|
|
Foundations of Robotics
Seminar, November 15, 2006
Time
and Place | Seminar Abstract | Speaker
Biography | Presentation Slides | Speaker
Appointments
Structure-from-motion
by convolution
Christopher Geyer
Smith Hall 100
Refreshments 4:45 pm
Talk 5:00 pm
In this talk I propose an efficient method
to compute the globally optimal hypothesis for the motion between two cameras
in the presence of extremely large numbers of outliers or ambiguities in
motion. By combining notions from multiple-view geometry and non-commutative
harmonic analysis, we can efficiently compute a Radon transform using a
convolution in the five-dimensional space of relative motions between two
cameras. This approach is analogous to the Hough transform and gives
robustness to large numbers of outliers, even all possible point pairs between
two views, and because it is non-parametric it yields useful marginals even in
the presence of ambiguous motions. We also discuss how this fits into a
number of applications of non-commutative harmonic analysis to robotics, often
used because of the ability to represent and efficiently compute non-parametric
probability distributions. This is joint work with Ameesh Makadia and
Kostas Daniilidis.
|
Speaker Bio |
Christopher Geyer is a Project Scientist in
the Field Robotics Center. He received his PhD from the GRASP Lab at the
University of Pennsylvania and spent time at U.C. Berkeley as a post-doc, among
other things working on computer vision for autonomous helicopters.
Recently he has been involved in projects in surveillance from UAVs, tracking
for BigDog, and collision avoidance of general aviation aircraft for UAVs.
|
Speaker Appointments |
For appointments, please contact Christopher
Geyer (cgeyer@cs.cmu.edu).
The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.