You are on page 1of 17

I.

INTRODUCTION

For the positioning purpose, tracking the time-delay or


Array-Based GNSS Signal the time-of-arrival (TOA) of a navigation satellite signal
has been challenging in global navigation satellite systems
Tracking With a Reduced State (GNSS). Conventional time-delay tracking methods,
embedded into commercial GNSS receivers, rely on
Signal Model feedback loop architectures such as delay-locked loop
(DLL) or early-minus-late (EML) correlator [1]. However,
it is well known that a classical 1-chip wide DLL can
introduce a bias up to hundreds of meters in the presence
of multipath [2, 3]. A large amount of effort has been
SANGWOO LEE made to enhance the DLL-based time-delay tracking
Hanyang University accuracy; for example, narrow EML [2], double-delta
Seoul, Republic of Korea () correlator [4], early1/early2 (E1/E2) tracker [5],
ELENA SIMONA LOHAN, Senior Member, IEEE early-late-slope (ELS) [6], improved ELS [7], multipath
Tampere University of Technology estimating delay lock loop (MEDLL) [8], and multiple
Tampere, Finland gate delay (MGD) [9]. More of the different DLL-based
SUNWOO KIM, Member, IEEE algorithms can be found in [10, 11].
Hanyang University Recently, applications of recursive Bayesian filters
Seoul, Republic of Korea
have been addressed for high-precision GNSS signal
tracking under multipath environments [3, 12–15]. In
particular, the use of particle filters [16] has been
This paper introduces a reduced state signal model in which
emphasized to provide an outstanding performance in
interfering signals are suppressed with the help of beamforming. nonlinear and non-Gaussian settings. Considering a single
Based on our model, a particle filtering algorithm integrated with antenna GNSS receiver, Closas et al. proposed several
beamformer for satellite navigation signal tracking is proposed to particle filtering approaches [13, 14] that track the signal
increase the computational efficiency and the robustness to the signal parameters (i.e. amplitudes and TOAs) of the line-of-sight
model mismatch from unknown interference. Through extensive
simulations and flop complexity analysis, we verify the superiority
(LOS) signal and its multipath replicas. In [3], a
of the proposed algorithm against existing Bayesian filtering Rao-Blackwellized particle filter was applied to reduce the
algorithms without beamforming. computational burden of the particle filtering algorithm.
A recent work [15] by Liu et al. presented a particle
filter-based tracking algorithm for an array antenna GNSS
receiver. However, this algorithm was developed under the
assumption that the direction-of-arrivals (DOAs) for the
LOS and multipath signals are perfectly known, which is
not realistic. Other applications of particle filters for
positioning and multipath mitigation can be found in
[17–19]. In practice, the GNSS signal can be also
Manuscript received January 13, 2015; revised July 24, 2015, October adversely affected by intentional interference (e.g.
22, 2015; released for publication November 24, 2015. jamming and spoofing) and/or external (unintentional)
interference from spurious emissions of terrestrial
DOI. No. 10.1109/TAES.2016.150024.
transmitters, in addition to multipath [20–22]. In [23], a
Refereeing of this contribution was handled by A. Dempster. mean-value theorem particle filter-based algorithm [24]
This work has been supported by the National Research Foundation was introduced to track the parameters of the LOS and
of Korea (NRF) grant funded by the Korea government jamming signals.
(NRF-2013R1A1A2062728) and by the National GNSS Research Center Although these particle filtering approaches enhanced
Program of Defense Acquisition Program Administration and Agency the time-delay tracking accuracy in multipath or
for Defence Development.
interference environments, these approaches may fail
Authors’ addresses: S. Lee, Hanyang University, Electronics and when multipath and interference signals coexist due to the
Computer Engineering, ITBT #822, 222 Wangsimni-ro, Seongdong-gu, model mismatch (error), defined as a difference between
Seoul, 133-791 Republic of Korea; E. S. Lohan, Tampere University
of Technology, Department of Electronics and Communications the true received signal and the presumed received signal
Engineering, Korkeakoulunkatu 1, Tampere, 33720 Finland; S. Kim, model. In the Bayesian filtering theory, the model
Hanyang University, Electronics and Computer Engineering, ITBT #817, mismatch is known to cause the tracking error to
222 Wangsimni-ro, Seongdong-gu, Seoul, 133-791 Republic of Korea. monotonically increase with time [25]. It seems that the
Corresponding author is S. Kim, E-mail: (remero@hanyang.ac.kr). model mismatch can be simply handled by using an
appropriate signal model, which takes into account both
0018-9251/16/$26.00 
C 2016 IEEE multipath and interference signals, and by tracking all

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016 1267
incoming signals. However, the risk of the model the use of array antennas is increasing for many military
mismatch still remains high because of unexpected operations such as autonomous navigation of unmanned
changes of multipath and interference (e.g. sudden aerial/ground vehicles, missile launches, and signal
appearance/disappearance). Moreover, tracking all intelligence in combat areas, our approach will facilitate
incoming signals is computationally inefficient since the successful operations also in jamming attacks.
complexity increases with the state space dimension [26]. The remainder of the paper is organized as follows.
The computational complexity and the model Section II describes the signal models considered in this
mismatch are important considerations for the real paper. Section III reviews the conventional particle
implementation of a Bayesian filtering algorithm. It has filtering approach to multipath and interference
been reported that the computational efficiency of the mitigation. Section IV presents a reduced state signal
Bayesian filters can be enhanced by using reduced state model with beamforming and introduces the proposed
models in which marginal states with little contributions to particle filter-based tracking algorithm combined with a
the system response are eliminated [27, 28]. However, the novel robust hybrid beamformer. Section V discusses
state reduction increases the risk of the model mismatch in simulation results in comparison with the proposed
the GNSS scenarios where the interfering signals algorithm and existing Bayesian filtering algorithms that
frequently remain dominant over the LOS signal. In [3], track the LOS and multipath signals without
the complexity issue was tackled for a single antenna beamforming. The paper is concluded in Section VI.
receiver from the particle filter perspective, and the Notation: Throughout the paper, vectors and matrices
robustness of a particle filtering algorithm to the model are denoted by lowercase and uppercase bold letters. The
mismatch from sudden appearance/disappearance of real and complex fields are indicated by R and C. IK is the
multipath was merely examined through computer identity matrix of size K × K. 0K is the zero vector of size
simulations. However, no investigation on the complexity K. The operator diag{a} with a ∈ CK results in a K × K
reduction and the model mismatch avoidance under matrix in which the diagonal entries correspond to the
multipath and interference environments has been reported elements of a and the others are all zero. The function D(·)
so far to the best of our knowledge. returns the dimension of a vector. The superscripts T, H,
Recent studies [29–31] have shown that array and ∗ denote the transpose, Hermitian transpose, and
antenna-based beamforming techniques can effectively conjugate operations, respectively. The superscript ◦
suppress the interfering signals in GNSS scenarios. represents an angle in degree. R(·) and I (·) provide the
Motivated by this fact, in this paper, we present a reduced real and imaginary parts of a complex number. The
state signal model where the received signal is defined Euclidean norm, the Frobenius norm, and the Kronecker
with only the LOS signal, while interfering signals are product are denoted by  · ,  · F , and ⊗. E[·] represents
suppressed and approximated as part of the thermal noise the expectation of the argument. N (a, C) and CN (a, C)
by the array antenna-based beamforming. The reduced are Gaussian and complex Gaussian densities with mean a
state signal model can enhance the computational and covariance C.
efficiency by greatly reducing both the state dimension
and the measurement size. However, the model mismatch II. SIGNAL MODEL
can arise when beamforming fails to reject the interfering We consider that an array of M antennas receives the
signals. To track the LOS signal with the effective direct sequence spread spectrum (DSSS) signal from a
avoidance of the model mismatch, we develop a particle given satellite with Nm multipath replicas and other
filtering algorithm integrated with a robust hybrid interferences. Using a different pseudorandom noise code
beamformer where the uncertainty bounds in the state per satellite, the DSSS signals from other satellites are
estimates are measured according to divergence among the considered as Gaussian noise included in the thermal
particles. noise. It is worth noting that the number of multipath
In our previous work in [32], we developed a particle signals can be different for each satellite due to their
filtering algorithm with a modified version of the deployments, but the number of interferences is assumed
minimum variance (Capon) beamformer [33, 34] for to be the same for all the satellites.
tracking the LOS and multipath signals. However, the At the time t, the continuous received signal y(t) at the
issues of the complexity reduction and the model front-end of the receiver in the presence of multipath and
mismatch avoidance were not investigated in our interference can be modeled as
preliminary study in [32]. The work presented in this

