sensors

Article
Signal Source Positioning Based on Angle-Only Measurements
in Passive Sensor Networks
Yidi Chen 1,† , Linhai Wang 2,† , Shenghua Zhou 2, * and Renwen Chen 1
1 State Key Laboratory of Mechanics and Control of Mechanical Structures, Nanjing University of Aeronautics
and Astronautics, No. 29, Yudao Street, Nanjing 210016, China; 18001296740@189.cn (Y.C.);
rwchen@nuaa.edu.cn (R.C.)
2 National Laboratory of Radar Signal Processing, Xidian University, No. 2 Taibai Road, Xi’an 710071, China;
xd2020lhwang@163.com
* Correspondence: shzhou@mail.xidian.edu.cn; Tel.: +86-29-88201220
† The authors contributed equally to this work.

Abstract: Some passive sensors can measure only directions of arrival of signals, but the real positions
of signal sources are often desirable, which can be estimated by combining distributed passive
sensors as a network. However, passive observations should be correctly associated first. This
paper studies the multi-target data association and signal localization problem in distributed passive
sensor networks. With angle-only measurements from distributed passive sensors, multiple lines
in a 3-dimensional (3D) scenario can be built and then those that will intersect in a small volume in
3D are classified into the same source. The center of the small volume is taken as an estimate of the
signal source position, whose statistical distributions are formulated. If the minimum distance is less
than an association threshold, then two lines are considered to be from the same signal source. In
numerical results, the impacts of angle measurement accuracy and platform self-positioning accuracy
are analyzed, indicating that this method can achieve a prescribed data association rate and a high



positioning performance with a low computation cost.


Citation: Chen, Y.; Wang, L.; Zhou, S.; Keywords: passive sensor network; signal localization; data association; angle-only measurements;
Chen, R. Signal Source Positioning accuracy analysis
Based on Angle-Only Measurements
in Passive Sensor Networks. Sensors
2022, 22, 1554. https://doi.org/
10.3390/s22041554 1. Introduction
Academic Editors: Ratnasingham Unlike active sensors such as radars, passive sensors do not transmit signals and thus
Tharmarasa and Mahendra Mallick have no anti-jamming problem [1,2]. However, some passive sensors, such as infrared
Received: 10 January 2022
sensors, photoelectric sensors, electronic counter measurement (ECM) and cameras, can
Accepted: 14 February 2022
estimate only angles of signal sources. Therefore, their signal source positioning perfor-
Published: 17 February 2022
mances are typically poor since they have no accurate range information of signal sources.
In order to estimate the positions of signal sources, passive sensors with angle-only ob-
Publisher’s Note: MDPI stays neutral
servations can be connected with communication links into a network to measure signals
with regard to jurisdictional claims in
sources from different spatial locations. In this case, an algorithm to combine the angle-only
published maps and institutional affil-
observations is needed [3–7]. Compared with the time of arrival (TOA) [8] and the time
iations.
difference of arrival (TDOA) localization , angle-only localization does not require accurate
time synchronization between distributed passive sensors for signal sources with low
speeds [6].
Copyright: © 2022 by the authors.
In a passive sensor network, there may be multiple signal sources, and before accurate
Licensee MDPI, Basel, Switzerland. localization, one should first correctly associate observations regarding the same sources [9].
This article is an open access article The multi-dimension assignment model is a classical method for data association in passive
distributed under the terms and sensor networks [10–12], but it needs the locations of signal sources, which is unavailable
conditions of the Creative Commons before correct data association. A geometry-based localization algorithm for a distributed
Attribution (CC BY) license (https:// sensor network is presented in [13], which constructs a test statistic based on the minimal
creativecommons.org/licenses/by/ distance between the lines of sight for data association. The measurement errors are
4.0/). considered, but the platform self-positioning errors are not considered.

Sensors 2022, 22, 1554. https://doi.org/10.3390/s22041554 https://www.mdpi.com/journal/sensors

Sensors 2022, 22, 1554 2 of 18

In this paper, how to perform data association and signal source localization in a
3-dimensional (3D) scenario for distributed passive sensors with only angle measurements
is studied. We improve the intersection localization algorithm for passive sensor networks
in a multi-target scenario. We first consider the data association problem and then the
signal source position estimation problem. The basic concept is to construct a set of lines
in a 3D scenario according to angle measurements of signal sources. In data association,
measurement lines that will intersect within a small space volume are categorized into the
same group. The statistical distribution of the minimal distances of the lines are formulated
and the minimum distance between any two observation lines is a random variable, proved
to follow the Chi-square distribution. The threshold for correct association is formulated
by the misassociation probability; namely, two observations are from the same signal
source but are classified into two groups. In the test statistics, not only the measurement
errors but also the platform positioning errors are considered, which makes the association
performance robust when the platform positioning errors exist.
After data association, observations regarding the same signal sources are grouped,
based on which the location of signal sources can be estimated. Three positioning algo-
rithms are considered. It is known that angle measurements are nonlinear functions of
coordinates of signal sources. With the Taylor expansion, the least square (LS) algorithm
linearizes the nonlinear angle measurements about the target position and then uses the LS
method to obtain the target position estimate [6,14–19]. In real applications, different pas-
sive sensors may obtain observations of different signal-to-noise ratios (SNRs) and then the
weighted least squares algorithm (WLS) [6,14] and total least square (TLS) algorithm [20]
can be used to obtain a better estimate. Another source location method is the intersection
localization algorithm [1,4,5,21–23]. The basic concept is that if multiple passive sensors
simultaneously measure the signal sources without measurement error, these measurement
lines of sight will intersect to the target position. The geometric method and algebraic
solution method use this property to estimate the positions of signal sources.
The data association process and target location process of this method are closely com-
bined, which ensures a lower algorithm complexity and a better positioning performance.
In numerical results, the improvement of data association and signal-source positioning
are analyzed. The impact of the target-sensor geometry on the localization accuracy is also
studied, indicating that the localization performance will be better if the lines associated
with different observations are perpendicular to each other.
We follow the convention that bold lower and upper case letters denote column vectors
and matrices, respectively. A symbol with an upper script o denotes the true value. For
instance, ao denotes the true value of a. diag(·) with a vector entry denotes a diagonal
matrix with the entry vector as diagonal elements. The notation diag(A1 , A2 , . . . , A N )
stands for the block-diagonal matrix formed by the matrices A1 , A2 , . . . , A N .

