See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4256278

A New Statistical Measure of Signal Similarity

Conference Paper · March 2007

CITATIONS READS

15 18,949

1 author:

Hugh Lachlan Kennedy

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Digital Control Systems View project

High-Temperature Superconductors View project

All content following this page was uploaded by Hugh Lachlan Kennedy on 09 March 2015.

The user has requested enhancement of the downloaded file.

Abstract

A new measure of similarity, suitable for both signal processing and image processing applications, is presented. The measure
of similarity ( Z M ) is an F-distributed test statistic that quantifies the degree of alignment, correlation or coincidence between
two or more (time- or space-dependent) waveforms. The test statistic may be used to automate detection, classification,
localisation, association and registration operations. When used to estimate time-of- arrival differences, which is typically done
using cross correlation or beam formation, the technique provides an efficient means of: maintaining a low and constant false-
alarm rate, detecting weak signals in noise and reducing angular errors. The method is applied to the problem of acoustic
source detection and localisation. Real data are used to compare the method with cross correlation and normalised cross
correlation. The new method yields prominent and narrow peaks, at the true source locations, with false peaks due to
misalignment, misassociation, and background noise, suppressed. When processing continuous data streams, the test statistic is
computed efficiently in the time domain.

1. INTRODUCTION

Measures of similarity are required in a wide range of radar sonar, communications, remote sensing, artificial intelligence and
medical applications, where one signal or image is compared with another. Many basic processing operations, such as matched
filtering, cross correlation, and beam formation, may be interpreted as being based on measures of similarity. These related
operations typically form the foundation of the detection, classification, localisation, association and registration algorithms
employed in semi-autonomous sensor systems. Matched filtering involves the projection of an unknown waveform onto a
known template; cross correlation is used to determine the translation (in time or space) required to bring one waveform (signal
or image) into alignment with another; while beam formation brings multiple signals into phase so that they constructively
interfere. A new measure of similarity [1] is derived in this paper and applied to the problem of acoustic source localisation.
Beam-formation [2] and cross-correlation processing techniques [3] are used to compute Time-Of-Arrival-Differences
(TOADs) or Time Delay Estimates (TDE) in distributed networks of acoustic sensors [4&5]. Networks may be composed of
single-sensor nodes, where low-resolution digitised data are shared over the network [6], or of multi-sensor nodes, where high-
resolution digitised data are processed at each node, with only direction vectors [7] (and possibly range [8]) shared over the
network. Both processing techniques and both network topologies employ detection and localisation processing stages. Cross
correlation and beam formation processing are typically implemented using detect-then-localise and localise-then-detect
architectures, respectively. Detection prior to localisation is efficient because a detection filter is applied to only one or more
channels of digitised sensor data and the localiser is only applied to short data segments containing transient features of interest,
as identified by the detector. Localisation prior to detection, using time-domain delay-and-sum beam-formation, is
computationally more demanding, because the detector is applied to the outputs of a large parallel bank of localising filters,
however, this architecture allows both continuous and transient signals to be processed.
Localisation by beam formation can be applied to any number of sensors, while localisation by cross correlation is restricted
to sensor pairs, with the pairwise results combined to form a consolidated angular estimate during a subsequent (fusion)
processing stage. Both cross-correlation and beam-formation processes output a localisation function, computed in either the
time or frequency domain. Maxima (peaks) in this function are used to identify the most likely lag (in the case of cross
correlation) or beam (in the case of beam formation). The wrong lag or beam may be selected in the presence of: reverberation,
multi-path propagation, multiple sources, and background noise. Angular errors due to the selection of the wrong peak are much
greater than the errors in the location of the true peak. Unfortunately, when sampling rates are low, and when sound sources are
far from broadside, even small TOAD errors result in large Angle-Of-Arrival (AOA) errors. A variety of alternatives to cross
correlation, such as normalised cross correlation, normalised covariance and sum-of-squared differences have been proposed
and compared in the literature [9], in an effort to mitigate the problems described above. These two-channel methods will be
collectively referred to here as correlation, or lag-estimation techniques.
As distributed arrays become larger, relative to the frequencies and the pulse durations of the sounds that they are processing,
the issue of TOAD ambiguities becomes more significant. Reducing the quality of the sensor data (sampling rates, number of
channels and quantization bits), as required in capacity constrained and disadvantaged networks, only makes matters worse. To
meet these challenges, the new measure of similarity described in this paper is used as the basis of a robust and efficient TOAD
estimation algorithm.
 The TOAD estimation algorithm operates on two or more channels of sensor data and outputs an F-distributed test statistic
that is ideal for joint localisation and detection. It has the potential to enhance the performance of distributed acoustic networks,
when processing both impulsive transients and continuous waveforms. Computer simulation and extensive live testing using
gun-shot data (processed in real time) by PNV have demonstrated improvements in the following areas relative to conventional
cross correlation and beam formation techniques: increased angular resolution, reduced (and theoretically constant) false-alarm
rates in the presence of background noise, and an increased probability of detecting weak signals. The test statistic may also be
used as a measure of localisation quality (or reliability) as misaligned or misassociated waveforms, in the presence of signal
distortion or background interference, are likely to yield a low measure of similarity. Filtering out AOA estimates of low quality
reduces the average angular error and the false detection rate of the system, however, the probability of detection is also
reduced. A measure of AOA reliability has also been derived in [10]; however, it is applied at the array level where multiple
pairwise TOADs are combined to form a consolidated AOA vector, not at the signal level, during TOAD estimation.

