Direction of Arrival Estimation Algorithms

Nikhil Shetty

April 7, 2004
Contents

1 Software report 2
1.1 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 User’s Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Design features and observed results . . . . . . . . . . . . . . 4
1.5 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.6 Code Listing . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1
Chapter 1

Software report

1.1 The Problem

The aim of the software is to implement the MUSIC algorithm that enables
an antenna array to estimate the number of incident signals on the array
and their directions of arrival.
Let there be m elements in the array. Let xi be the input at the ith element
of the array (including noise added
 by the instruments and present in in-
coming signal) and let x = xi . According to the MUSIC algorithm , the
covariance matrix formed as S = xx∗ (where x∗ denotes conjugate trans-
pose) has n repeated minimum eigenvalues ( for uncorrelated noise ) that
represent the variance of the noise. The number of incident signals is given
by d = m − n. Further , the eigenvectors corresponding to these eigenvalues
are orthogonal to the array manifold ( the gain and phase change provided
by the array elements to the incoming signal directions ) at the values of
the incoming signal directions. Hence , on forming a matrix EN of these
eigenvectors and plotting
1
P M U (θ) = ∗ (1.1)
a(θ)∗ E N EN a(θ)

we get peaks at the values of theta corresponding to the incoming signal

directions.

1.2 User’s Guide

PART I: This part is a sample calling program designed to validate the
working of the music algorithm. A real life case will be considered

2
 later. The sample program is simulating a sine wave input that is
arriving at the array in the form of planar wavefronts ( the source
is assumed sufficiently far away ). The array itself is assumed to be
an m ( N OOF ELEM EN T ) element linear array with no directional
properties ( could be an array of dipoles ) in the given plane. The
number of elements can be varied. Further the frequency of the input
can be changed. df actor is the division factor of the wavelength that
represents the distance between the elements.(for eg: if df actor = 2
,distance between elements = λ2 ). Thus we have a handle on both
the frequency and the distance between elements in terms of the given
frequency.
The samples variable represents the number of samples that we wish
to take for the given input and f s represents the sampling frequency.
These two variables can be adjusted to simulate the processor available
, the amount of accuracy and the system refresh rate. For example ,
if we chose a high value of samples , amount of memory utilisation
increases and the refresh rate ( rate between consequent measurements
) decreases however the accuracy of calculations increases ( assuming
input signal lasts for that long a time ).
x is the matrix in which the signal recieved at the aray elements is
stored.
 
x1 (t1 ) x1 (t2 ) · · · x1 (tsamples )
 x2 (t1 ) x2 (t2 ) · · · x2 (tsamples ) 
 
 .. .. .. .. 
. . . .
 
 
xm (t1 ) xm (t2 ) · · · xm (tsamples )
is the format in which the array is stored where xi (tj ) represents the
value of the input function detected at the ith element at the jth time
instant(including noise). Thus f 1 is the required input function and
r is the noise matrix. The noise is zero mean with variable variance
that can be changed by changing the multiplier.
Once the array manifold function is defined in a different file (refer
PART II) all the user has to do is call the music1d function to get the
estimate of the number of signals and a plot of P M U (θ) versun θ.

PART II: The array manifold is defined as the gain and phase change pro-
vided by the array elements to a given signal direction. The array
manifold is a function of θ and is saved in the file manifold file. Note
that the variables N OOF ELEM EN T S and df actor are global and
hence can be accessed by the manifold too.

3
PART III: This part consists of the actual implementation of MUSIC al-
gorithm.The function returns an estimate of the dir3ection of arrival.
It takes as input the x matrix. From the x matrix , the number of
elements and the samples is obvious. The first step is the calculation
of the covariance matrix. The next step involves calculation of the
eigenvalues and eigenvectors of S.
On knowing the minimum eigenvalue, the multiplicity of this eigen-
value is obtained. The multiplicity is presently being obtained using a
naive comparison (refer to the code listing). However , the final aim is
to use the MDL criterion to find the number of incident signals. The
next step is to obtain the noise eigenvector matrix of these minimum
eigenvalues and form the matrix nev. Using this nev , we can plot
to get direction of arrival in either a single dimension or multiple di-
mensions. The program presented works only for a planar angle. The
final aim is to get direction of arrival in terms of azimuthal angle and
elevation angle.

1.3 Testing
Ensure that all the files ( main, manifold,music1d are in same directory.
Run the main program in matlab and enter the number of signals and the
angle in degrees for each of the signals. The program will give the number
of signals as output along with the plot. Now we just have to obtain the
values of the highest d peaks to get the direction of arrivals.
Test Input 1: Enter value for number of signals as 1. Now , change the
angle from 10 to 80 and observe the output. We observe sharp peaks for
single signal.
Test Input 2: Enter an arbitrary value for number of signals ( less than
about 3/4th of the number of elements for good results ). Enter the values
of the angles and observe the output.