2. Localization with Angle-Only Passive Sensors

2.1. Signal Model of Passive Observations
Consider a passive sensor network with N widely separated sensors and M targets in
the surveillance volume. Assume that a coordinate system is available for all the sensors,
such as earth-centered earth-fixed (ECEF) of the World Geodetic System 84 (WGS84). The
real position of the nth sensor at instant t is denoted as
son (t) = [ xn,s
o ( t ), yo ( t ), zo ( t )]T , n = 1, 2, . . . , N, where (·)T denotes the transpose op-
n,s n,s
eration, t denotes time, and xn,s o ( t ), yo ( t ), zo ( t ) denote the x, y, z coordinates of the nth
n,s n,s
sensor at instant t, respectively. The real position of the mth target at instant t is denoted
by gom (t) = [ xm,g
o ( t ), yo ( t ), zo ( t )]T , m = 1, · · · , M, where x o ( t ), yo ( t ), zo ( t ) denote
m,g m,g m,g m,g m,g
the x, y, z coordinates of the mth target at instant t, respectively. Assume that there is a
self-positioning device to measure the position of each sensor. There are different kinds of
instruments that can measure the position of a platform, such as the Global Positioning
System (GPS) and inertial sensors. The topology of the passive sensors and targets are
shown in Figure 1.
Sensors 2022, 22, 1554 3 of 18

sk,N gom (tk,n )

soN (tk,N ) goM (tk,n )
g1o (tk,n )

sk,n
o ϕom,k,n
sn (tk,n ) y
o sk,1
θm,k,n o
s1 (tk,n )

x
Figure 1. Measurement scenario of the passive sensors.

For the nth sensor, signals are detected at instants denoted by tk,n , k = 1, · · · , Nn ,
where Nn denotes the number of observations of the nth sensor. At instant tk,n , assume that
the position of the nth sensors is measured as

sk,n = son (tk,n ) + ∆sn (tk,n ) = [ xn,s (tk,n ), yn,s (tk,n ), zn,s (tk,n )]T , k = 1, · · · , Nn (1)

where ∆sn (tk,n ) denotes the sensor self-positioning error. For simplicity, we assume that
the positioning errors follow zero mean Gaussian distributions with covariance matrices
Ck,n = E(∆sn (tk,n )∆sTn (tk,n )), where E denotes the expectation operation. In practice,
the self-positioning error of the sensor is approximately subject to zero-mean Gaussian
distribution, so this assumption is reasonable and widely used.
For the angle-only sensors, the observations are directions of the signals and the
lth signal at instant tk,n is denoted by θl,k,n = [θl,k,n , ϕl,k,n ]T , where (l, k, n) ∈ Mk,n ,
k = 1, · · · , Nn and n = 1, · · · , N, and Mk,n denotes a set of triples of signals detected
at the kth measurement by the nth sensor. Assume that |Mk,n | = Mk,n , where | · | over a
set denotes the cardinality of the set. As the existence of miss detection, false alarms and
overlapping of signal sources, Mk,n may not be equal to M. Denote Mn = ∪kN=n 1 Mk,n and
A = ∪nN=1 Mn , where ∪ denotes the union operation. The total number of observations by
N sensors is denoted by
N Nn
Ns = |A| = ∑ ∑ Mk,n . (2)
n =1 k =1

At instant tk,n , the real position of the mth signal source is denoted by

gom (tk,n ) = [ xm,g (tk,n ), ym,g (tk,n ), zm,g (tk,n )]T , m = 1, · · · , M. (3)

For the mth signal source, the real azimuth angle and elevation angle regarding the
nth sensor can be expressed by
 
o
θm,k,n = tan−1 yom,g (tk,n ) − yon,s (tk,n ), xm,g
o o
(tk,n ) − xn,s (tk,n ) (4)
 q 
ϕom,k,n = arctan (zom,g (tk,n ) − zon,s (tk,n ))/ ( xm,go ( t ) − x o ( t ))2 + ( yo ( t ) − yo ( t ))2
k,n n,s k,n m,g k,n n,s k,n

respectively, where θm,k,no ∈ (−π, π ), ϕom,k,n ∈ (− π2 , π2 ), tan−1 (∗) is called the two-argument
inverse tangent function [24,25] and arctan(∗) is the inverse tangent function. Denote
θom,k,n = [θm,k,n
o , ϕom,k,n ]T .
Each observation is associated with one of M targets or the false alarm indexed by 0,
represented by a set M = {0, 1, · · · , M}. It can be considered as a mapping ψ : A → M.
Sensors 2022, 22, 1554 4 of 18

According to our setting, the index set A can be partitioned into M + 1 disjoint sets
A0 , A1 , · · · , A M defined by

Ai = { x |ψ( x ) = i, x ∈ A}, (5)

where A0 denotes the index of observations corresponding to false alarms, and Am denotes
the index set of observations from the mth signal source. As a partition of A, we have
Ai ∩ A j = ∅, i, j ∈ M, i 6= j, and A = ∪iM
=0 Ai , where ∩ denotes the intersection operation
of sets.
If the mth signal source is detected and indexed as the (l, k, n) observation, then the
azimuth angle and elevation angle measurements can be written as

θl,k,n = [θl,k,n , ϕl,k,n ]T (6)

= θom,k,n + ∆θl,k,n (7)
o
θl,k,n = θm,k,n + ∆θl,k,n (8)
o
ϕl,k,n = ϕm,k,n + ∆ϕl,k,n (9)
∆θl,k,n = [∆θl,k,n , ∆θl,k,n ]T (10)

where ∆θl,k,n and ∆ϕl,k,n represent the measurement noise of the azimuth angle and ele-
vation angle, respectively. For simplicity, we assume that observation noises ∆θl,k,n and
∆ϕl,k,n , l = 1, · · · , Mk,n , k = 1, · · · , Nn , n = 1, · · · , N are statistically independent and
follow zero-mean Gaussian distribution by assumption. The covariance matrices of ∆θ
are denoted by Cl,k,n = E(∆θ∆θT ) ∈ C2×2 , namely ∆θ ∼ N (0, Cl,k,n ), which is typically
affected by the SNR of the signal, where N (0, Cl,k,n ) denotes the zero-mean Gaussian distri-
bution with mean 0 and covariance matrix Cl,k,n . It should be noted that modeling the angle
measurement noise as a zero-mean Gaussian distribution is a commonplace assumption.

