Problem Solving by Teams of Heterogeneous Agents
Gregory D. Hager
Dept. of Computer Science
Yale University
2:30 pm Mon. Nov. 13 in 2310 CS&S
One of the common motivations for studying vision is to provide
sensory feedback systems that are able to move about and manipulate
their world. However, to date most "successful" systems have been
limited in scope, brittle to the point of being contrived, and
daunting in their hardware complexity. These limitations can usually
be traced to two implicit design principles: first, that vision is a
means for measuring geometry; and second, that general-purpose,
real-time vision computation must be performed on specialized
hardware.
In this talk, I will argue that by avoiding both of these
assumptions, it is possible to build software systems that provide
real-time vision processing and high positioning accuracy on common
desktop hardware and inexpensive mechanical components. In both
vision and control, the key idea is to design feedback mechanisms
which exploit any available geometric and/or photometric invariants of
the problem. As an illustration, I will describe an approach to
hand-eye coordination which leads to provably convergent positioning
without relying on accurate calibration. Given time and interest, I
will also present a second example in the domain of visual tracking
under changing illumination.
In order to facilitate the construction of vision-based tracking
and control systems, we have constructed XVision, a software toolkit
for performing visual tracking on common PC's and workstations. In
addition to illustrating XVision in the context of the above examples,
I will argue that our approach to vision is likely to lead to a
variety of novel and interesting (and fun!) vision applications.