1.4 Design features and observed results

There are a number of parameters for which changes can be made and
varying results can be obeserved. They are

1. Number of elements: Increasing the number of elements increases

the resolution for multiple sources. If the number of signals is large ,
then a higher element array predicts the directions of arrival better.

4
 2. Noise With lower noise , the peaks become sharper. The presence of
added noise causes a spreading effect on the peaks.

3. Number of samples Increasing the number of samples does not bring

about a greater improvement in the plots. However , number of sam-
ples must be greater than number of incident signals for workable
results. 100 samples work well for a 15-element array.

4. Distance between elements The distance between elements can

be changed. This effectively means that for a fixed frequency of the
input wave,the centre frequency is changing that is the frequency at
which the detection is best. A lower wavelength and hence higher
frewquency gives no resolution at all ( fig 1 ). On the other hand ,
a higher wavelength (lower frequency) gives a lot of spurious peaks.
It is observed that for best detection of given frequency the distance
between elements must be 2 .

5. Kind of input wave The kind of input wave can be varied . Resluts
were verified for sine waves and FM-modulated waves ( refer PART IV
). The plots obtained match those in theory. However the algorithm
could not detect sin2 waves even on removing the DC offset of the
waves.

6. Obtain 2-d direction We can also extend the MUSIC algorithm to

2-d with the two dimensions being the plane angle and the elevation
angle respectively (refer PARTS V and VI ). However the results ob-
tained in 2-d are not so encouraging in terms of the number of peaks
obtained.

1.5 Further Work

Further work can proceeed on the following counts

1. Using the MDL criterion to find the number of incident signals to get
a better estimate.

2. Delve into why the algorithm does not work satisfactorily for inputs
like sin2 waves.

3. To improve the performance of the 2-d DOA performance.

5
1.6 Code Listing
PART I:
clear

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%% PART I - THE CALLING PROGRAM %%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

global NOOFELEMENTS;%no of elements

NOOFELEMENTS=15;
m=NOOFELEMENTS;

w=1000000000; %frequency (not angular) taken 1 GHz

global dfactor; % distance between elements= wavelegth / dfactor

dfactor=2;

nos=input(’enter no of signals’)

angl(nos)=0; % array to store the angles of arrival

for i=1:nos
angl(i)=input(’enter angle in degrees’)
theta(i)=angl(i)*pi/180;
end

samples=100; % no of samples of the input signal taken

fs=(.0001*w-1); % sampling frequency - rate at which samples are taken

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%% SIMULATION OF A POSSIBLE INPUT TO THE %%%%%%%%%%%
%%%%%%%% MUSIC ALGO %%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

f1(m,samples)=0; % array of sampled signal ie signal values at different

% elements at diff instants
for k=1:samples % no of time instants
for i=1:m % no of elements

6
 for ctr=1:nos
f1(i,k)=f1(i,k)+sin(2*pi*w*(k/fs+(1-i)/dfactor*sin(theta(ctr))/w))
% sine wave sampled - this function can be changed

end
end
end

r=0.3162/1.414*randn(m,samples); % RANDOM VARIABLE REPRESENTING ADDED NOISE

x=f1+r; % ADDING NOISE TO THE INPUT SIGNAL ARRAY TO GET THE POSSIBLE OUTPUT

noofsignals=music1d(x) % calling the music1d function (1-d music)

PART II:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%% PART II - ARRAY MANIFOLD %%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function a=manifold(angle)
global NOOFELEMENTS dfactor
no=NOOFELEMENTS;
for k=1:no
a(k,1)=cos((1-k)*2*pi*sin(angle)/dfactor)+sin((1-k)*2*pi*sin(angle)/dfactor)*j
end

PART III:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%% PART III - MUSIC Algorithm %%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function estimate=music1d(x)

temp=size(x); % returns the size

m=temp(1); % m is the number of elements
samples=temp(2); % no of samples of the incident signals

%finding covariance matrix of the input x

7
s(m,m)=0;
for i=1:m
for k=1:m
for n=1:samples
s(i,k)=s(i,k)+x(i,n)*x(k,n);
end
s(i,k)=s(i,k)/samples;
end
end

% eigenvalues of s
[evec eval]=eig(s); % evec is the set of eigenvectors and eval is the
% set of eigenvalues

minevalue=eval(1,1);% finding the minimum eigenvalue

for i=2:m
if(minevalue>eval(i,i))
minevalue=eval(i,i);
end
end

% finding the multiplicity of the minimum eigenvalue

n=0; %multiplicity
for i=1:m
if(eval(i,i)<(1+minevalue)) % assuming that all the roots less than
% 1+minevalue are minimum roots
% have to get a better approximator of
% multiplicity
n=n+1;
end
end