2.2. The Distance of Observation Lines

The angle-only observations provide information on the directions of signal sources,
and in theory, the directions can be expressed by 3D lines. Therefore, it is important to study
the properties of the lines. Consider a simple scenario where the target can be deemed to
be static when the observations are recorded and the location of the mth signal is simplify
denoted by gom . For passive observations, each observation, say (l, k, n), will contribute a
line in space and the line can be written as Ll,k,n

Ll,k,n : x = sk,n + αl,k,n el,k,n , αl,k,n ∈ R (11)

where αl,k,n , a parameter indicating the distance to the origin sk,n , en,m = [el,k,n,x , el,k,n,y , el,k,n,z ]T ∈
R3×1 , is the normalized direction vector associated with the angle observation θl,k,n, and

el,k,n,x = cos(θl,k,n ) cos( ϕl,k,n ) (12)

el,k,n,y = sin(θl,k,n ) cos( ϕl,k,n ) (13)
el,k,n,y = sin( ϕl,k,n ). (14)

It can be seen that the subscripts of the denotations are complicated. Therefore, for
two observations i, j ∈ Am , we simplify the expression of lines by

Li,m : x = si + αi,m ei,m (15)

0
L j,m : x = s j + α j,m e j,m (16)

where the locations of two sensors regarding the two observations are denoted by si , s j ,
αi,m , α j,m are two scalars indicating the distances to two origins, and ei,m and e j,m are two
normalized vectors associated with two observations.
Sensors 2022, 22, 1554 5 of 18

The difference between two points over the two lines are

⇒ x0 − x = s j − si + α j,m e j,m − αi,m ei,m (17)

= s j − si + (−ei,m , e j,m )α, (18)

where  
αi,m
α= . (19)
α j,m
Without measurement error, then there will be two αi,m and α j,m such that x0 − x = 0, i.e.,

gm = si + αi,m ei,m = s j + α j,m e j,m . (20)

In general, due to inevitable measurement errors, the lines even regarding the same
signal source may not coincide to each other. Therefore, we calculate minimal distance
between those lines. The distance between two points over two lines can be expressed by

d = k x 0 − x k2
= ks j − si + (−ei,m , e j,m )αk2 (21)
= αT ETi,j Ei,j α + 2αT ETi,j (s j T
− si ) + ( s j − si ) ( s j − si )

where k · k denotes the `2 -norm, α = [αi,m , α j,m ]T , and Ei,j = [−ei,m , e j,m ]. For simplicity,
the distance d is actually the squared distance, instead of the distance.
In particular, if ei,m = e j,m , namely two lines are parallel to each other, then

d = ks j − si + (α j,m − αi,m )ei,m k2 , (22)

and the minimal distance between points in two lines will be achieved if

α j,m − αi,m = eTi,m (si − s j ). (23)

It can be proved that the minimal distance is

d = k(ei,m eTi,m − I)(si − s j )k2 . (24)

If ei,m 6= e j,m or |Ei,j | 6= 0, the distance is a second order function of αi,m and α j,m , thus
the minimal value of d is unique, where | · | over a matrix denotes the determinant of the
input matrix. To obtain the minimal value, we take a derivative of d with respect to α,
" #T
dd dd dd
= , (25)
dα dαi,m dα j,m
= 2ETi,j Ei,j α + 2ETi,j (s j − si ) (26)

where
dd
= −2eTi,m (s j − si + Eα) (27)
dαi,m
dd
= 2eTj,m (s j − si + Eα). (28)
dα j,m

Let the derivative be zero and then we obtain a solution

dd
=0⇒ ETi,j Ei,j α = −ETi,j (s j − si ). (29)
dα
Sensors 2022, 22, 1554 6 of 18

Under the assumption that |Ei,j | 6= 0, Ri,j = ETi,j Ei,j has a reverse matrix and then the
solution can be immediately obtained by

αopt = −(ETi,j Ei,j )−1 ETi,j (s j − si ). (30)

The minimal distance can be expressed by

−1 T
dmin = (s j − si )T (I − Ei,j Ri,j Ei,j )(s j − si ) (31)
−1 T
= k(I − Ei,j Ri,j Ei,j )(s j − si )k2 . (32)

In particular, if
−1 T
Ei,j Ri,j Ei,j (s j − si ) = s j − si , (33)
−1 T
namely, s j − si is an eigenvector of Ei,j Ri,j Ei,j and the eigenvalue is 1, then

dmin = 0. (34)

It means that two lines will intersect at a point. Without measurement errors, observa-
tions regarding the same target will form lines intersecting at the target position.
In practice, the real mapping ψ should be estimated through an association algorithm.
Due to measurement errors, the estimated mapping may not be correct, and then, the
positioning error may raise. Therefore, an accurate data association method is important,
which will be studied subsequently.

3. Data Association Based on Minimal Distance

3.1. Data Association Model
In order to perform data association, we first need to build the statistical model of the
minimal distance of the observation lines. Because of both the platform positioning error
∆si , ∆s j and the angle measurement error ∆θi,m and ∆θj,m , i, j ∈ Am , the minimal distance
dmin is not zero and follows a certain distribution depending on the measurement errors,
where ∆si , ∆s j denote the sensor self-positioning errors of the ith and the jth observations,
respectively, and ∆θi,m , ∆θj,m denote the angle measurement error of the ith and the jth
observations on the mth target.
In practice, the ith and the jth observations may belong to the same target or not, and
the data association problem is treated as a test to make a decision here. For i, j ∈ A, the
data association problem for a target can be formulated as the following hypothesis problem
(
H0 : dmin = 0 ψ(i ) = ψ( j)
(35)
H1 : dmin > 0 ψ(i ) 6= ψ( j).

