The following film shows how an object is identified using a model stored in the JANUS model database. The method is similar to the one described in A model based approach to recognition and measurement of partially hidden objects in complex scenes.
In the first phase, the contours of the interesting objects are
detected. This is done by a grid search over the picture. Afterwards
a flood fill method is applied to get contours of the objects.
In the next phase, a model based fitting is applied to to selected
objects parameters (distance, orientation, ...).
As a result of this, one gets classification and a scale model
of the object, this includes position and orientation in the camera
coordinate system.
To enable the observation everything is much slower than in reality.
The methods described in
A model based approach to recognition
and measurement of partially hidden objects in complex scenes
have been
further improved, so that we achieve 1-3 seconds on a single
PC out of our PC cluster. This
time is still quite long, but there is a rich set of promising
heuristics,
which we have not yet included into the speed up process.
Once a good first hypothesis has been found, the fine iteration is very
fast (0.05 sec). In the video this results in a very fast fitting at
the
end of the iteration. Therefore tracking of an already recognized but
moving object should be possible in real time.