|
Since 1st September 2006, I am postdoctoral scientific assistant (Akademische
Rätin) in the
Intelligent Vision Systems Group
in the
Institut of Computer Science III at the
university of Bonn.
Research interests
Visual Attention, Cognitive Computer Vision, Artificial Intelligence, Robot
Vision, Robot Localization, Visual SLAM, Omnidirectional Sensing
My main research interest is in computational visual attention, that
means in the detection of salient regions in images. The concept of visual
attention is adapted from the human visual system: humans look automatically at
salient regions that "pop out" of an image.
To illustrate what I mean by "pop out", have a look at the following image which was kindly
provided by Christoph Dose (© 2005
Christoph Dose):
I guess, you looked immediately at the red sheep. This should even work if you
are color blind, since there is not only a color but also
an intensity contrast. So, the
strong contrast to the environment is one aspect that attracts our
attention. The other important aspect is the uniqueness: there is only one red
sheep in the image, it differs considerably from all the other sheep. That
means, everything that is different attracts our gaze.
During my phd-thesis at the Fraunhofer Institute AIS, I investigated this
topic in detail and
developed a computer system VOCUS that finds these
salient regions automatically.
It creates a "saliency map", a map in which salient regions are displayed bright
and other regions dark. Have a look how such a saliency map looks like for
the previous sheep image:
As you can see, the region of the red sheep "pops out" in the saliency map.
VOCUS is based on psychological models as the
Feature Integration Theory by Anne Treisman and on computational models like
the Neuromorphic Vision Toolkit by Laurent Itti and colleagues.
One special aspect about VOCUS is that it is able to perform "Goal-directed
Search", that means it can search for a target object in a scene. For example,
if you provide a training picture of an object to VOCUS, like a picture of a fire
extinguisher, it can search for this object in test images.
The system is real-time capable (50ms for a 400 x 300 pixel image on a 2,8GHz
PC) and thus useful for robot applications.
More information on VOCUS and my thesis can be found here.
After my phd, I spent one year in Stockholm in the group of Henrik Christensen
at KTH (the Royal Institute of
Technology, or in Swedish: Kungliga Tekniska Högskolan).
I was employed within the EU project NEUROBOTICS and investigated
the use of VOCUS for landmark detection
and gaze strategies in robot navigation.
Together with Patric Jensfelt, I developed a visual SLAM system (simultaneous
localization and mapping) in which a robot had to build a map of the
environment based on camera data. The selection of the visual landmarks was done by VOCUS. Here, you can see me with the robot Dumbo in the
"livingroom" at KTH:
Have a look at my Publications to see the results of our work.
Currently, I am working on visual matching strategies, that means I try to find
correspondences in images. This is especially important in loop closing
situations in visual SLAM when image regions in a current frame have to be
matched to regions from a database.
Academic record
- Senior Scientific Assistant (Akademische Rätin) at the
Institute of Computer Science III,
University of Bonn, Germany, since June 2007.
- Scientific Assistant (Wissenschaftliche Mitarbeiterin)
at the
Institute of Computer Science III,
University of Bonn, Germany, Sept. 2006 - June 2007.
- Post-doc at the
computer vision and active perception lab (CVAP) at the
school of computer science and communications (CSC) at
KTH, the Royal Institue of Technology, Stockholm,
Sweden, Aug. 2005 - July 2006.
- Dissertation at the
University of Bonn, July 2005
- PhD student at
Fraunhofer Institute AiS, February 2002 - June 2005.
- Diploma in Computer Science,
University of Bonn, 2001.
Former Research projects
- NEUROBOTICS (EU Project FP6-IST-001917): The fusion of Neuroscience and Robotics
- PhD thesis: VOCUS: A Visual Attention System for Object Detection and Goal-directed search
- MACS (EU Project FP6-004381 (STREP)): Multi-sensory Autonomous
Cognitive Systems Interacting with Dynamic Environments for Perceiving
and Using Affordances
-
Robust Robot Localization with Omnidirectional Vision (former
project, diploma thesis)
Publications
see here
|