To determine the statistical distribution of dmin under the H0 hypothesis, we first define
−1 T
Ki,j = I − Ei,j Ri,j Ei,j (36)

and then
dmin = kKi,j (s j − si )k2 . (37)

Next, we discuss the eigenvalues of Ki,j . It can be proved that KTi,j Ki,j = Ki,j and
KTi,j = Ki,j . Therefore, Ki,j is a positive semi-definite matrix and its possible eigenvalues
are either 0 or 1. As Ki,j is not an all-zero matrix, then there is at least an eigenvalue of 1.
Sensors 2022, 22, 1554 7 of 18

The trace of the matrix K(i, j) satisfies

−1 T
tr(Ki,j ) = tr(I − Ei,j Ri,j Ei,j ) (38)
−1 T
= tr(I) − tr(Ei,j Ri,j Ei,j ) (39)
−1
= 3 − tr(ETi,j Ei,j Ri,j ) (40)
−1
= 3 − tr(Ri,j Ri,j ) (41)
2×2
= 3 − tr(I2 ∈ R )=1 (42)

where tr(·) with a matrix input denotes the trace of the matrix.
Therefore, it can be inferred that three eigvenvalues of Ki,j are 1, 0, 0 and the rank of
Ki,j is 1. In other words, Ki,j can be written as

Ki,j = es eTs (43)

where es is a unity direction vector perpendicular to both ei and e j , defined by

q
es = e i × e j / 1 − (eTi e j )2 (44)

and ei and e j denote the unity direction vectors associated with the ith and the jth observa-
tions, respectively.
With this fact, the minimal distance can be rewritten as

dmin = |eTs (s j − si )|2 . (45)

In practice, both s j − si and es may be inaccurate. A statistical distribution is necessary

to determine the impact of the measurement errors. In order to formulate the statistical
distribution of dmin under the H0 hypothesis, we define

r = eTs (s j − si ) (46)

and its relationship to dmin is dmin = |r |2 .

The minimal distance varies with a total of 10 parameters; namely θi,m , θj,m , si,m and
s j,m , and the corresponding partial derivatives can be written as

∂r T
 
∂r ∂r
= , (47)
∂θi,m ∂θi,m ∂ϕi,m
" #T
∂r ∂r ∂r
= , (48)
∂θj,m ∂θ j,m ∂ϕ j,m
∂r
= − es (49)
∂si
∂r
= es . (50)
∂s j

Consequently, under the H0 hypothesis, with denotation v = [θTi,m , θTj,m , sTi,m , sTj,m ]T , we
have an approximation

∂r ∂r ∂r ∂r
r≈ T
∆θi,m + T ∆θj,m + T ∆si,m + T ∆s j,m (51)
∂θi,m ∂θj,m ∂si ∂s j
∂r
= ∆v (52)
∂vT
Sensors 2022, 22, 1554 8 of 18

where ∆v = [∆θTi,m , ∆θTj,m , ∆sTi,m , ∆sTj,m ]T , and

" #
∂r ∂r ∂r ∂r ∂r
= , , , . (53)
∂vT ∂θTi,m ∂θTj,m ∂sTi ∂sTj

Under the assumption that the measurement errors ∆θi,m , ∆θj,m , ∆si,m , ∆s j,m are statisti-
cally independent of each other and follow a zero-mean Gaussian distribution, r follows the
Gaussian distribution and then dmin follows the central weighted Chi-square distribution
with 1 degree of freedom, namely dmin ∼ χ21 (λ), where λ is the variance

∂r ∂r ∂r ∂r
λ = E( T
∆v∆vT ) = T
Cv (54)
∂v ∂v ∂v ∂v
and
Cv = E(∆v∆vT ). (55)
The probability density function (PDF) and cumulative distribution function (CDF)
can be written as
1 d 1 d
pdmin (d) = √ ( )− 2 exp(− ), d ≥ 0 (56)
2λ π λ 2λ
1 d 1
Fdmin (d) = √ γ( , ), d ≥ 0 (57)
π 2λ 2

respectively, where Γ(·) denotes the Gamma function and γ(·, ·) denotes the incomplete
Gamma function. If the decision rule is to keep the misassociation probability P( H1 | H0 ) as
a constant, say pf , then the decision threshold can be obtained as
√ 1
ρ = 2λγ−1 ( π (1 − pf ), ) (58)
2

where γ−1 (·, 12 ) denotes the inverse incomplete Gamma function with 12 a degree
of freedom.
Therefore, the key is to derive the variance λ. In practice, as the measurements are
carried out by different sensors, it is reasonable to assume that the measurement errors are
statistically independent of each other. In this case, Cv is a block diagonal matrix and the
variance can be formulated conveniently.

3.2. Measurement Errors and Association Threshold

For simplicity, we first consider the self-positioning error ∆si , ∆s j , whose covariance
matrices are assumed to be

Ci,s = E(∆si ∆sTi ) (59)

C j,s = E(∆s j ∆sTj ). (60)

Consequently, from (49) and (50), the variances due to the two terms are

∂r ∂r T
λi,s = E( ∆s ∆sT ( ) ) = eTs Ci,s es (61)
∂si i,m i,m ∂si
∂r ∂r
λ j,s = E( ∆s j,m ∆sTj,m ( )T ) = eTs C j,s es . (62)
∂s j ∂s j
Sensors 2022, 22, 1554 9 of 18

Next, we consider the variance caused by the angle measurement error ∆θi,m and
∆θj,m . Assume that the covariance matrices of ∆θi,m and ∆θj,m are

Ci,θ = E(∆θi,m ∆θTi,m ) (63)

C j,θ = E(∆θj,m ∆θTj,m ) (64)

respectively. The variances caused by the two terms are

∂r ∂r ∂r ∂r ∂r ∂r
λi,θ =E( ∆θi,m ∆θTi,m ) = T E(∆θi,m ∆θTi,m ) = T Ci,θ (65)
∂θTi,m ∂θi,m ∂θi,m ∂θi,m ∂θi,m ∂θi,m
∂r ∂r ∂r ∂r ∂r ∂r
λ j,θ =E( T
∆θj,m ∆θTj,m ) = T E(∆θj,m ∆θTj,m ) = T C j,θ (66)
∂θj,m ∂θj,m ∂θj,m ∂θj,m ∂θj,m ∂θj,m