Nm
paper extends our previous work to address the above y(t) = hα (t) + g(t) + u(t) (1)
challenges. The reduced state signal model is derived and α=0
validated. An efficient tracking algorithm is proposed that
where α = 0 represents the LOS signal and α > 0 denotes
performs joint signal tracking and robust hybrid
its multipath signal. The α-th DSSS signal is denoted by
beamforming in the framework of a single filter. This
paper also provides an extended analysis and evaluation hα (t) = Aα (t)q(t − τα (t)) cos(2πfc t + φα (t))a( α (t))
of the Bayesian filtering algorithms under a realistic (2)
measurement-based satellite propagation model [35]. As where φα (t) = 2πfα (t)t + ϕα (t),

1268 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
where hα ,k,n , gk,n , and uk,n are the down-converted
and discretized versions of hα (t), g(t), and u(t) at
t = (n – 1)KTs + kTs . Defining
 
Aα,n = Aα (nKTs ), τα,n = τα (nKTs ),
 
θα,n = θα (nKTs ), φα,n = φα (nKTs ), (6)

qk (τα,n ) = q((n − 1)KTs + kTs − τα,n ),

hα ,k,n can be defined as hk (α ,n ), which is a function of the


state vector α,n = [τα,n , θα,n , R(ψα,n ), I (ψα,n )]T ∈ R4
where ψα,n = Aα,n ej φα,n represents the complex amplitude
of the signal1 [3, 29]:
Fig. 1. Total of K samples are collected in one observation with sample hα,k,n = Aα,n ej φα,n qk (τα,n )a(θα,n )
period of Ts .
= ψα,n qk (τα,n )a(θα,n )

1) Aα (t), τα (t), fα (t), ϕα (t), and α (t)=[θα (t), ξα (t)]T = hk (α,n ). (7)
are the amplitude, TOA, Doppler frequency, carrier phase, We note that in this model, the Doppler frequency and the
and DOA [elevation angle θ α (t) and azimuth angle ξ α (t)] carrier phase are incorporated into the complex amplitude.
of the α-th signal; Hence, the Doppler shift and the carrier phase are also
2) q(t) is the DSSS signal (composed of data symbols trackable in this model. Specifically, they can be estimated
and spreading code [3, 13, 14, 29]) from a given satellite; from the estimated complex amplitude in the sense of the
and, least squares problem (see details in [3]). The n-th
3) fc is the carrier frequency of the GNSS signal. observation block yn ∈ CKM of the received baseband
The vector a( α (t)) ∈ CM represents the steering vector signal samples is constructed by stacking K array sample
for the α-th signal with the DOA  α (t) and without any vectors:
mismatches:
yn = [yT1,n , yT2,n , . . . , yTK,n ]T
a( α (t)) = [a1 ( α (t)), a2 ( α (t)), . . . , aM ( α (t))] T
(3)

Nm
where = h(0,n ) + h(α,n ) + gn + un (8)


2π α=1
am ( α (t)) = e−j λ s( α (t)) pm ,
T LOS

⎡ ⎤ multipath & interference
sin θα (t) cos ξα (t)
⎢ ⎥ where
s( α (t)) = ⎣ sin θα (t) sin ξα (t) ⎦ , (4)
cos θα (t) h(α,n ) = [hT1 (α,n ), hT2 (α,n ), . . . , hTK (α,n )]T ,
pm ∈ R3 is the position of the √ m-th element in a gn = [gT1,n , gT2,n , . . . , gTK,n ]T , (9)
three-dimensional space, j = −1, and λ is the carrier
wavelength [36]. In this paper, a uniform linear array un = [uT1,n , uT2,n , . . . , uTK,n ]T .
(ULA) receiver where the elements are placed along the The covariance matrix of un is denoted by Cun = E[un uH n ]
z-axis with separation d is assumed for simplicity, ∈ CKM×KM , which is known to the receiver.
that is to say, pm = [0, 0, (m – 1)d]T . In this case, The DSSS signal parameters {α,n }N m
−1 α=0 are modeled
am ( α (t)) = e−j 2π λ (m−1)d cos θα (t) . Hence, in this paper, by time-varying processes. It is assumed that the initial

we assume that  α (t) = θα (t) and a(θα (t)) = a( α (t)). state α ,0 of each signal is drawn from N ( ¯ α,0 , Cα,0 ). At
The vector g(t) represents the mixture of unknown the receiver, ¯ α,0 can be obtained through acquisition
interferences except the multipaths, and u(t) is the methods (see [37, 38] and references therein), which are
zero-mean complex white Gaussian noise. out of the scope of this paper, and Cα ,0 is determined
After down-conversion, the receiver records K according to the uncertainties in  ¯ α,0 . The dynamics of
snapshots with the sampling period of Ts , as depicted in each state can be approximated by a first-order
Fig. 1. It is assumed that the channel remains stationary autoregressive model [3, 39, 40]:
during one observation interval KTs . Then, the k-th array
sample vector yk,n ∈ CM of the received baseband signal α,n =  α,n α,n−1 + vα,n , (10)
during the n-th observation period is written as

Nm 1 When using different array configurations, the steering vector is defined
yk,n = hα,k,n + gk,n + uk,n (5) with both elevation and azimuth angles of a signal. In this case, the state
α=0 vector α ,n should be extended to include the azimuth angle.