’Multiplicity of min root’

n

’no of incident signals’

m-n

8
 % forming the eigenvector matrix corresponding to the minimum eigenvalues
nev(m,n)=0; % noise eigenvector matrix

for i=1:n
for l=1:m
nev(l,i)=evec(l,i);
end
end

% now to plot using this matrix

ctr=1;
for angle=1:0.25:360 %in degrees
indep(ctr)=angle;
rad=angle*pi/180;
a=manifold(rad); % returns the array behaviour for a signal
% incident from certain direction
c=a’*nev;
d=c*c’;
pmu(ctr)=1/abs(d); % pmu(theta) is the function that is to be
% plotted as a function of theta
ctr=ctr+1;
end

plot(indep,pmu) % will plot PMU(theta) versus theta

grid
estimate=m-n % estimate of the number of incident signals

PART IV:
clear
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%% PART IV - THE CALLING PROGRAM %%%%%%%%%%%%%%%%
%%%%%%%%%%% FOR FM-MODULATED INPUT %%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

global NOOFELEMENTS;%no of elements

NOOFELEMENTS=15;
m=NOOFELEMENTS;

9
w=1000000000 ; %frequency (not angular)

global dfactor; % distance between elements= wavelegth / dfactor

dfactor=2;

nos=input(’enter no of signals’)

angl(nos)=0;

for i=1:nos
angl(i)=input(’enter angle in degrees’)
theta(i)=angl(i)*pi/180;
end

fs=(.0001*w-1); % sampling frequency - rate at which samples are taken

samples=100

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%SIMULATION OF A FM MODULATED INPUT TO THE %%%%%%%%%
%%%%%%%% MUSIC ALGO %%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

for k=1:samples % no of time instants

for i=1:m % no of elements
for ctr=1:nos
f1(i,k)=f1(i,k)+sin(2*pi*w*((1+.01*(sin(2*pi*w*(k/(.001*w-1)))))
+(1-i)/dfactor*sin(theta(ctr))/w));

end
end
end

r=0.3162/1.414*randn(m,samples); % random variable .

10
 x=f1+r;

noofsignals=music1d(x)

PART V:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%% PART V - MUSIC ALGORITHM %%%%%%%%%%%%%%%%
%%%%%%%%%%%% FOR 2-D DOA %%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function estimate=music1d(x)

temp=size(x); % returns the size

m=temp(1); % m is the number of elements
samples=temp(2); % no of samples of the incident signals

%finding covariance
s(m,m)=0;
for i=1:m
for k=1:m
for n=1:samples
s(i,k)=s(i,k)+x(i,n)*x(k,n);
end
s(i,k)=s(i,k)/samples;
end
end

% eigenvalues of s
[evec eval]=eig(s); % evec is the set of eigenvectors and eval is the
% set of eigenvalues

minevalue=eval(1,1);% finding the minimum eigenvalue

for i=2:m
if(minevalue>eval(i,i))
minevalue=eval(i,i);
end
end

n=0; %multiplicity
for i=1:m

11
 if(evalues(i)<(1+minevalue)) % have to get a better approximator of
% multiplicity
n=n+1;
end
end
’Multiplicity of min root’
n

’no of incident signals’

m-n

nev(m,n)=0; % noise eigenvector matrix

%k=1
for i=1:n%m
%if abs(eval(i,i))<(1+abs(minevalue))
for l=1:m
nev(l,i)=evec(l,i);%nev(l,i)=evec2(l,i);
end
%k=k+1
%end
end

% now to plot using this matrix

ctr1=1;
for angle1=1:1:90 %in degrees
indep1(ctr1)=angle1;
rad1=angle1*pi/180;
ctr2=1;
for angle2=1:1:90 %in degrees
indep2(ctr2)=angle2;
rad2=angle2*pi/180;
for k=1:m
a(k,1)=manifold3(rad1,rad2);
end
c=a’*nev;
d=c*c’;
pmu(ctr1,ctr2)=1/abs(d);
ctr2=ctr2+1;
end

12
 ctr1=ctr1+1;
end
’columns of I represent the elevation angles’
[Y,I]=max(pmu)
mesh(indep1,indep2,pmu)

PART VI:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%% PART VI - ARRAY MANIFOLD %%%%%
%%%%% FOR 2-D DOA %%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function a=manifold(angle1,angle2)
global NOOFELEMENTS dfactor
no=NOOFELEMENTS;
for k=1:no
a(k,1)=cos((1-k)*2*pi*sin(angle1)/dfactor*cos(angle2))
+sin((1-k)*2*pi*sin(angle1)/dfactor*cos(angle2))*j;
end

13