It can be seen that the distance is not a linear function of θ. As a cross product of
ei,m and e j,m , es is a complicated function. From (47) and (48), ∂θ∂r and ∂θ∂r are proved in
i,m j,m
Appendix A.
Under these assumptions, Cv = diag(Ci,θ, C j,θ, Ci,s , Ci,s ). Consequently, the variance
λ can be expressed by
λ = λi,s + λ j,s + λi,θ + λ j,θ. (67)
It should be noted that under the H1 hypothesis, the distance may be arbitrary, and
for simplicity, we assume that the distance is uniformly distributed over the surveillance
volume. In this case, we can determine whether two observations are from the same target
by the following decision rule

H1
dmin ≷ ρ. (68)
H0

The association algorithm is a mapping ψ0 : A → M, which partitions A into M + 1

groups and this mapping may disagree with real ψ due to inevitable errors. In practice,
some knowledge about the targets of interest may be available and thus can be used for
better performance. For instance, if only targets on the ground are of interest, then we may
use this information and discard observations in which the cross points are obviously apart
from the ground, which will be studied in the future.

3.3. Localization Algorithms

With mapping ψ0 , we obtain another partition A = ∪m
M A0 , where A0 = { x | ψ0 ( x ) =
=0 m m
0
m, x ∈ A}. For observations in Am , we can conduct signal source positioning, and three
target positioning estimation methods will be introduced subsequently.
We first consider the intersection method. In presence of measurement errors, we
often have dmin 6= 0 even if two observations are from the same target. In this case, over
two lines, the two points with minimal distance are

xi = si + αopt (1)ei,m (69)

x j = s j + αopt (2)e j,m . (70)

In this case, it is reasonable to take the middle of two points as the estimate of the
signal position, namely

1
ĝi,j =(x + x j )
2 i (71)
1 1 −1 T
= (si + s j ) − (ei,m , e j,m )Ri,j Ei,j (s j − si ).
2 2
Sensors 2022, 22, 1554 10 of 18

With more observations available, there will be an estimate of the target location for
each observation pair , and a simple estimate of the target location can be expressed by
their average, namely
1
Nm ( Nm − 1) i,j∑
ĝm = ĝi,j (72)
∈A m

where Nm denotes the number of observations associated with the mth signal source.
In practice, another widely used localization algorithm is the LS algorithm. It stems
from the delta method concerned in [26]. From (4), the ith angle observation denoted by
θi,m actually contributes a geometric relationship, which can be expressed by

Gi g m = Gi si , i ∈ A m (73)

where
 
sin θi,m − cos θi,m 0
Gi = . (74)
cos θi,m sin ϕi,m sin θi,m sin ϕi,m − cos ϕi,m

In (74), we used the equality

o o o
( xm,g − xi,s ) cos θi,m + (yom,g − yi,s
o o
) sin θi,m (75)
q
o − x o )2 + ( y o − y o )2
= ( xm,g (76)
i,s m,g i,s
o o
=di,m cos ϕi,m
o is the true range of go to so and can be expressed by do = k go − so k.
where the value di,m m i i,m m i
Next, we can combine the observations into an equation

Ggm = y (77)

where
   
G1 G1 s1
G =  ... , y= ..
. (78)
   
.
G Mm G Mm s Mm

It should be noted that after the data association operation, the number of samples
associated with a target may not agree with the real number. For simplicity, we still use
Mm = |A0m | to denote the number of observations associated with the mth target.
With Mm observations, there are in total 2Mm equations and three unknown parame-
ters in gm . Therefore, as if Mm ≥ 2, we can use the LS method to obtain an optimal solution
in the sense of the mean square error, as

ĝm = (GT G)−1 Gy. (79)

In practice, different observations may have different SNRs, and then, the WLS algo-
rithm can also be used in this framework. Then, the angle measurement error and sensor
positioning error are equally weighted in the process of the LS algorithm. Assume that the
distribution of the angle measurement error and the positioning error of sensors are known
a priori. With this information, we can impose different weights over the observations,
which is the WLS algorithm.
With the following approximations
o o
sin θi,m ≈ sin(θi,m ) + ∆θTi,m cos θi,m (80)
o o
cos θi,m ≈ cos(θi,m ) − ∆θTi,m sin θi,m (81)
Sensors 2022, 22, 1554 11 of 18

we can put (80) and (81) into (73) and then write (73) as

ei = Gi (gom − si ) (82)

where

ei = [ei,1 , ei,2 ]T (83)

o o oT
ei,1 = ∆θi,m di,m cos ϕi,m + ai,m ∆si (84)
o oT
ei,2 = ∆ϕi,m di,m + bi,m ∆si (85)
o o o
ai,m = [− sin θi,m , cos θi,m , 0]T (86)
o o o o o o T
bi,m = [− cos θi,m sin ϕi,m , − sin θi,m sin ϕi,m , cos ϕi,m ] . (87)

In (84), we have used the equality (76). In (85), we have used the equality
o o o
( xm,g − xi,s ) sin θi,m − (yom,g − yi,s
o o
) cos θi,m =0 (88)
(gom − sio )T [cos θi,m
o
cos o
ϕi,m o
, sin θi,m cos o
ϕi,m , sin o T
ϕi,m ] = o
di,m (89)

which can be easily verified in the angle measurement equations.

The vector ei can be rewritten as

ei = Bio δi (90)

where
" #
o cos ϕo o,T
di,m 0 ai,m
Bio = i,m
o o,T (91)
0 di,m bi,m
δi = [∆θi,m , ∆ϕi,m , ∆sTi ]T . (92)

Under the assumption that the self-positioning error and the angle measurement error
are decorrelated, it can be proven that the covariance matrix of ei can be written as

Qn = E(eeT ) = diag(Cn,θ, Cn,s ). (93)

Putting (90) into (82) and combining the observations into an equation yield

Bo δ = Ggom − y. (94)

where

B1o
   
δ1
 . 
Bo =  .. , δ =  ... . (95)
 

BoMm δ Mm

In the WLS algorithm, the goal is to minimize the objective function J (gm ) as

J (gm ) = (Ggm − y)T W(Ggm − y) (96)

where W is the weighting matrix with the expression

W = E[Bo δδT BoT ]−1 = (Bo QBoT )−1 (97)

and Q is the error covariance matrix with an expression Q = diag(Q1 , Q2 , . . . , Q N ).

