|
This research attempts to improve on the techniques
available to estimate the 3-dimensional direction of
arrival of a far-field acoustic source using an array
of microphones, in the presence of room-reverberation.
The idea is to perform the estimation reliably using
a small frame of recorded data, say 256 samples from
each microphone. The DOA is estimated by first computing
the time delays of arrival of signals between each pair
of mcirophones. These time-delays are then converted
to a direction of arrival estimated from the known geometry
of the array. The reliability of the time delay estimates
is reduced because of multiple reflections of the signal
from the walls of the room, a phenomenon known as reverberation.
The generalized cross-correlation with the phase transform
is used to improve the reliability of the time-delay
estimates by making the peaks of the cross-correlation
sharper, thus separating the true delay from those due
to reflections.
Using the phase transform gives rise to multiple peaks,
and thus multiple time-delay estimates for each pair.
Our algorithm, named Time Delay Selection (TIDES), attempts
to select the correct set of time delays over all pairs
of microphones. This is done by first estimating all
candidate time delays for eah pair and then constructing
all possible sets of time-delays over all pairs. We
developed two sets of criteria for selecting an appropriate
time delay set. One criterion selects that set that
has the minimum weighted least squares error associated
with it. The second criterion selects that set that
has minimum weighted distance from a previously selected
set. The previously selected set that was used here
was a median filtered set from the previous three frames.
In either case the weighting was done so that the sets
corresponding to stronger peaks had a greater chance
of being selected. Both these methods improved the reliability
of the direction estimates. This was shown using both
simulations and actual recorded data using recording
equipment built in our lab.
|