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1269
The transitional matrix  α,n ∈ R4×4 is given by posterior probability density function p(xn |y1:n ) given
all available measurements up to time nKTs , y1:n =
 α,n = diag{γα,τn , γα,θn , γα,Rn , γα,In }, (11) {y1 , y2 , . . . , yn } [16]. Particle filters characterize the
where γα,(·) is the dynamic parameter for the posterior density of the state with a set of S random
corresponding signal parameter. The zero-mean white samples (particles) taken from an importance density
Gaussian noise vector is denoted by vα ,n with covariance function with associated weights, which is given by
Cα,vn = E[vα,n vTα,n ] = diag{σα,τ 2 2
, σα,θ 2
, σα, Rn , σα,In },
2 
S
n n
(12) p(xn |y1:n ) ≈ ωns δ(xn − xsn ) (13)
2 s=1
where σα,(·) is the process noise variance for the
corresponding parameter. The matrices { α,n , Cα,vn }N m where δ(·) is the Dirac delta function and the s-th particle
α=0
are design parameters and should be adjusted properly is denoted by
according to satellite and receiver dynamics [3].  α ,n xsn = [(s0,n )T , (s1,n )T , . . . , (sNm ,n )T , (sn )T ]T (14)
controls the average rate of change of each state and
vα,n ∼ N (04 , Cα,vn ) provides additional randomness on with its weight ωns .
the change. In the related literatures [3, 13–15], it is The initial particles {xs0 }Ss=1 are drawn from
generally assumed that  α ,n = μI4 where μ ≈ 1. This p(x0 ) = N (x̄0 , C0 ) where x̄0 is the initial estimate of x0 ,
assumption is valid for both stationary and high-speed obtained in the acquisition phase, and C0 is determined by
receivers as far as the values in Cα ,vn are determined to be the uncertainties in x̄0 .
proportional to the amounts of uncertainties in the receiver The weights for the initial particles are uniform, that
dynamics. In fact, the average rate of change of each state is, ωns = 1/S. When the prior density p(x0 ) is unavailable,
(i.e.  α ,n ) can be theoretically calculated with the particles can be initialized with a uniform distribution
receiver’s relative speed, acceleration, and angle towards over an uncertainty region of the state. Without loss of
the satellite and the satellite’s position. More details on the generality, the dynamic model for α ,n in (10) is extended
computation of the change of the state with respect to the for xn . Let Dn and Cv n represent the transitional matrix for
geometry of the receiver and satellite are provided in [41]. xn and the noise covariance matrix, respectively, and they
Also, the design parameters can be computed in a heuristic are block diagonal matrices. Based on this, each particle
manner through prior experiments. Due to the simplicity, propagates as
this simple model has been used in several related xsn ∼ N (Dn xsn−1 , Cvn ). (15)
literatures [3, 12–15, 23, 40].
Assuming that the transitional prior p(xn |xsn−1 ),
which
III. CONVENTIONAL PARTICLE FILTERING follows a Gaussian distribution, is used as the importance
APPROACH TO MULTIPATH/INTERFERENCE density function, the weight of each particle is updated to
MITIGATION: SIMULTANEOUS TRACKING OF be proportional to the likelihood p(yn |xsn ) [16]:
LOS AND UNDESIRED SIGNALS
ωns ∝ ωn−1
s
p(yn |xsn ) (16)
This section describes the conventional particle
filtering approach to multipath and interference mitigation, where p(yn |xsn )
is given by
which tracks the entire state of all incoming signals. For N 
 m

ease of explanation, let us suppose that the receiver also p(yn |xn ) = CN
s
h(α,n ) + g(n ), Cun .
s s
(17)
has a prior information on the interferences (i.e. the signal α=0
numbers, types, and dynamics) in addition to that of the Then,
LOS and multipath signals, and denote the state vector  s the updated weight is normalized such that
s ωn = 1. Following the minimum mean square error
defining the interferences at the n-th observation by n .2 (MMSE) criterion, the parameters of all the received
Then, we can define the interference term as a function of signals at the n-th observation are estimated as

n , gn = g(n ), as well as the GNSS signal. We denote
the entire state vector at the n-th observation by 
S
x̂n = ωns xsn . (18)
xn = [T0,n , T1,n , . . . , TNm ,n , Tn ]T ∈ R4(Nm +1)+D(n ) .
s=1
From the Bayesian filtering perspective, the GNSS
signal tracking problem in the presence of multipath and In the GNSS scenarios, it is assumed that K M.
interference is defined as the problem of estimating the In this case, the likelihood computation is the most
overall state xn at the n-th observation by constructing the computational intensive task, which requires
O(K 3 M 3 + 32 K 2 M 2 S) flops [42].
When the information not only on the LOS signal, but
2 The element and dimension of the state vector  for the interference
n also on the multipath and interference signals is available,
are different according to the signal numbers and types. For a single-tone the GNSS signal tracking problem can be solved by
continuous wave (CW) interference,n contained the DOA and complex
amplitude (including the amplitude, frequency offset, and initial phase) using the conventional approach. However, the signal
of the interference. For a multitone CW interference, the state vector for parameters of the multipath and interference signals are
each tone should be included into n . nuisance parameters [43], which are not our concern in the

1270 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
be inversely proportional to the difference between the
beamformer outputs of the received signal and the
predicted LOS signal, denoted as the measurement and the
reduced state signal in Fig. 2. The parameters of the LOS
signal are then estimated with the weighted sum of all the
particles following the MMSE criterion. When a new
received signal vector is available, the particles are
updated based on the dynamic model, and the same
process described above is repeated.

A. Reduced State Signal Model with Robust Hybrid


Beamformer and Particle Filter Integration
Since the LOS signal has a prime importance in the
GNSS applications, our concern is to estimate 0,n by
approximating the posterior probability function of 0,n
from all available measurements y1:n . In the ideal case
where perfect information on multipath and interference
signals is available, the posterior probability function of
Fig. 2. Proposed algorithm incorporates robust hybrid beamformer and
0,n can be approximated by eliminating the multipath
particle filter in single framework to track LOS signal based on reduced
state signal model with avoidance of model mismatch. and interference signals from the measurements. However,
it is practically impossible to obtain accurate knowledge
of the unwanted signals.
GNSS applications. The nuisance parameter estimation
Adaptive beamforming techniques [29, 33, 44–46]
increases the state space dimension, and consequently,
possess the advantage of neglecting the unwanted signals.
increases the computational complexity and the required
Assume the unwanted signals are successfully suppressed
number of particles in order to achieve a given tracking
and considered as part of the thermal noise with the aid of
accuracy for the LOS signal [26]. The model mismatch
beamforming. Then, we can yield the output of the
may be caused from the inaccurate knowledge on
beamformer zn ∈ CK as
multipath and interference signals, which often change
rapidly and unpredictably in practice. In the following zn = W H
n yn
subsection, a novel integrated robust hybrid beamforming 