Sensors 2022, 22, 1554 12 of 18

The variable gm to minimize the objective function J (gm ) can be calculated by the least
square method and the estimate of the target positioning can be expressed by

ĝm = (GT WG)−1 GT Wy (98)

which is the WLS algorithm.

In WLS, the covariance matrices of different observations should be known a priori. If
the observations can be deemed to have close SNRs, then one can simply use the LS algo-
rithm. If the covariance matrices are not known with certain accuracy, some performance
loss may occur, which should be analyzed with numerical analysis.

4. Numerical Results
We first study the localization performance in the presence of a platform of self-
positioning errors and angle measurement errors. Then, the data association performance
will be analyzed, followed by the analysis of the impact of sensor-target geometry.
Consider a scenario where two sensors are installed on two aircraft and three targets of
interest are in the scope. Assume that two aircraft move at the same speed. The parameter
settings are shown in Table 1. For simplicity, we assume that two passive sensors collect
their observations on the same instants, namely, tk,i = tk,j for all k and i, j ∈ {1, · · · , N }.
Meanwhile, the sampling interval is 0.1 s and the simulation runs for 10 seconds. Assume
that the covariance matrices of angle measurement error for all sensors and all targets are the
same as Ck,θ = σθ2 I2 , ∀k. The self-positioning error covariance is Cn,s = σs2 I3 , n = 1, · · · , N.

Table 1. Positions and velocities of sensors and targets.

Position (m) at t = 0 s Velocity (m/s) Position (m) at t = 10 s

Sensor #1 so1 (0) = [0, 0, 0]T [50, 100, 0]T so1 (10) = [500, 1000, 0]T
Sensor #2 so2 (0) = [12,000, 10,000, −800]T [50, 100, 0]T so2 (10) = [12,500, 11,000, −800]T
Target #1 go1 (0) = [18,000, 12,000, 8000]T [20, 30, 0]T go1 (10) = [18,200, 12,300, 8000]T
Target #2 go2 (0) = [15,000, 13,000, 7000]T [20, 30, 0]T go2 (10) = [15,200, 13,300, 7000]T
Target #3 go3 (0) = [13,000, 12,000, 5000]T [20, 30, 0]T go3 (10) = [13,200, 12,300, 5000]T

4.1. Impact of Self-Positioning Error and Angle Measurement Error

We first study the impacts of sensor positioning error and angle measurement error in
the proposed intersection method. In order to make a comparison between the intersection
method, the LS algorithm [19] and the WLS algorithm [6], we first consider the positioning
accuracy of target #1 with the two sensors at t = 0 s.
In order to analyze the positioning performance of the concerned algorithms, we
measure the performance with the average root mean square error (RMSE), defined for a
target, e.g., target #1, by
v
Ns
u
u 1
RMSE = t
Ns ∑ kĝ1 (k) − go1 k2 (99)
k =1

where ĝ1 (k) denotes the estimate of the target position at the kth simulation run for the
target #1, and Ns denotes the number of simulation runs.
With Ns = 10,000 Monte Carlo runs, the RMSE of different self-positioning errors
are shown in Figure 2, under the assumption that the covariance matrix is C1,s = σs2 I3 ,
where we always set σθ = 0.03◦ in this simulation. It can be seen that as σs rises from
0 to 20 m, the three algorithms have close performances and the WLS performs better,
Sensors 2022, 22, 1554 13 of 18

under the assumption that the covariance matrices of Ck,θ , k = 1, 2, 3 and Cn,s , n = 1, 2 are
known exactly.

Figure 2. Average RMSE of target localization algorithms based on angle measurements of two
sensors at a fixed angle measurement error.

The impacts of the angle measurement error are shown in Figure 3, where σθ raises
from 0◦ to 1.5◦ , where we set σs = 5 m in this simulation. It shows that the three algorithms
have close performances and the intersection method and the LS method perform a little
worse than the WLS method. The rise of both the self-positioning error and the angle
measurement error will cause the rise of the target positioning error. However, with a better
weighting, the WLS algorithm often performs better than the intersection method and the
LS method.

Figure 3. Average RMSE of target localization algorithms based on angle measurements of two
sensors at a fixed self-positioning error.

4.2. Data Association Performance

Consider the parameters of the two sensors and the targets, as shown in Table 1. As the
targets are moving in this scenario, we set the sampling interval to 0.1 s. At each sampling
instant, each sensor has three measurements corresponding to the three targets. Therefore,
the two sensors will totally have nine measurement combinations at each sampling instant,
of which three combinations are correct. In data association, we set pf = 1% and then the
probability of correct association pc = 99%.
With Ns = 2000 Monte Carlo runs, the averaged correct association probability is
shown in Figure 4, where the self-positioning error is σs = 5 m, and the angle measurement
error is σθ = 0.03◦ . It can be seen that during the whole sampling period, the correct associ-
Sensors 2022, 22, 1554 14 of 18

ation probability of the three targets is close to the present value 99%, namely 20 wrong
combinations, on average, for each instant.

Figure 4. Probability of correct association.

Next, we consider the location performance with observations after the data associa-
tion operation. The RMSEs of the algorithms over the three targets based on the proposed
method are shown in Figure 5. In Figure 5, the RMSEs of the three targets are 23.43, 30.05
and 14.45 m. Note that the data association error may affect the localization performance
in this case. From Figure 5, the relative position of the sensor and the target will affect
the positioning performance, so the impact of the geometry of sensors and targets on the
positioning performance will be studied subsequently.

Figure 5. The localization error of the three targets.

