![](bg6.png)
the segmentation distance threshold s
d
is adapted to
prevent future false segmentations. As the reliability
of a perception depends on its perspective, the new
distance threshold s
f d
is defined as a function of the
perspective, strictly speaking, of the angle between
surface normal and the connecting line of the position
of the surface part and the sensor. Initialising s
f d
with
s
d
for all angles, the value assigned to a certain angle
has to be reduced in case of under-segmentation and
to be increased in case of over-segmentation. Analo-
gously, a threshold function h
f d
has to be applied for
object hypothesis generation instead of h
d
.
4 FUTURE WORK
Currently, we are working on the object hypothe-
sis generation, which partially includes the proposed
segmentation. Thus, the main parts of the concept,
class generation and feedback, have still to be set up,
verified and evaluated. For the sake of simplicity,
we will restrict our system to a static environment,
which would not result in loss of generality. Addition-
ally, we will start feedback verification with simply
shaped objects like boxes and self-generation of ob-
ject classes with a set of well distinguishable shapes
with slight variations within each class, like those il-
lustrated in figures 2 and 3. The final part of our fu-
ture work will be the evaluation of the overall system.
For further improvement of the classification scheme,
it will include the feedback of shape part descrip-
tiveness, derived while generating the shape class de-
scriptions, into the hypothesis classification step.
5 CONCLUSIONS
In this paper, we introduced our idea for a joint seg-
mentation and classification with feedback. We are
confident that our approach can significantly con-
tribute to a more robust as well as a more autonomous
object classification thus overcoming traditional clas-
sification methods which either rely on an initial train-
ing set or some other specific information. As this
kind of information has a priori to be provided by a
human, the robot cannot act truly autonomously. Ad-
ditionally, this kind of information is difficult to ob-
tain even for humans, given the incompletely-known
and ever-changing environment, in which the robot
typically operates. In contrast, our approach au-
tonomously generates object classes. Consequently,
we expect a considerable improvement of autonomy
due to our self-generating object classes fom environ-
ment perceptions based only on a simple set of rules.
Furthermore, the joint segmentation and classification
feeds back classification results into both the segmen-
tation and the object hypothesis generation, and thus
is able to prevent many cases of over- and under-
segmentation, which typically occur due to incorrect
assumptions and thresholds in both segmentation and
hypothesis generation.
REFERENCES
Chen, Z. and Ellis, T. (2011). Multi-shape descriptor vehi-
cle classification for urban traffic. In Proceedings of
Digital Image Computing: Techniques and Applica-
tions. IEEE.
Endres, F., Plagemann, C., Stachniss, C., and Burgard, W.
(2009). Unsupervised discovery of object classes from
range data using latent dirichlet allocation. In Pro-
ceedings of Robotics: Science and Systems. The MIT
Press.
Farmer, M. and Jain, A. (2004). A wrapper-based approach
to image segmentation and classification. In Transac-
tions on Image Processing Volume 14 Issue 12. IEEE.
Guyet, T., Garbay, C., and Dojat, M. (2007). A human-
machine cooperative approach for time series data in-
terpretation. In Proceedings on Artificial Intelligence
in Medicine. IEEE.
Huhle, B., Magnusson, M., Strasser, W., and Lilienthal, A.
(2008). Registration of colored 3d point clouds with
a kernel-based extension to the normal distributions
transform. In Proceedings on Robotics and Automa-
tion. IEEE.
Knopp, J., Prasad, M., and Van Gool, L. (2010). Orien-
tation invariant 3d object classification using hough
transform based methods. In Proceedings of the ACM
workshop on 3D object retrieval. ACM.
Lai, K. and Fox, D. (2009). 3D laser scan classification us-
ing web data and domain adaptation. In Proceedings
of Robotics: Science and Systems. The MIT Press.
Lecumberry, F., Pardo, A., and Sapiro, G. (2010). Simul-
taneous object classification and segmentation with
high-order multiple shape models. In Transactions on
Image Processing Volume 19 Issue 3. IEEE.
Marton, Z.-C., Pangercic, D., Blodow, N., and Beetz, M.
(2011). Combined 2d-3d categorization and classifi-
cation for multimodal perception systems. In Interna-
tional Journal of Robotics Research. SAGE Publica-
tions.
Schwartzkopf, W., Bovik, A., and Evans, B. (2005).
Maximum-likelihood techniques for joint
segmentation-classification of multispectral chromo-
some images. In Transactions on Medical Imaging
Volume 24 Issue 12. IEEE.
Triebel, R., Shin, J., and Siegwart, R. (2010). Segmenta-
tion and unsupervised part-based discovery of repet-
itive objects. In Proceedings Robotics: Science and
Systems. The MIT Press.
AJointSegmentationandClassificationofObjectShapeswithFeedbackfor3DPointClouds
393