Accuracy and noise characteristics of the sensors must be considered. 
The first problem was solved by converting each single view into a true 3-D geometric description, i.e., a 
triangular tessellation of the graph surface. The second one is taken care of by the first step of the merging 
algorithm, the “mutual approximation”. 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995