|
We
will begin with a short introduction on the motivation of our
work (combining acoustical and visual cues for perception), and
an outline of ongoing work in our lab in this area. Next we will
present a new source localization algorithm that we have developed
in the context of a prototype audio-video teleconferencing system.
Source
localization using the difference in time of arrival of a signal
(TDOA) at an array of receivers is used in many fields including
speaker location using acoustical arrays. Using three TDOA values
from four non-collinear receivers one can, in principle, solve
for the unknown source coordinates in terms of the receiver locations.
However, the equations are non-linear, and in practice, signals
are contaminated by noise. Practical systems often use multiple
receivers for accuracy and robustness, and improved S/N, and solutions
must be obtained via nonlinear minimization.
We
have developed a new family of exact solutions, for source localization
for the case of four receivers located in a plane. These solutions
can be evaluated using a small number of arithmetic operations.
These
solutions, however, are sensitive to noise in the delays measured,
and multiple microphones (>4) can be used to improve the source
location estimate. We have developed a new formulation for the
source localization problem for multiple microphones. Our approach
solves the source localization problem in stages, with appropriate
optimization algorithms used for each stage, and incorporates
the developed exact solutions. The present estimator is compared
with ones from the literature, and found to be robust and accurate,
and more efficient.
(work
with Dmitry Zotkin and Larry S. Davis.)
|
