Fac. Science |
IAS | W. Zajdel | Demos
Below, you will find some movie clips demonstrating the techniques
that we develop for our project. All files are MPEG-1 video.
- tracking and identification with multiple static cameras
This clip contains three simultaneous video streams from three
cameras. We have placed the cameras in the faculty building, in such a
way that their fields of view remain disjoint. In the bottom panel,
you will see five people appearing in the camera views. The goal is to
track these people locally (that is: within the filed of
view) and globally - that is: re-identify a person when
he/she moves between cameras. The estimated identities of people are
indicated with numerical labels. In the top panel the ground-truth
labels are given. See the ICPR 2004
paper for details.
- tracking and identification from a mobile robot
Here we have a camera placed on a mobile robot that roams through the
building. The task of the robot is to keep track of the people it has
seen, even after they (temporarily) disappear from the view. The two
clips present two runs, each with three people appearing multiple
times in the view. Both runs have taken place in a relatively small
area, but with significant illumination changes. The estimated
identities of the people are indicated with numerical labels. In the
corner you will see a status info indicating whether the robot is
in motion or stationary. See the ICRA 2005
paper for details.
- background/foreground scene segmentation
This video accompanies the CVPR 2005
paper , where we describe a foreground/background classification of
pixels in video. The classification follows from a standard assumption
that the camera and the background scene are static, whereas the
interesting foreground object are in motion. Our method is one of the
few approaches, where the spatial coherence of segments is enforced by
an appropriate prior distribution on segments. This distribution
prefers foreground segments that have smooth border.
- more to come ... !