Nm
and particle filtering approach based on a reduced state = WH
n h(0,n ) + h(α,n ) + gn + un
signal model is introduced to overcome the drawbacks of α=1
the conventional approach. ≈ WH
n h(0,n ) + ũn (19)
IV. AN INTEGRATED ROBUST HYBRID where Wn = IK ⊗ wn ∈ CKM×K , wn ∈ CM is the
BEAMFORMING AND PARTICLE FILTERING beamforming vector, and ũn = WH n un is the zero-mean
APPROACH FOR STATE SPACE REDUCTION AND complex Gaussian noise with covariance Cũn = E[ũn ũH n ]
MODEL MISMATCH AVOIDANCE = WH n Cun Wn ∈ C
K×K
. It is worth noting that particle
This section presents an integrated robust hybrid filters have the robustness to the small model mismatch3
beamforming and particle filtering algorithm based on a [3]. Now, the received signal is defined in terms of only
reduced state signal model to increase the computational the components of the LOS signal, and the state of our
efficiency and to avoid the model mismatch in the LOS interest can be reduced from xn to 0,n . However, if
signal tracking under multipath and interference. The beamforming fails to reject the unwanted signals, a
framework of the proposed algorithm is illustrated in considerably large mismatch is observed and the
Fig. 2. The notations used in the figure are defined later. In approximation becomes invalid. Consequently, any
the proposed algorithm, once random particles containing Bayesian filtering algorithms fail to track the LOS signal.
possible solutions to the parameters of the LOS signal are In this paper, a novel robust hybrid beamformer is
generated, they are used to predict the LOS signal in the presented to avoid the model mismatch. The reduced state
received signal. According to the divergence of the signal model (19) with the proposed beamformer is
particles, the uncertainty bounds of the parameters validated through experiments in Section V.
estimates based on the previous estimates are computed. Following (19), the state in the received signal is
The beamforming vector is derived subject to the reduced as 0,n from xn , and the GNSS signal tracking
uncertainty bounds to have the robustness to the errors in
the parameter estimates and to avoid the model mismatch. 3 Each particle represents a candidate solution for the interest, and the
The beamforming vector is then multiplied with the estimate is obtained within the region where the particles are deployed.
received signal and the predicted LOS signals according to Thus, the error from the model mismatch is limited according to the
the particles. The weight of each particle is determined to particle distribution.

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1271
problem is subject to tracking only the LOS signal given the proposed algorithm, the degeneracy problem can
the beamformer output set z1:n = {z1 , z2 , . . ., zn }. Hence, reoccur with the use of beamforming. The recurrence of
the objective of the proposed algorithm is to estimate the the degeneracy problem may be more frequent with a
posterior probability function of 0,n given z1:n : larger array aperture. More specifically, the beamwidth
becomes narrower as the entire array aperture increases

S
p(0,n |z1:n ) ≈ ωns δ(0,n − s0,n ) (20) [48], and many particles can be worthless even if there is a
s=1
small gap between the DOA in each particle and the beam
direction.
with the s-th particle s0,n = [τ0,n
s s
, θ0,n , R(ψ0,n
s
), I (ψ0,n
s
)]T To overcome the sampling impoverishment problem
s
and the corresponding weight ωn . The procedures are the and the recurrence of the degeneracy problem, the Markov
same as those described in Section III, but the particle chain Monte Carlo (MCMC) move step is applied, which
propagation and the likelihood computation are different is based on the Metropolis-Hastings algorithm [49]. The
according to the states: MCMC move step improves diversity among particles by
s0,n ∼ N ( 0,n s0,n−1 , C0,vn ), (21) introducing an additional random noise (jitter) into a
resampled particle ˜ s0,n with the acceptance probability
  αn:
p(zn |s0,n ) = CN WH s
n h(0,n ), Cũn . (22)  
p(zn |˜ s )p( ˜ s |s
At each iteration, the beamforming matrix Wn is 0,n 0,n 0,n−1 )
αn = min 1, (24)
determined by using the set of particles {s0,n , ωns }Ss=1 and p(zn |sn )p(s0,n |s0,n−1 )
the parameter estimates based on the previous estimates
where the function min(·) gives the smallest value of the
ˆ 0,n−1 , and the details on the derivation of the
argument. The detail procedure for the MCMC move step
beamforming matrix are discussed in the following
resampling can be found in [16]. The overall procedure of
subsection. Based on the MMSE criterion, the signal
the proposed algorithm is summarized in Algorithm 1.
parameters of the LOS signal at the n-th observation can
be estimated as the conditional mean of 0,n , and we
Algorithm 1 LOS signal tracking using the integrated robust hybrid
denote the estimate by  ˆ 0,n .
beamformer and particle filter-based algorithm
We recall that the likelihood computation is the most
computationally intensive part for large K. Using the Initialization
signal measurement vector of size K in the proposed 1: set n ← 0
2: for All s do
particle filtering algorithm, O(K 3 + (2M 2 + 2M+ 32 S)K 2 )
3: generate s0,0 ∼ N ( ¯ 0,0 , C0,0 )
flops are required for the likelihood computation. More 4: assign ω0s = 1/S
details on the computational complexity analysis are 5: end for
provided in Section V. 6: ˆ 0,0 =  ¯ 0,0
It is well known that the particle filters suffer from the Particle propagation & Weight update
degeneracy problem [16] where all but one particle 7: n ← n + 1
8: compute εd,n and εa,n in (35) and (36)
become worthless as their weights converge to nearly zero. 9: derive wn using  ˆ 0,n−1 according to algorithm 2
In other words, a small portion of particles contribute to 10: for all s do
the approximation of the posterior probability function. A 11: propagate s0,n ∼ N ( 0,n s0,n−1 , C0,vn )
severe degeneracy is observed when the effective sample 12: update ωns = ωn−1s
p(zn |s0,n )
size Neff is small. The effective sample size at the n-th 13: end for 
14: normalize ωns such that s ωns = 1
observation is estimated as
Resampling

1 15: calculate N̂eff,n = 1/ s (ωns )2
N̂eff,n = S . (23)
s 2 16: if N̂eff,n < Nth then
s=1 (ωn )
17: resample {s0,n }Ss=1 using MCMC move step
The degeneracy problem can be solved via resampling, 18: assign ωns = 1/S
which eliminates particles with low weights and clones 19: end if
particles with high weights. The resampling is executed Estimation

20: estimate  ˆ 0,n = S ωs s
when N̂eff,n < Nth where Nth represents a preset threshold. s=1 n 0,n
21: go to Particle propagation & Weight update phase and repeat
In general, it is assumed that Nth = 2S/3 in the literature if new yn is available
[3, 15, 47].
Duplicating the particles with high weights in the
resampling phase, the resultant particles contain many
B. Robust Hybrid Beamformer for Signal Model
repeated points, and this brings a loss of diversity among
the particles. This is referred to as the sampling Mismatch Avoidance from Reduced State Signal
Model
impoverishment problem [16]. The sampling
impoverishment problem becomes severe in the GNSS This subsection presents a novel hybrid beamformer to
scenarios where the process noise vα ,n is considered very avoid the model mismatch from the state space reduction.
small (see simulation parameters in [3, 15]). Moreover, in Hence, the approximation in (19) holds. For ease of