2. DERIVATION

A statistical treatment of a delay-and-sum beam-former will now be described and used to derive the new measure of signal
similarity. The derivation is based on a few standard statistical relationships [11]. A hypothesis test is performed with the null
hypothesis being that there is no signal present and that the waveforms entering the beam former contain only zero-mean
Gaussian-distributed noise. It is assumed that any Direct Current (DC) offset in the data (e.g. sensor bias) or frequencies that are
of no interest (e.g. wind or self noise) have been removed by a pre-whitening stage. If the null hypothesis is rejected then it is
assumed that a localisable signal is present. The test statistic for all possible lag combinations corresponding to all physically
measurable angles is computed. The most likely direction of the source is set equal to the angular coordinate for which the null
hypothesis is least likely, i.e. the test statistic is maximised.
The delay-and-sum beam-former is applied as
M 1
y n  x
m 0
m n , (1)

where xm  n  is the nth sample output from the mth delay channel and y  n  is the beam-formed output. In Eq. (1) it is
assumed that the appropriate delays have been applied to steer a beam in a desired direction. The noise statistics of every
sample from all sensors are assumed to be identical, so the nth sample in each delay channel is assumed to be an independent
observation of the random variable X n . Analysing the digitised waveforms (in x) over a window of length N, gives a total of N
different random variables, with M observations of each variable. Under the null hypothesis the variables have a Gaussian
(Normal) distribution


X n ~ N n ,  n2 .  (2)

At a given n, using the data from all M channels, the Maximum Likelihood Estimates (MLEs) ˆ n and ˆ n2 , of the (true) mean
and variance  n and  n2 , are computed using

y n
ˆ n  (3)
M
and


M 1
y  n 
2


1
m 
2
ˆ n2   x n  . (4)
M  M 
 m  0 
Under the null hypothesis the following relationships hold:

 ˆ n  n 2
If Z a  M then Z a ~  2 1 ; (5)
 n2

ˆ n2
If Zb  M then Zb ~  2  M  1 . (6)
n 2
Under the null hypothesis it is also assumed that the noise statistics of the sensor outputs are zero mean and time invariant so the
parameters of each distribution are the same:
1  2 ,...,  N    0
(7)
and

12   22 ,...,   N2   2 . (8)

Using the reproductive property of  2 variables, the following aggregate test statistics can be formed and analysed:
N 1

 ˆ
M
If Z c  2
n then Zc ~  2  N  ; (9)
2 n 0

N 1

ˆ then Z d ~  2  N  M  1  .
M
If Z d  2
n (10)
2 n 0

So far it has been assumed that the true variance (  2 ) of the (white) noise is known. This is an inconvenient and unnecessary
assumption. It can be eliminated by dividing (9) by (10); furthermore, if the numerator and the denominator are scaled by the
inverse of their respective degrees of freedom, i.e.
Zc N
ZM  , (11)
Zd  N  M  1 
then a variable distributed according to Snedecor’s F distribution results; that is, after substituting (9) and (10) into (11):
N 1

