|
Sponsored by:
|
|
The robot was driven
through the environment three times, one "clean" run
with as low noise level as possible, one "noisy" run with people
walking through the environment and one "home-tour" run, in which a
person is guiding the robot around the house. Each run took around 5
minutes.
Sensors:
Omnidirectional
images |
The omnidirectional images were taken by a
camera with a hyperbolic mirror. For
more information on the omnidirectional vision sensor, look here.
On average 7.5 images per second were taken amounting in +/- 2000
images per run.
|
Laser range
readings |
A SICK-laser (LMS-200) was mounted on the Nomad to record 180 degree
range scans at the front of the robot. Approximately 3.5 scans were
conducted per second.
|
Odometry
+sonar |
On average 12 odometry measurements per second were taken. Because the robot
has solid wheels the odometry is quite accurate. At the same time
the current values of the 16 ultrasonic sonar sensors were recorded giving a 360
degrees range scan.
|
For more information on the time-stamping, the format of
the files given for download and specifications of the sensors used,
have a look at the respective README's.
Annotation:
We will constrain ourselves here to a simple but still rich set of
spatial concepts. We start with the human concept of different
rooms in a home environment. Furthermore, within each room we
selected a number of prominent objects. The objects are manually
roughly segmented in the omnidirectional images. The region close
to an object is defined as the region from where the object can be
used in a manner common for that object. The data was annotated by
an inexperienced person. The person has never visited the
environment and relied on the omnidirectional image and a rough
drawing of the environment map. Every second frame was annotated.
See example movie(900KB). The annotation is provided
in XML format. The XML structure for one frame is described in frame.xml. .
Download:
The data is available via:
http://www2.science.uva.nl/sites/cogniron . Some useful MATLAB functions
are also provided: reading the annotation, geometric transformations for the omnicam,
etc. Detailed description is
here.
Credits: Following persons were involved in making this dataset (alphabetical order):
Anne Doggenaar,
Bas Terwijn,
Ben Krose,
Edwin Steffens,
Elin Anna Topp,
Henrik Christensen,
Matthijs Spaan,
Olaf Booij,
Ruben Boumans,
Zoran Zivkovic.
Special thanks to UNET for making their space available for the experiments.
|
Contact: Zoran Zivkovic zivkovic at science dot uva dot nl