![](bg5.png)
To determine the scale, the 30 lowest features were
located and evaluated with the method described in
sec. 3.5. In fig. 7 the resulting features are shown, the
bigger red dots indicate the assumed ground plane.
It is seen that a few lower points are filtered out. The
height of the camera system is also seen on the y-axis.
The measured height of the system is 140 cm. The
vertical field of view is also displayed well.
In fig. 6 our method’s output can be visually com-
pared to that of the laser scanner. The middle im-
age is one of the tracked images, the left image is
our method’s output and the right image is the laser
scanner’s output. The circled areas are correspond-
ing areas showing obstacles. It can be seen that our
method gives much more information about the sur-
rounding area than the laser scanner, it may though
appear noisy. This is due to that features from up to
108
◦
vertically are projected down on to the ground
plane. Looking at the top left ellipse in fig. 6 it is seen
that the laser scanner only detects very few features
around the table and chair while our method is able
to detect the full shape. It can also be seen that the
feature tracker we use is somewhat limited to sharp
edges. Compared to the laser scanner it still detects
more features and obstacles in the scene. When the
laser scanner is limited to one half plane our algorithm
detects obstacles in a bigger area of the sphere, i.e.
the lower left ellipse in fig. 6. It can also be seen that
while the laser scanner is more accurate, our method
detects more features on every obstacle, i.e. a more
reliable detection of obstacles.
Figure 7: Side view of the 3D-positions of the features.
The bigger red dots indicate features on the assumed ground
plane.
5 CONCLUSIONS
This paper presents a successful approach of deter-
mining absolute 3D-distances in the scenes. The pro-
posed method uses a single omnidirectional camera
to compute absolute distances in the scene. The exact
scale is determined by assuming that the robot moves
in one plane and measuring the height of the cam-
era. The height is easily accessible, which makes our
method cheap and easy to implement. Our algorithm
is compared with the output from a SICK laser range
scanner. The results show that our approach gives
more information about the scene and is able to locate
more obstacles than the laser scanner. The laser scan-
ner is more accurate but limited to data in only one
half plane, while our algorithm is able to detect obsta-
cles in 360
◦
×108
◦
of the surrounding sphere. This
makes our approach more reliable in ways that it de-
tects a wider range of obstacles, e.g. table tops, over-
hanging cupboards etc. The current version might not
be accurate enough for industrial applications, but for
service applications, e.g. wheelchairs, the accuracy is
not crucial. It is though a good candidate to replace
the laser scanner with for some applications.
Advantages of a laser range scanner compared to
our method are as follows; It detects obstacles regard-
less of translation, speed or rotation. It works in any
light condition and it detect obstacles regardless of the
texture of the scene. Our method relies on detecting
corners. In order to detect corners our system needs
scene with texture.
The speed of our method is currently 1Hz. This
can be seen as the lower bound since we have not yet
performed any optimisation. The optimised algorithm
will have a speed comparable to the laser range scan-
ner’s 5Hz.
Future work includes, apart from optimisation, to
combine our algorithm with the laser scanner and in
this way make it reliable enough for industrial appli-
cations. We also need to perform additional experi-
ments to see if the wheelchair can navigate solely on
the camera.
ACKNOWLEDGEMENTS
This work has been supported by the Inter-University
Attraction Poles, Office of the Prime Minister (IUAP-
AMS), the Institute for the Promotion of Innovation
through Science and Technology in Flanders (IWT-
Vlaanderen), the Fund for Scientific Research Flan-
ders (FWO-Vlaanderen, Belgium), and the K.U. Leu-
ven Research Fund (GOA-MARVEL). A. H
¨
untemann
is research assistant of the research foundation Flan-
ders FWO.
REFERENCES
T. Goedem
´
e , M. Nuttin, T. Tuytelaars, L. Van Gool 2005
Omnidirectional Vision based Topological Navigation
15th International Symposium on Measurement and
Control in Robotics, ISMCR 2005, 2005.
A. N
¨
uchter, K. Lingeman, J. Hertzberg and H. Surmann
2005 Heuristic-based laser scanner matching for out-
ICINCO 2006 - ROBOTICS AND AUTOMATION
370