 ˆ 2
n
Z M   M  1 Nn 01 , (12)
 n 0
ˆ n2

with

Z M ~ F  N , N  M  1  . (13)

Substituting the expressions for ˆ n and ˆ n2 , given in (3) and (4) respectively, into (12) yields
N 1

 y  n
1 2
M
Z M   M  1 M 1 N 1 n 0
N 1
. (14)
 x  y n
1
m n 2

2

m 0 n 0
M n 0

3. IMPLEMENTATION

Note that the above expression (14), contains three sum-of-squares terms that are equivalent to power measurements computed
in the time domain. Alternatively, (14) may be written in terms of moments:
N 1

 E  x  n 
2

Z M   M  1 N 1 n 0
N 1
(15)
 E  x  n     E  x  n 
2 2
 
n 0 n 0

using
 M 1

x
1
E  x  n    m n , (16)
M m 0

M 1

x
1
E  x  n     n 2
2
(17)
  M m
m 0

and
M 1
y n  x
m 0
m n . (18)

The two sum-of-squares terms (16&17) are computed efficiently as moving averages, using sliding windows, implemented
using circular buffers. In this way, the Z M statistic, for each beam, is calculated using only the samples that enter and exit two
sliding windows, each of length N; as a result, the computational effort does not depend on N.

4. INTERPRETATION

The Z M test statistic is the ratio of two sum-of-squares quantities (12). If the square of the estimated mean (numerator) is
regarded as the (delay-and-sum) signal power, and the variance (denominator) the noise power, then it may be convenient to
convert Z M into a Signal-to-Noise Ratio (SNR) in dB. Images may then be formed using many closely-spaced beams, and
presented to an operator for visual inspection.
Knowing that under the null hypothesis the Z M statistic is F distributed, allows a detection threshold to be computed to give
the desired probability of falsely rejecting the null hypothesis when it is indeed true (a false alarm). If the computed Z M value
exceeds the threshold then a localisable signal is instead assumed to be present. The necessary threshold is determined using the
inverse Cumulative Density Function (CDF) of the F distribution. The two parameters (degrees of freedom) of the function
automatically adjust the threshold (increase it) to compensate for the higher variability of the test statistic when low channel
counts (M) are used and when the data window length (N) is small.
In practice, the null hypothesis is rarely entirely true, and false alarms due to nuisance sources are common, so a larger
detection threshold is usually appropriate, giving a negligible theoretical false-alarm probability (the size of the test), an
acceptable practical false-alarm probability and a reasonable probability of detection (the power of the test).
This method is able to suppress peaks on background noise and produce a single sharp peak on localisable signals because
the Z M statistic for poorly aligned signals is low. The null hypothesis is rejected when the estimated means at each sample in
the analysis window ( ˆ n ) are far from zero and when the dispersion around each of the means is small (as indicated by small
values of ˆ n2 ). Applying the appropriate delays to steer the beam in the direction of the source results in large ˆ n values and
small ˆ n2 values because the waveforms in each delay channel are nearly identical (except for disturbances caused by additive
white noise). In the presence of non-localisable signals (e.g. flow noise or other diffuse sound sources), or when the incorrect
delays are applied to steer the beam in the wrong direction, the ˆ n values remain large but so are the ˆ n2 values (due to
waveform misalignment), which ensures that Z M remains small.
The Z M method is able to detect and localise weak signals in noise because a small variance term in the denominator of (14)
yields a large Z M test statistic, even when the power in the numerator is small.
 5. APPLICATIONS