1272 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
explanation, let d0,n ∈ CK and a0,n ∈ CM define the LOS particles. In this sense, we approximate the bounds εd,n
signal vector and the steering vector for the LOS signal: and ε a,n with the set of particles {s0,n , ωns }Ss=1 ,

d0,n = d(ψ0,n , τ0,n ) 
S
εd,n = ωns d(ψ0,n
s s
, τ0,n ) − d̂0,n )2 , (35)
= ψ0,n [q1 (τ0,n ), q2 (τ0,n ), . . . , qK (τ0,n )]T , (25) s=1


S

a0,n = a(θ0,n ). (26) εa,n = ωns a(θ0,n
s
) − â0,n 2 . (36)
s=1
Our goal is to minimize the mean squared error (MSE)
between the beamformer output WH According to (35) and (36), the upper bounds depend on
n yn and the LOS
the divergence among the particles. Hence, when there are
n a0,n = 1
signal vector d0,n subject to the constraint wH
[29]. Recall that Wn = IK ⊗ wn . Hence, the hybrid large uncertainties in the state dynamics (i.e. large
2 2
beamforming problem is formulated as σ0,τ n
, σ0,θn
, σ0,2 Rn , and σ0,2 In ), larger bounds are obtained.
 2 For small uncertainties in the state dynamics, the bounds
min MSEn = K1 WH n yn − d0,n  can be very small, and the beamformer strongly relies on
wn (27) d̂0,n and â0,n , which can contain large errors. In this case,
subjectto wH n a0,n = 1. the bounds can be improper, but this can be fixed in the
Ignoring the constant, the objective function in (27) can be proposed algorithm with the MCMC move step-based
written as resampling that improves diversity among the particles by
 H  imposing a random noise. In this regard, the adaptive
W yn − d0,n 2
n computations of the upper bounds by using the particles
∗ ∗
= wH
n Rn wn − wn Yn d0,n − d0,n Yn wn + d0,n d0,n (28)
H T H T
can provide reliable results.
Under the uncertainty conditions, the problem (27) is
where Rn = Yn YHn ∈C
M×M
and the received signal reformulated as the robust hybrid beamforming problem:
matrix Yn ∈ CM×K
is given by  H 
  min W yn − d0,n 2
n
Yn = y1,n , y2,n , . . . , yK,n . (29) wn ,d0,n ,a0,n

The true LOS signal vector d0,n and the true steering n a0,n = 1,
subject to wH (37)
vector a0,n are unknown, but they can be substituted d0,n − d̂0,n  ≤ εd,n ,
2

with their estimates d̂0,n = d(ψ̃0,n , τ̃0,n ) and a0,n − â0,n 2 ≤ εa,n .
â0,n = a(θ̃0,n )where
The problem (37) is a quadratically constrained quadratic
ψ̃0,n = γ0,Rn R(ψ̂0,n−1 ) + j γ0,In I (ψ̂0,n−1 ), programming problem, which can be solved by using the
(30)
τ̃0,n = γ0,τn τ̂0,n−1 , θ̃0,n = γ0,θn θ̂0,n−1 , Lagrange multiplier method [50]:
∗ ∗
Ln = wH
n Rn wn − wn Yn d0,n − d0,n Yn wn + d0,n d0,n
H T H T
ψ̂0,n−1 , τ̂0,n−1 , and θ̂0,n−1 are the LOS signal parameter
estimates at the (n –1)-th observation. + λ1 (wH a0,n − 1)
Due to the estimation errors in d̂0,n and â0,n , d0,n and  n H

+ λ2 (d0,n − d̂0,n ) (d0,n − d̂0,n ) − εd,n
a0,n can be expressed as
 
d0,n = d̂0,n + d,n , (31) + λ3 (a0,n − â0,n )H (a0,n − â0,n ) − εa,n (38)
where λ1 = 0, λ2 ≥ 0, and λ3 ≥ 0 are the Lagrange
a0,n = â0,n + a,n , (32) multipliers. Let w̄n , d̄0,n , and ā0,n be wn , d0,n , and a0,n
satisfying (37). They can be simply obtained with partial
where d,n ∈ CK and a,n ∈ CM are the uncertainties in differentiations of (38) with respect to wn , d0,n and a0,n :
d̂0,n and â0,n , respectively. Let us assume that the
w̄n = R−1 ∗
n (Yn d0,n − λ1 a0,n ), (39)
uncertainties are upper bounded by
 2  2 λ2  
 
d0,n − d̂0,n  = d,n  ≤ εd,n , (33) d̄0,n = YTn w∗n + d̂0,n − YTn w∗n , (40)
λ2 + 1
    λ1
a0,n − â0,n 2 = a,n 2 ≤ εa,n . (34) ā0,n = â0,n −
wn (41)
λ3
The concept of upper bound conditions is used in many where the Lagrange multipliers are given by
robust beamforming problems under the assumption that −1 −1 H −1 ∗
the bounds are priorly known [33, 44, 46]. However, since λ1 = (aH
0,n Rn a0,n ) (a0,n Rn Yn d0,n − 1), (42)
the predefined bounds may be too small or too large, the  ⎛  ⎞
 
derived solutions can be inaccurate. The particles define d̂0,n − YTn w∗n 
the feasible regions of the parameters of the LOS signal, λ2 = max ⎝0, √ − 1⎠ , (43)
and the weights are proportional to the likelihoods of the εd,n

