Figure 6: Segmentation results on the test scene. On the left the results after initial classification are shown. On the right 
the result after region growing and removal of sliver regions is shown. 
The most dominant regions, i.e. regions above a certain 
size threshold, are selected as seed regions for a region 
growing process. Region growing is implemented as a 
morphological operation. A 3 x 3 mask is moved over 
the dataset. When a neighbor to the point of interest (the 
center of the mask) has a label assigned, the point of in- 
terest is checked for compatibility to that region. In case 
it is found to be compatible it is assigned the label of the 
corresponding region. If there are conflicting regions, i.e. 
there is different regions adjacent to the point of interest, 
the largest region is preferred. This is also the case if the 
center pixel is already labeled. 
The compatibility check is performed by a least squares fit 
to a second degree explicit polynomial as described above. 
If the error of fit is below a certain threshold the point is ac- 
cepted as compatible. The threshold has to be established 
beforehand, when evaluating the sensor system. The result 
of the region growing is shown in figure 6. 
5 CONCLUSION 
We have presented an efficient technique for the model- 
based segmentation of dense range scans. The joint use of 
model-based classification and region growing results in a 
reliable segmentation and overcomes most of the problems 
caused by misclassification. Curvature estimation still is a 
crucial part of the process and remains a topic of intense 
research. The proposed method is aimed at inspection and 
measurement task of industrial objects, but also has poten- 
tial for the application in the automated segmentation of 
laser scans. In the future we plan to extend the process 
by improving the compatibility check during region grow- 
ing. Since the surface type is assumed to be known from 
initial classification, the surface fit can be constrained to a 
specific surface type. 
REFERENCES 
Besl, P., 1988. Surfaces in Range Image Understanding. 
Springer- Verlag. 
Besl, P. and Jain, R., 1986. Invariant surface characteris- 
tics for 3d object recognition in range images. Computer 
Vision, Graphics, Image Processing 33, pp. 33-80. 
Bóhm, J., Brenner, C., Gühring, J. and Fritsch, D., 2000. 
Automated extraction of features from cad models for 3d 
object recognition. In: ISPRS Congress 2000, Vol. 33num- 
ber 5, Amsterdam, Netherlands. 
Brenner, C., Bóhm, J. and Gühring, J., 1999. Photogram- 
metric calibration and accuracy evaluation of a cross- 
pattern stripe projector. Proc. SPIE Videometrics VI 3641, 
pp. 164-172. 
Cantzler, C. and Fisher, R., 2001. Comparison of hk and 
sc curvature description methods. In: Third international 
conference on 3-D Digital Imaging and Modeling, IEEE, 
Quebec City, Canada, pp. 285-291. 
do Carmo, M., 1976. Differential Geometry of Curves and 
Surfaces. Prentice-Hall. 
Flynn, P. and Jain, A., 1989. On reliable curvature es- 
timation. In: IEEE Conference on Pattern Recognition, 
pp. 110-116. 
Karbacher, S., Labourex, X., Schoen, N. and Haeusler, 
G., 2001. Processing range data for reverse engineering 
and virtual reality. In: Third international conference on 
3-D Digital Imaging and Modeling, IEEE, Quebec City, 
Canada. 
Koenderink, J. and van Doorn, A., 1992. Surface shape 
and curvature scales. IVC 10, pp. 557-565. 
3442.