As this technique outputs large values when the delayed waveforms are large in magnitude and aligned with each other (in time
or space), it may be used in any application where measures of similarity are required. One of the advantages of this technique
relative to other measures of similarity (like cross-correlation) is that it can be applied to an arbitrary number of channels.
When the Z M test statistic is applied to two data segments (M=2) of finite length, extracted from different channels by a
detector it may be used to estimate TOADs, as an alternative to cross correlation. When applied to the problem of lag
estimation, using two finite data segments of equal length, one data segment is shifted relative to the other and the summations
are performed only over the overlapping portion (of length N) of the data segments. A lag-dependent threshold (as determined
by the inverse of the F-distribution CDF) provides an alternative way of managing the problem of high variance at large lags,
due to reduced overlap, without biasing results (towards zero) with a triangular window function [12]. The lag-dependent
detection thresholds are computed once only on system initialization. In this role, the Z M method may be used to correlate
transmit and receive signals for range estimation in active sensor systems, or to correlate multiple pairs of received signals for
AOA estimation in passive systems.
The method may also be used to compare an unknown transient waveform (in a continuous data stream or a finite data
segment) with a set of pre-specified signal templates (each a finite data segment), for the purpose of joint detection and
classification. In this case, the channel count is also equal to two (M=2). One of the channels contains the unknown waveform,
the other contains the template used to represent a given class. The Z M test statistic is computed for all possible classes. A
detection is declared when any of the test statistics exceeds a specified threshold. The class, for which the test statistic is
maximised, is assigned to the detection.
The above operations of correlation, detection and classification are also required when analysing imagery data. When the
Z M test statistic is applied to frames of imagery, the summations over time are replaced by summations over all (spatial and
spectral) dimensions of the image space, within the region of image overlap. Two or more images may be aligned (correlated or
registered) by finding the set of spatial transforms for which the test statistic is maximised.

6. ILLUSTRATION

Using real acoustic data, a single two-channel lag-estimation problem will be used to illustrate the operation, and demonstrate
the performance, of the new measure of similarity described in this paper ( Z M ), relative to cross-correlation ( R12 ) and
normalised cross-correlation ( 12 ). A pair of ¼” pre-polarised free-field array-microphones (± 1 dB from 100 to 10000 Hz)
were used, each simultaneously digitised at a rate of 50,000 samples per second per channel ( Fs ), with 24bit quantization,
using a data acquisition device (DAQ) with in-built anti-aliasing filters. The microphones, separated by a distance of 0.4m ( D ),
were placed in the corner of a 3m x 5m room, 1m from each wall. A speed of sound of 335m/s ( vc ) was assumed. The
microphones were exposed to a transient signal, generated using a mechanical clicker (approx. pulse frequency = 2000Hz-
2500Hz, approx. pulse duration = 7ms). The emitter was placed 2.5m from the inter-sensor mid-point, at 90° to the inter-sensor
axis (broadside, where l = 0). Wave-front curvature for this sensor-source geometry is not insignificant, it was however ignored
in this study. The digitised sensor signals were pre-filtered using a DC notch filter (-2dB at 25Hz) implemented in software.
Upon detection of the signal, data segments from both channels were extracted and processed. The segment from channel 1 was
correlated with the segment from channel 2 at all physically feasible lags ( L  l   L ), where L  60 (see Eq. 18). The
segments were extracted so that the wave-front arrival point, as determined automatically by the detector, was at the centre of
both segments. The segment from the first channel was of length N; the segment from the second channel was extended
(padded) by L samples at either end, to ensure that the analysis window length (N) is constant for all lags (for uniform variance),
giving a total segment length of N+2L. Two analysis window lengths were used to process the same detection ( N  201 and
N  101 ) and their results compared. The results are plotted in Figs. 1 to 4.
 Fig. 1: Localisation functions (with N=201).

Fig. 2: Digitised waveforms over the analysis window, aligned using the most likely lag (as determined using ZM with N=201).

Fig. 3: Localisation functions (with N=101).

Fig. 4: Digitised waveforms over the analysis window, aligned using the most likely lag (as determined using ZM with N=101).