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1273
|λ1 | of the proposed algorithm is compared with that of the
λ3 = √ wn  , (44)
εa,n extended Kalman filtering (EKF) and particle filtering
(PF) algorithms [15], referred to as the EKFw/o BF and
and the function max(·) returns the largest value. Details
PFw/o BF algorithms in this paper, that track the states of
on the derivation of (39)–(44) are provided in the
the LOS and multipath signals without beamforming.4
Appendix.
Since we consider a jamming scenario in simulations, the
By substituting (40) and (41) into (39), the
EKFw/o BF and PFw/o BF algorithms that do not track the
beamforming vector can be obtained. However, d̄0,n and
states for interference experience the model mismatch. The
ā0,n are functions in terms of wn , which is unknown. Thus,
proposed algorithm may also have the model mismatch
w̄n , d̄0,n , and ā0,n are derived with the iterative procedure.
problem if multipath and interference signals are not
The pseudocode for the robust hybrid beamformer is
sufficiently suppressed with beamforming. We note that it
described in Algorithm 2 where the superscript κ denotes
is not possible to apply the reduced state signal model
the number of iterations. Note that the proposed robust
with the proposed robust hybrid beamformer to the EKF
hybrid beamformer is combined with the particle filtering
algorithm. More specifically, in the EKF context, the error
algorithm as depicted in Fig. 2. More specifically, the
bounds (35) and (36) correspond to the second central
previous output  ˆ 0,n−1 of the particle filtering algorithm
moments of functions of Gaussian random variables,
and the current particles {s0,n , ωns }Ss=1 are used to set the
which cannot be simply computed. The error bounds can
initial inputs and to adaptively compute the bounds of the
be computed in the unscented [51] and sigma-point [52]
uncertainties in d̂0,n and â0,n in the beamformer. In the
Kalman filters, which are modified versions of the EKF,
particle filtering algorithm, zn is obtained with the derived
but applications of such filters for further complexity
beamforming vector w̄n , and the weight of each particle is
reduction lie beyond the scope of this paper. DLL-based
updated with the signal measurement zn .
algorithms are not considered for comparison because
For K M, the computational complexity of the
they do not deal with the nonmultipath interference. As
robust hybrid beamformer is bounded by O((2M2 + 6Mρ
the performance of the Bayesian filtering algorithms in
+ 4S + 4ρ)K + 4MS) flops where ρ is the number of
multipath environments has been investigated in [3, 15],
iterations for convergence (see Table III in Section V). The
our analysis is more focused on the effect of interference
convergence of the proposed iterative robust hybrid
on tracking accuracy of the Bayesian filtering algorithms.
beamformer is proved with extensive simulations, but an
A GPS L1 C/A signal is used for simulations, but other
analytical proof (if possible) is not available.
GNSS signals can be also processed. We consider ULA
receivers of M = 5 and M = 10 with half-wavelength
Algorithm 2 Iterative robust hybrid beamformer for model mismatch
avoidance
spacing (d = λ/2) for simulations. The sampling period is
set as Ts = Tc /3 where Tc = 1/(1.023 MHz). The number
Require: Yn , Rn , εd,n , εa,n , ψ̃0,n , τ̃0,n , θ̃0,n of samples is K = 3069 in one observation. The incoming
Ensure: w̄n , d̄0,n , ā0,n signal is filtered by the low pass filter with 2 MHz
1: set κ ← 0
κ) κ) κ) bandwidth. A carrier-to-noise density ratio (CNR) is
2: wn ← 0M , d̂0,n ← d(ψ̃0,n , τ̃0,n ), â0,n ← a(θ̃0,n )
assumed to be 38 dB-Hz. Considering the correlation
3: repeat
4: κ ←κ + 1 between uk,n and ui,n for k = i from oversampling, the
κ)
κ−1) H
(â0,n ) R−1
κ−1) ∗
n Yn (d̂0,n ) −1
noise covariance Cun is modeled for simulations as
5: λ1 = κ−1) H follows:
(â0,n ) R−1
κ−1)
n â0,n
wn = R−1
κ)
(Yn (d̂0,n )∗ − λ1 â0,n )
κ−1) κ−1) ⎡ ⎤
6: n 
κ−1) κ) ∗
1 ζ · · · ζ K−1
7:
κ) d̂
λ2 = max 0, 0,n √ε n
−YT (wn ) 
−1 ⎢ .. .. ⎥
d,n ⎢ ζ . . ⎥
⎢ 1 ⎥
κ)
= YTn (wn )∗
κ)
+
κ)
λ2
(d̂0,n − YTn (wn )∗ )
κ−1) κ) Cun = σu2n ⎢ ⎥ ⊗ IM (45)
8: d̂0,n κ)
λ2 +1 ⎢ .. .. .. ⎥
κ)
κ)
|λ | κ)
⎣ . . . ζ ⎦
9: λ3 = √1 wn 
εa,n
λ
κ) ζ K−1 ··· ζ 1
κ) κ−1) κ)
10: = â0,n − 1κ) wn
â0,n
λ3 38−10 log10 (2×106 )
where σu2n = 10− and ζ = e− 3.069 . The Jahn’s
2
κ) κ−1)
11: until wn − wn  ≤ a small value 10
κ) κ) κ)
12: w̄n ← wn , d̄0,n ← d̂0,n , ā0,n ← â0,n channel model [35], which is the measurement-based
satellite propagation channel model, is used to evaluate the
tracking algorithms in a realistic environment. The values
V. PERFORMANCE EVALUATION AND DISCUSSIONS for the channel model parameters are given only for the
LOS elevation angles of 15, 25, 35, 45, and 55 deg in [35].
This section provides a comprehensive performance For other elevation angles, the values are obtained with
evaluation and analysis of the proposed algorithm. The linear interpolation or extrapolation as described in [53].
aim of this paper is to analyze the effect of the model
mismatch on the performance of Bayesian filtering
algorithms and to validate the robustness of the proposed 4 Inthe EKFw/o BF and PFw/o BF algorithms, {α,n }Nm
α=0 are tracked by
algorithm against the model mismatch. The performance using the array signal samples yn ∈ CKM .

1274 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
The initial channel is generated by following the Jahn’s
channel model, and it varies over time following the
first-order autoregressive model described in Section II.
The parameters for the channel evolution are given below
(see the paragraph below (47)). The initial DOA (elevation
angle) of the LOS signal is assumed to follow N (60◦ , υ 2 )
where υ 2 is the corresponding variance. Simulations are
conducted under υ = 5◦ and υ = 10◦ to investigate its
effect. Under this assumption, the number of multipath is
determined with the Poisson distribution with a mean of
3.8, and the relative delay of each multipath from the LOS
signal is drawn from the exponential distribution with a
mean of 0.079Tc in urban situations (see Table 2 in [35]).
We consider that the receiver is jammed with a single-tone
CW interference [54]:
gk,n = Ag,n ej 2π fg,n ((n−1)KTs +kTs ) a(θg,n ) (46)
where Ag,n is the amplitude, fg,n = 5000 Hz is the
frequency offset, and θ g,n is the DOA of the interference.
The signal-to-interference ratio (SIR) after despreading is
assumed to be –20 dB.
The initial TOA and DOA of the LOS signal are drawn
from N (15Tc , Tc2 ) and N (60◦ , υ 2 ). The initial complex
amplitude for the LOS signal is statistically generated by
following the Jahn’s model [35]. Accordingly, in the
proposed algorithm, initial particles are drawn from
⎛ ⎞
⎜⎡ ⎤ ⎡ 2 ⎤⎟
⎜ 15Tc Tc 0 0 0 ⎟
⎜ ⎟
⎜⎢ 60◦ ⎥ ⎢ 0 υ2 0 0 ⎥⎟

⎜⎢ ⎥ ⎢ ⎟
s0,0 ∼ N ⎜⎢ ⎥, ⎢ ⎥⎟ .
⎜⎣ R(ψ0,0 ) ⎦ ⎣ 0 0 0.152 0 ⎦ ⎟
⎜ ⎟
⎜ I (ψ0,0 ) 2 ⎟