4.3. Impact of Target-Sensor Geometry

The geometry will play an important role in the positioning performance. Consider
the scenario shown in Figure 6, where a target is probed with two passive sensors and the
intersection angle of two azimuth lines is denoted by φ. In the 3D scenario, we also define
a new elevation angle η for convenience and the distance of the target is the same for both
the sensors, i.e., 8000 m.
It is assumed that the positions of the two sensors are accurately measured and
target #1 can be detected by both two sensors. The two sensors are symmetrically dis-
tributed on both sides of target #1. We explore the impact of geometry on the positioning
accuracy by changing φ and η. In order to illustrate that the errors caused by geometry on
the three coordinate axes are different, we define ∆x, ∆y and ∆z to represent the RMSE on
the three coordinate axes, respectively, which can be expressed by
Sensors 2022, 22, 1554 15 of 18

v
u 1 Ns
u

Ns k∑
∆x = t o k2
k x̂1,g (k) − x1,g (100)
=1
v
u 1 Ns
u

Ns k∑
∆y = t kŷ1,g (k) − yo1,g k2 (101)
=1
v
u 1 Ns
u

Ns k∑
∆z = t kẑ1,g (k) − zo1,g k2 . (102)
=1

With Nps = 10,000 Monte Carlo runs, the ∆x, ∆y, ∆z and the spatial error sum denoted
by sum = ∆x2 + ∆y2 + ∆z2 are shown in Figure 7, under the assumption that η = 15◦ ,
and the angle measurement error σθ = 0.03◦ . It can be seen that when the intersection
angle φ is changed, the positioning error ∆x, ∆y, ∆z are different. When the intersection
angle φ is equal to 82.58◦ , the spatial error sum is the best, about 5.58 m, and ∆x, ∆y and
∆z are approximately equal to each other. In fact, the angle around φ = 90o will all lead to
high accuracy.

s1 s2

O y
η
φ

x target

Figure 6. Definition of the intersection angle φ and pitch angle η of the plane.

Figure 7. Relationship between localization error and intersection angle φ.

The impacts of intersection φ and η on spatial error sum are shown in Figure 8. For η
in the range of 0◦ to 45◦ and φ in the range of 25◦ to 175◦ , the spatial error sum is less than
Sensors 2022, 22, 1554 16 of 18

20 m. Therefore, in practice, one can look for geometry with sensors and targets nearly
perpendicular to each other to improve the positioning performance.

Figure 8. The relationship between spatial error sum and intersection angle φ, elevation angle η.

5. Conclusions
This paper studies the data association and signal source localization problems with
distributed passive sensors with angle-only observations. A geometry-based data associa-
tion method is considered, and the concept is that real targets will contribute observations
with a small minimal distance. The statistical distribution of the minimal distance of two
lines associated with the same target is formulated, based on which a data association
method based on hypothesis testing is also developed. The decision threshold is formu-
lated. Meanwhile, for observations that are classified into the same class, three positioning
algorithms are studied, namely the intersection method, the LS method and the WLS
method. Two kinds of measurements errors are considered, namely sensor self-positioning
error and angle measurement error.
In numerical results, we analyze the data association performance of the concerned
positioning algorithms and the signal source positioning performance in different scenarios,
indicating that the data association algorithm works well and the positioning performances of
the algorithms are very close to each other. Meanwhile, if the observation lines are approxi-
mately perpendicular to each other, then the localization performance is more accurate.
This algorithm can be used in laser, infrared, and other passive sensors with angle-
only measurements. In practice, other information, such as range, ground surface and sea
surface, may be available, which can be incorporated into the positioning and association
algorithms to improve performance.

Author Contributions: Conceptualization, S.Z.; formal analysis, Y.C. and L.W.; methodology, Y.C.,
L.W. and S.Z.; software, Y.C. and L.W.; supervision, R.C. and S.Z.; writing—original draft, Y.C. and
L.W.; revising, S.Z. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: This study did not report any data.
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2022, 22, 1554 17 of 18

Appendix A
∂r ∂r
In this section, we derive the partial derivative vectors ∂θi,m and ∂θj,m from (47) and
∂r ∂r
(48). The elements in vector ∂θi,m and ∂θj,m can be expressed by

 T
∂r ∂r ∂ei ∂r ∂r
= = − sin θi,m cos ϕi,m + cos θi,m cos ϕi,m (A1)
∂θi,m ∂ei ∂θi,m ∂ei,x ∂ei,y
 T
∂r ∂r ∂ei
= (A2)
∂ϕi,m ∂ei ∂ϕi,m
∂r ∂r ∂r
= − cos θi,m sin ϕi,m − cos θi,m sin ϕi,m + cos ϕi,m (A3)
∂ei,x ∂ei,y ∂ei,z
!T
∂r ∂r ∂e j ∂r ∂r
= = − sin θ j,m cos ϕ j,m + cos θ j,m cos ϕ j,m (A4)
∂θ j,m ∂e j ∂θ j,m ∂e j,x ∂e j,y
!T
∂r ∂r ∂e j
= (A5)
∂ϕ j,m ∂e j ∂ϕ j,m
∂r ∂r ∂r
= − cos θ j,m sin ϕ j,m − cos θ j,m sin ϕ j,m + cos ϕ j,m (A6)
∂e j,x ∂e j,y ∂e j,z

where
∂r 1 3
= X − 2 (e j,y zi,j − e j,z yi,j ) − X − 2 Y (e2j,y ei,x + e2j,z ei,x − ei,y e j,y e j,x − ei,z e j,z e j,x ) (A7)
∂ei,x
∂r 1 3
= X − 2 (e j,z xi,j − e j,x zi,j ) − X − 2 Y (e2j,z ei,y + e2j,x ei,y − ei,z e j,z e j,y − ei,x e j,x e j,y ) (A8)
∂ei,y
∂r 1 3
= X − 2 (e j,x yi,j − e j,y xi,j ) − X − 2 Y (e2j,y ei,z + e2j,x ei,z − ei,y e j,y e j,z − ei,x e j,x e j,z ) (A9)
∂ei,z
∂d 1 3
= X − 2 (ei,z yi,j − ei,y zi,j ) − X − 2 Y (ei,y
2 2
e j,x + ei,z e j,x − ei,y e j,y ei,x − ei,z e j,z ei,x ) (A10)
∂e j,x
∂d 1 3
= X − 2 (ei,x zi,j − ei,z xi,j ) − X − 2 Y (ei,z
2 2
e j,y + ei,x e j,y − ei,z e j,z ei,y − ei,x e j,x ei,y ) (A11)
∂e j,y
∂d 1 3
= X − 2 (ei,y xi,j − ei,x yi,j ) − X − 2 Y (ei,y
2 2
e j,z + ei,x e j,z − ei,y e j,y ei,z − ei,x e j,x ei,z ) (A12)
∂e j,z
xi,j yi,j zi,j
Y= ei,x ei,y ei,z (A13)
e j,x e j,y e j,z
2 2 2
ei,y ei,z ei,z ei,x ei,x ei,y
X= + + (A14)
e j,y e j,z e j,z e j,x e j,x e j,y

where [ xi,j , yi,j , zi,j ]T = s j − si .