When N  201 (see Fig. 1), the maxima in the Z M and 12 localisation functions are at the correct lag ( l  0 ). After
quadratic interpolation around the Z M maximum, a non-integer lag of 0.5 is obtained, yielding an AOA (  ) of 89.5°, using (for
far-field sources):
 l  F D
  cos 1   with L  ceil  s  . (18)
L  vc 
In the R12 localisation function, one of the peaks at an incorrect lag ( l  22 ) is slightly larger than the peak at the correct lag
(see Fig. 1). Use of this incorrect (integer) lag yields an AOA of 111.5°.
Reducing the length of the analysis window from N  201 to N  101 , excludes the newer data from the computations,
where the waveforms begin to lose coherence (compare Fig. 2 and Fig. 4). All methods now yield maxima at the correct lags
(see Fig. 3). The value of the Z M localisation function at the true lag ( l  0 ) is now much greater than values at the adjacent lags
( l  1 ), and much greater than maxima of adjacent peaks at the wrong lags ( l  22 ). The interpolated lag value of 0.1 yields
an improved AOA of 89.9°.
In general, the ambiguous peaks in the R12 localisation function tend to be similar in height; as a result, random signal
perturbations may cause the height of a false peak to be greater than that of the true peak. In comparison to the 12 and R12
methods, the Z M method yields true peaks that are tall and narrow, with flattened false peaks. Further investigations (not
shown here) with different sound sources, inter-sensor separations, angles of arrival, and noise environments, have revealed that
the 12 and the Z M methods are much less likely to fail (i.e. to identify the incorrect lag) than the R12 method; however, they are
not perfect and do fail in some cases. Even though the 12 false peaks are still tall, relative to the true peak, it was noted that the
12 and Z M methods tend to succeed and fail in the same cases; therefore, the subtle suppression of the 12 false peaks is
sufficient in most cases. The more obvious suppression of the Z M false peaks is still a desirable characteristic when the
localisation function is converted into an image because the image is sharper and visually less ambiguous.

7. DISCUSSION

According to Eq. (18), the theoretical angular resolution is increased by increasing the inter-sensor separation, by increasing the
sampling rate, or both (assuming the speed of sound is constant). However, when the source frequency is above the array
frequency, as is the case in the above example, the cyclic ambiguity gives rise to multiple peaks during TOAD estimation,
which may result in misalignment errors, yielding very large AOA errors. By reducing the incidence of misalignment errors
(spatial aliasing), the Z M measure of similarity allows the use of large arrays, with greater angular resolution, to be used to
localise transient sounds, over a greater range of frequencies, with a greater degree of accuracy.
In the preceding example, the Z M measure of similarity resolves the cyclic ambiguity through the effective utilization of
envelope information. If the detector is late in determining the onset of the signal, so that the rise of the envelope is not captured
within the analysis window, preferably near its centre, then the height difference between the true peak and the ambiguous
peaks, in the localisation function, is reduced. The performance of the detector and the length of the analysis window (relative
to the signal wavelength, the duration of the modulating envelope, and the extent of signal coherence) are critical design
parameters of all correlation methods, when processing transients using a detect-then-localise architecture.
The Z M measure of similarity reduces the incidence of misalignment errors when processing data collected from a compact
array. It is also able to reduce the incidence of misassociation errors when processing data collected from widely dispersed
arrays, where the sensors may be separated by hundreds of meters. Association is used here to describe the process whereby all
signals, due to the same source, measured using different sensors, are identified. The Z M test statistic will be small, if a signal
from sensor 1, due to source A, is correlated with a signal from sensor 2, due to source B, if the waveforms of A and B are
dissimilar.
A disadvantage of the Z M measure of similarity is that it requires the signal power levels in each channel to be
approximately equal. Signal power levels are determined by the received Sound Pressure Level (SPL), the sensor/preamplifier
sensitivity, and the signal conditioner gain. Even if the digitised sensor signals due to a common source have been correctly
aligned (in time), a high variance (in amplitude) will result if the power of the signals in each channel are different. This effect
is minimised by sensor calibration or pre-normalization of the data, in each channel, over the analysis window. Normalised
cross correlation is unaffected by signal power level differences, however, like the Z M method, it is sensitive to DC offsets.
This shortfall motivated the development of the normalised covariance method.
The sensitivity of the Z M method is an advantage in most cases; however, detections on weak but well correlated signals due
to nuisance or low-interest sources (acoustic clutter) or internal noise (e.g. electromagnetic interference) may also be generated.
Normalised cross correlation is similarly affected, while cross correlation is relatively unaffected. In practice, when a localise-
then-detect architecture is adopted, and the Z M statistic computed, a constant minimum variance term is added to the
denominator in Eq. (14), to reduce the sensitivity and compensate for the finite measurement resolution of the DAQ.
 One of the undesirable characteristics of cross correlation is that detection thresholds must be arbitrarily set. This is not a
problem if each data segment is known to contain one localisable signal; in this case, the lag of the maximum in the cross
correlation vector is selected, regardless of its level. Normalised (covariance and cross correlation) methods are more
favourable in this respect because perfectly aligned waveforms have a coefficient of unity. On this basis, an appropriate
detection threshold (0.90 or 0.99, for example) may be chosen. When the Z M method is used, the detection threshold is chosen
to give the desired theoretical false-alarm rate, when the null hypothesis is assumed to be true, using the inverse of the F-
distribution CDF.