0 0 0 0.15 ⎠



¯ 0,0 C0,0
(47)
The dynamic model parameters are given by  0,n = I4 and
C0,vn = diag{(0.005Tc )2 , (0.05◦ )2 , (0.03)2 , (0.03)2 }.
(48)
Fig. 3. RMSE results of tracking algorithms under M = 5 antennas and
For multipath and interference, the initial DOAs are υ = 5◦ .
assumed to be uniformly distributed over (10◦ , 40◦ ). The
other initial states for multipath are generated according
to the statistic models in [35]. Similarly, we make the
A. Tracking Accuracy of LOS Signal’s TOA and DOA
initial mean vector  ¯ α,0 with the true values, and we
assume Cα ,0 = C0,0 and Cα,vn = 2C0,vn for multipath 1) Effect of the Variance of the Initial DOA for M = 5:
(0 < α ≤ Nm ). For a half-wavelength spacing ULA, the Figs. 3 and 4 present the root mean square error (RMSE)
width of the mainbeam is approximately θ BW = 2 sin–1 results of the EKFw/o BF, PFw/o BF, and proposed
(0.891/M) [48]. θ BW ≈ 20◦ for M = 5, and θ BW ≈ 10◦ for algorithms for υ = 5◦ and υ = 10◦ . In the simulations, the
M = 10. Hence, it is considered that the beamwidth is number of antenna elements was M = 5. Despite υ, the
narrow enough to reject the unwanted signals in proposed algorithm had better tracking performance in
simulations. both TOA and DOA of the LOS signal compared with the
The posterior Cramér-Rao lower bound (PCRLB) is EKFw/o BF and PFw/o BF algorithms. The performance of
also evaluated as benchmark. Details on the derivation of the proposed algorithm was improved as the number of
the PCRLB under the Gaussian noise assumption are particles increased. The proposed algorithm obtained
omitted in this paper, but they are well described in the higher tracking accuracy for υ = 5◦ than for υ = 10◦ since
related literature [3, 16]. The simulation results are more precise beams towards the LOS signal can be formed
averaged over 500 Monte Carlo simulations, which are under the small DOA variation. The results showed that
sufficient to present the trend. the required number of particles (i.e. overall complexity)

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1275
Fig. 5. RMSE results of TOA estimates with proposed algorithm under
varying number of antenna elements.

accuracy of DLL in multipath and CW interference, but in


noise-free environments [55].
Through the simulations, we showed the superiority of
the proposed algorithm over the EKFw/o BF and PFw/o BF
algorithms in the presence of multipath and interference.
For a receiver equipped with an antenna array, the steering
vector or DOA estimation is also an important issue.
While our focus is concentrated on tracking the TOA of
the GNSS signal through the LOS path, it has been shown
that the DOA of the LOS signal estimation problem can be
effectively solved with the proposed algorithm.
2) Effect of the Number of Antenna Elements on TOA
Tracking: Fig. 5 shows the RMSE results of LOS signal’s
TOA estimates by using the proposed algorithm with
S = 1000 particles under varying the number of antenna
elements M and the initial DOA variance υ 2 . The results
Fig. 4. RMSE results of tracking algorithms under M = 5 antennas and
in the figure were captured at n = 40 ms. In order to focus
υ = 10◦ . on how the performance of the proposed algorithm
changes with M, the results of the EKFw/o BF and PFw/o BF
algorithms are not presented.
for a given accuracy can be reduced in the proposed We found an interesting observation that the PCRLB
algorithm with accurate beamforming. As seen in the for M = 10 was lower than that for M = 5, but the
figures, the EKFw/o BF and PFw/o BF algorithms provided proposed algorithm showed lower performance when
poor results, and they lost their tracking function due to the using a larger number of antenna elements. We remind the
model mismatch from unknown interference. In particular, reader that the beamwidth for M = 10 is almost half of the
although the PFw/o BF algorithm used 1000 particles, it had beamwidth for M = 5 in the ULA context. In general, it is
a performance even lower than that of the EKFw/o BF and expected that the performance is improved for a narrower
showed convergence at the points larger than the initial beam with a higher array gain. However, in the proposed
uncertainties from the start. This implies that a small algorithm, a narrower beam does not guarantee higher
portion of particles having large errors became dominant performance in TOA tracking if υ > θ BW /2. As υ becomes
because of unknown interference, and other particles were larger than θ BW /2, more particles are likely to have the
duplicated with the dominant ones through resampling. DOAs out of the beam and those particles have near zero
The TOA tracking accuracy of the proposed algorithm weights even though their TOAs are accurate. In the
with 1000 particles was approximately 0.15 chips, which resampling phase, they are substituted with others, and
is deemed significant for the GNSS applications. However, consequently, the TOA tracking performance is limited
it is worth noting that this accuracy is comparable to the with a loss of diversity of TOAs in particles. In

1276 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
conclusion, as υ is smaller, the proposed algorithm
enables to track the LOS signal more accurately by using a
larger number of antenna elements (i.e. by increasing the
entire array aperture).

B. Reduced State Signal Model Validation


To validate the reduced state signal model in
(19), we verify if zn = WH n yn has the distribution
CN (WH n h(0,n ), Cũn ) with an empirical goodness-of-fit
test introduced in [56]. According to the fit test, the
equality between two multivariate distributions can be
shown by comparing their empirical distributions of the
Mahalanobis distance  (b, ), which is given by
 
(b, ) =  −1/2 (b − E[b]) (49)
where b is a random vector from a multivariate distribution
and  = E[(b − E[b])(b − E[b])H ]. In other words, when
the empirical distributions of (WH n h(0,n ) + ũn , Cũn )
and (zn ,  zn )with the covariance matrix  zn of zn are Fig. 6. Empirical distributions of normalized Mahalanobis distance
approximately the same, the distribution of zn is said to with proposed algorithm.
follow CN (WH n h(0,n ), Cũn ). For the sake of
convenience, we refer to the former as the 1-distribution TABLE I
and the latter as the 2-distribution. Parameters for Distributions of Normalized Mahalanobis Distance under
Gaussian Fitting
It is well known that an empirical distribution is
#
simply estimated with a sufficiently large number of E[] E[( − E[])2 ]
random samples. Since the distribution estimation
is very straightforward, we do not provide the details. Type S = 100 S = 1000 S = 100 S = 1000
However, given Wn , only a single random vector M=5
(a sample from a multivariate distribution) is available 1-dist. ν = 5◦ 0.4648 0.4925 0.1157 0.1259
in the simulations. Since ũn = WH ν = 10◦ 0.5158 0.4637 0.1285 0.1202
Nm n un and
2-dist. ν = 5◦ 0.4650 0.4933 0.1037 0.1120
zn = Wn ( α=0 h(α,n ) + gn + un ), the randomness in
H
ν = 10◦ 0.5173 0.4644 0.1153 0.1073
vectors WH n h(0,n ) + ũn and zn stems from the same M = 10
noise term un . Hence, we generate additional random 1-dist. ν = 5◦ 0.4675 0.4801 0.1137 0.1167
noise vectors from CN (0KM , Cun ) in order to derive the ν = 10◦ 0.4973 0.4939 0.1179 0.1161
empirical 1- and 2-distributions and to validate the 2-dist. ν = 5◦ 0.4682 0.4806 0.1015 0.1044
ν = 10◦ 0.4980 0.4949 0.1053 0.1038
proposed model by comparing them. Let uin and Ns be the
i-th random noise vector and the total number of noise
vectors. Accordingly, ũin = WH i i
n un and zn denote the i-th
noise and received signal vectors after beamforming. Now, To ease the comparison, the Mahalanobis distances
the Mahalanobis distance for the i-th random vector bi of were normalized such that  ∈ [0, 1]. Fig. 6 shows the
the (·)-distribution is expressed as  (bi , ) where empirical 1- and 2-distributions of the normalized
Mahalanobis distance for the 5-element ULA receiver.
bi = W H
n h(0,n ) + ũn ,
i
Table I describes the parameter estimates for Gaussian
 = Cũn = Wn Cun Wn ,
H
(50) distributions, which were obtained by using dfittool of
MATLAB [57]. It is clearly seen that the 1- and
E[b] = WHn h(0,n )
2-distributions were very close in appearance to each other
for the 1-distribution function, and despite υ. From the fact that the WHn h(0,n ) + ũn
(1-distribution) follows CN (WH n h(0,n ), Cũn ), we can
bi = zin , conclude that zn is also Gaussian distributed and that our
1  proposed model holds.
Ns
= (bi − E[b])(bi − E[b])H , (51)
Ns i=1
C. Convergence of Iterative Robust Hybrid Beamformer
1 
Ns
To show the convergence of the proposed robust
E[b] = bi
Ns i=1 hybrid beamformer, the MSE between the beamformer
output and the LOS signal vector, which is defined as the
for the 2-distribution function. We set Ns = 50 000 to objective function in (27), was measured with respect to
estimate the distribution functions. iterations of algorithm 2. The MSE results are shown in

