|
Time delay estimation (TDE)-based algorithms for estimation
of direction of arrival (DOA) have been most popular
for use with speech signals. This is due to their simplicity
and low computational requirements. Though other algorithms,
like the steered response power with phase transform
(SRP-PHAT), are available that perform better than TDE
based algorithms, the huge computational load required
for this algorithm makes it unsuitable for applications
that require fast refresh rates using short frames.
In addition, the estimation errors that do occur with
SRP-PHAT tend to be large. This kind of performance
is unsuitable for an application such as video camera
steering, which is much less tolerant to large errors
than it is to small errors.
We propose an improved TDE-based DOA estimation algorithm
called time delay selection (TIDES) based on either
minimizing the weighted least squares error (MWLSE)
or minimizing the time delay separation (MWTDS). In
the TIDES algorithm, we consider not only the maximum
likelihood (ML) TDEs for each pair of microphones, but
also other secondary delays corresponding to smaller
peaks in the generalized cross-correlation (GCC). From
these multiple candidate delays for each microphone
pair, we form all possible combinations of time delay
sets. From among these we pick one set based on one
of the two criteria mentioned above and perform least
squares DOA estimation using the selected set of time
delays. The MWLSE criterion selects that set of time
delays that minimizes the least squares error. The MWTDS
criterion selects that set of time delays that has minimum
distance from a statistically averaged set of time delays
from previously selected time delays.
Both TIDES algorithms are shown to out-perform the
ML-TDE algorithm in moderate signal to reverberation
ratios. In fact, TIDES-MWTDS gives fewer large errors
than even the SRP-PHAT algorithm, which makes it very
suitable for video camera steering applications. Under
small signal to reverberation ratio environments, TIDES-MWTDS
breaks down, but TIDES-MWLSE is still shown to out-perform
the algorithm based on ML-TDE.
|