8. CONCLUSION

The method described in this paper has the potential to improve a wide range of signal processing algorithms where a measure
of signal similarity is required. The method is particularly well-suited to the problem of source localisation, where it has been
shown to improve angular accuracy, and reduce false-alarm rates. The method was used in this paper to correlate two channels
of acoustic data containing transient signals; however, it may also be used to estimate TOADs in more than two channels.

ACKNOWLEDGEMENT

This work was (in part) conducted under the sponsorship of a Capability and Technology Demonstrator (CTD) program (Land
148), provided by the Commonwealth of Australia (CoA) and administered by the Land Engineering Agency (LEA). Initial
technical guidance was provided by the Defence Science and Technology Organization (DSTO), Maritime Operations Division.

REFERENCES

[1] Patent Pending, Provisional Application No. 2006901479, “Signal Analysis Methods”, Mar. 2006.
[2] B.D. Van Veen and K.M. Buckley, “Beamforming: a versatile approach to spatial filtering,” IEEE ASSP Mag., vol. 5, pp.
4-24, Apr. 1988.
[3] C.H. Knapp and G.C. Carter, "The Generalized Correlation Method for Estimation of Time Delay," IEEE Trans. Acoust.,
Speech, and Sig. Proc., vol. 24, no. 4, pp. 320-327, Aug. 1976.
[4] J.C. Chen, Kung Yao and R.E. Hudson, “Source localization and beamforming”, Sig. Proc. Mag. IEEE, vol. 19, no. 2, pp.
30–39, Mar. 2002.
[5] Qing Zhao, A. Swami, Lang Tong, “The interplay between signal processing and networking in sensor networks”, Sig.
Proc. Mag. IEEE, vol. 23, no. 4, pp. 84–93, Jul. 2006.
[6] P. W. Boettcher and G. A. Shaw, “A Distributed Time-Difference of Arrival Algorithm for Acoustic Bearing Estimation”,
in Proc. 2001 Int. Conf. on Information Fusion, vol. 1, Montreal, Aug. 2001.
[7] B.G. Ferguson, L.G. Criswick and K.W. Lo, “Locating far-field impulsive sound sources in air by triangulation”, J. Acoust.
Soc. Am., vol. 111, no. 1, pt. 1, pp. 104-16, Jan 2002.
[8] B.G. Ferguson and K.W. Lo, “Passive ranging errors due to multipath distortion of deterministic transient signals with
application to the localization of small arms fire”, J. Acoust. Soc. Am., vol. 111, no. 1, pp. 117-128, Jan. 2002.
[9] F. Viola and W.F. Walker, “A comparison of the performance of time-delay estimators in medical ultrasound”, Ultra.,
Ferro. Freq. Contr., IEEE Trans. on, vol 50, no. 4, pp. 392–401, Apr. 2003.
[10] T. Pirinen, P. Pertila, A. Visa, “Toward intelligent sensors - reliability for time delay based direction of arrival estimates”,
in Acoustics, Speech, and Signal Processing, 2003. Proceedings. (ICASSP '03). 2003 IEEE International Conference on,
2003, vol. 5, pp. 197-200.
[11] L.J. Bain and M. Engelhardt, Introduction to Probability and Mathematical Statistics, 2nd Ed., California, Duxbury Press,
1992.
[12] J.G. Proakis and D.G. Manolakis, Digital Signal Processing. Principles, Algorithms and Applications, 3rd. Ed. New Jersey,
Prentice-Hall International, 1996.

View publication stats