Hi,
below is a text I intend to include in a report. Do I get the basic ideas right?
Regards,
Rene
Object handling consists of several steps, which are performed by the
tabletop_object_detector (and other) nodes.
• Get an image of the object (e.g. a pool ball) from e.g. a stereo
camera with texturing or e.g. a laser range finder.
• Transform the image (either a stereo-photo or a “sphere” of
distances) into a point cloud. A point cloud is a collection of
points, where all points have a number of pre-defined attributes, like
spatial coordinates, color, or neighborhood density. Point clouds are
the universal data structures for performing 3-dimensional algorithms
within ROS. TODO: figure out the algorithm
• Transform the point cloud into a wireframe. TODO: figure out which
algorithm is used for this.
• Now that an approximate wireframe model of the object is available,
(because the camera images are probably imperfect), this model can be
compared to a database of existing, perfect wireframe models. Because
the models are stored as blobs in the database, the approximation is
done in the ROS node. The database is thus only used a storage
backend.
• The comparison algorithm, given the actual model at one side and the
database models at the other side, is able to calculate the likelihood
that the presented object matches a or multiple objects in the
database (e.g. for a ping-pong ball, it could say “90% pool ball, 8%
bottle cap, and 2% unknown”).
--
http://www.rene-ladan.nl/
GPG fingerprint = E738 5471 D185 7013 0EE0 4FC8 3C1D 6F83 12E1 84F6
(subkeys.pgp.net)
(text/plain)