References
1. Qiang, Z.; Jiaqi, C.; Rijie, Y.; Zhichao, S. Research on airborne infrared location technology based on orthogonal multi-station
angle measurement method. Infrared Phys. Technol. 2017, 86, 202.
2. Sioutis, M.; Tan, Y. User Indoor Location System with Passive Infrared Motion Sensors and Space Subdivision; Springer International
Publishing: Berlin/Heidelberg, Germany, 2014.
3. Cheng, X.; Huang, D.; Wei, H. High precision passive target localization based on airborne electro-optical payload. In Proceedings
of the 2015 14th International Conference on Optical Communications and Networks (ICOCN), Nanjing, China, 3–5 July 2015.
4. Bai, G.; Liu, J.; Song, Y.; Zuo, Y. Two-UAV Intersection Localization System Based on the Airborne Optoelectronic Platform.
Sensors 2017, 17, 98. [CrossRef] [PubMed]
Sensors 2022, 22, 1554 18 of 18

5. Peng, S.; Zhao, Q.; Ma, Y.; Jiang, J. Research on the Technology of Cooperative Dual-Station position Based on Passive
Radar System. In Proceedings of the 2020 3rd International Conference on Unmanned Systems (ICUS), Harbin, China, 27–28
November 2020.
6. Wang, Y.; Ho, V. An Asymptotically Efficient Estimator in Closed-Form for 3-D AOA Localization Using a Sensor Network. IEEE
Trans. Wirel. Commun. 2015, 14, 6524–6535. [CrossRef]
7. Li, M.; Lu, Y. Angle-of-arrival estimation for localization and communication in wireless networks. In Proceedings of the 2008
16th European Signal Processing Conference, Lausanne, Switzerland, 25–29 August 2008.
8. Stolnik, M. Radar Handbook, 3rd ed.; McGraw-Hill Professional Publishing: New York, NY, USA, 2008.
9. Kang, S.; Kim, T.; Chung, W. Multi-Target Localization Based on Unidentified Multiple RSS/AOA Measurements in Wireless
Sensor Networks. Sensors 2021, 21, 4455. [CrossRef] [PubMed]
10. Pattipati, K.R.; Deb, S. A new relaxation algorithm and passive sensor data association. IEEE Trans. Autom. Control 1992,
37, 198–213. [CrossRef]
11. Ouyang, C.; Ji, H. Modified cost function for passive sensor data association. Electron. Lett. 2011, 47, 383–385. [CrossRef]
12. Li, H.; Lu, C.; Feng, X.; Di, Z.; Wei, W. Data association algorithm for multi-infrared-sensor system. Infrared Laser Eng. 2014, 79,
511–517.
13. Wang, X.; Cai, W.; Qiu, L.; Yuan, A.; Cao, Z. Research on Data Association Algorithm Based on Line of sight Distance in
Photoelectric Two-dimensional Detection System. Air Space Def. 2019, 2, 51.
14. Yin, J.; Wan, Q.; Yang, S.; Ho, K.C. A Simple and Accurate TDOA-AOA Localization Method Using Two Stations. IEEE Signal
Process. Lett. 2015, 23, 144–148. [CrossRef]
15. Dogancay, K. On the bias of linear least squares algorithms for passive target localization. Signal Process. Amst. 2004, 84, 475–486.
[CrossRef]
16. Doanay, K. Bearings-only target localization using total least squares. Signal Process. 2005, 85, 1695–1710.
17. Dogancay, K. Relationship Between Geometric Translations and TLS Estimation Bias in Bearings-Only Target Localization. IEEE
Trans. Signal Process. 2008, 56, 1005–1017. [CrossRef]
18. Yu, Z.; Fu, Y. A passive location method based on virtual time reversal of cross antenna sensor array and Tikhonov regularized
TLS. IEEE Sens. J. 2021, 21, 21931–21940 [CrossRef]
19. Li, C.; Zhuang, W. Hybrid TDOA/AOA mobile user location for wideband CDMA cellular systems. IEEE Trans. Wirel. Commun.
2002, 1, 439–447.
20. Loan, G. An Analysis of the Total Least Squares Problem. SIAM J. Numer. Anal. 1980, 17, 883–893.
21. Wu, W.; Jiang, J.; Fan, X.; Zhou, Z. Performance analysis of passive location by two airborne platforms with angle-only
measurements in WGS-84. Infrared Laser Eng. 2015, 44, 654–661
22. Frew, E.W. Sensitivity of Cooperative Target Geolocalization to Orbit Coordination. J. Guid. Control Dyn. 2008, 31, 1028–1040
[CrossRef]
23. Yi, Z.; Li, Y.; Qi, G.; Sheng, A. Cooperative Target Localization and Tracking with Incomplete Measurements. Int. J. Distrib. Sens.
Netw. 2014, 2014, 1–16.
24. Malick, M. A Note on Bearing Measurement Model. 2018. Available online: https://www.researchgate.net/profile/Mahendra-
Mallick/publication/325214760_A_Note_on_Bearing_Measurement_Model/links/5afe2d230f7e9b98e0197b3f/A-Note-on-
Bearing-Measurement-Model.pdf (accessed on 8 January 2022).
25. Mallick, M.; Nagaraju, R.M.; Duan, Z. IMM-CKF for a Highly Maneuvering Target Using Converted Measurements. In
Proceedings of the 2021 International Conference on Control, Automation and Information Sciences (ICCAIS), Xi’an, China, 14–17
October 2021; pp. 15–20. [CrossRef]
26. Casella, G.; Berger, R.L. Statistical Inference, 2nd ed.; Duxbury Thomson Learning Press: Pacific Grove, CA, USA, 2002.