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1277
TABLE II
Computational Complexity for Basic Operations for K M

Operation Size Cost

aa Vector scaling a∈ ∈
C1 , a CM O(M)
a + b Vector addition a, b ∈ CM O(M)
aH b Inner product a, b ∈ CM O(2M)
Ac Matrix × Vector A ∈ CM×K , c ∈ CK O(2MK)
AB Matrix × Matrix A ∈ CM×K , B ∈ CK×N O(2MKN)
aH Ra Hermitian form a ∈ CM , R ∈ CM×M O( 23 M 2 )
R–1 Matrix inversion R ∈ CM×M O(M3 )

TABLE III
Computational Complexity of Tracking Algorithms

Algorithm Instruction Cost

EKFw/o BF State prediction O(2χ 2 )


Predicted covariance O(3χ 3 )
Kalman gain O(K3 M3 )
State update O((2χ + 1)KM)
Updated covariance O(2KMχ 2 )
Fig. 7. Convergence of proposed iterative beamformer: MSE converges PFw/o BF Particle propagation O((χ 2 + χ )S)
after 15 iterations for M = 5 antennas and after 40 iterations for M = 10 Likelihood O(K3 M3 + 32 K2 M2 S)
antennas. Weight update & Estimation O((χ + 1)S)
Proposed Particle propagation O(36S)
Inv. sample covariance R−1
n O(2KM2 )
Fig. 7. Since a larger array gain is obtained with the Adaptive error bounds O(4KS + 4MS)
increase of antenna elements in the ULA, the MSE results Lagrange multipliers O((2M + 3)ρK)
for M = 10 were lower than that for M = 5. The Beamforming optimization O((4M + 1)ρK)
convergence was observed nearly after 15 iterations for Likelihood O(K3 +
M = 5 and after 40 iterations for M = 10. Compared with (2M2 + 2M + 32 S)K2 )
Weight update & Estimation O(5S)
the results for M = 5, the distinctive oscillation was found
at early iterations when M = 10. Since a larger value of
ε a,n , which is the error bound on the steering vector, is
However, in the GNSS scenarios, the complexity in terms
obtained with the increase of the number of antenna
of the measurement size is dominant in the overall cost
elements, λκ) κ) κ)
1 /λ3 increases and â0,n is more dependent on due to the fact that K χ > D(0,n ). Detailed analysis
the beamforming vector wκ) n than the previous estimate on the complexity of the Bayesian filtering algorithms is
âκ−1)
0,n . ε a,n also increases with υ. These bring the available in [42]. The amounts of time for generating
oscillation and the increase of time for convergence when random numbers in particle propagation and in resampling
M and υ are large. Nevertheless, it should be emphasized are proportional to the state dimension and the number of
that the complexity for convergence is negligibly small particles. However, compared with other operations, the
compared with the complexity for the likelihood complexity of the random number generation is
computation in the proposed tracking algorithm. insignificant. Therefore, in our analysis, it is assumed that
the resultant complexity is constant [i.e. O(1)] for
D. Complexity Analysis simplicity.
The equivalent flop complexity for each tracking Recall thatK M. Considering only the most
algorithm is provided for the complexity analysis [42]. significant terms, the computational complexity of each
The equivalent flop complexity is defined as the number of tracking algorithm is given by
flops for an operation. The complexity for the basic 1) O(K3 M3 ) flops for the Kalman gain computation in
operations is summarized in Table II. the EKF algorithm,
Table III shows the computational complexity for the 2) O(K3 M3 + 32 K2 M2 S) flops for the likelihood
main operations of the EKFw/o BF, PFw/o BF, and proposed computation in the PF algorithm, and
algorithms. The state dimension for the EKFw/o BF and 3) O(K 3 + (2M 2 + 2M + 32 S)K 2 ) flops for the
PFw/o BF algorithms is denoted by χ. In the simulations, χ likelihood computation in the proposed algorithm.
= 4(Nm + 1) as the states of the LOS and multipath signals
are tracked, but χ = D(xn ) = 4(Nm + 1) + D(n ) when According to Table II, the most demanding operation of
the state vector includes the interference parameters as the proposed algorithm is the computation of
described in Section III. In general, the computational Cũn = WH n Cun Wn that requires O((2M + 2M)K ) flops:
2 3

complexity of a Bayesian filtering algorithm is determined O(2K M ) flops for (K × KM) × (KM × KM) and
3 2

by the state space dimension and the measurement size. O(2K3 M) for (K × KM) × (KM × K). However, since

1278 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
order to improve the robustness to the errors in the LOS
signal estimates and to avoid the model mismatch from the
reduced state signal model. The proposed algorithm was
analyzed under varying the numbers of particles and
antenna elements and the initial DOA variance. It was
shown that the proposed algorithm is computationally
more efficient (nearly 1000 times better for M = 10
antennas and S = 1000 particles) and retains the tracking
function, while the existing Bayesian filtering algorithms
that track the LOS and multipath signals without
beamforming lose their function under the model
mismatch from strong interference.
Even though the proposed algorithm was developed
and verified for the GPS signal, it is applicable to the
general wireless communication systems. Highly efficient
computational complexity with the large number of array
antenna elements allows the proposed approach to be an
attractive solution for highly directive future
communication such as millimeter-wave systems, which is
Fig. 8. Computational complexity of EKFw/o BF, PFw/o BF, and left as our future work.
proposed tracking algorithms for K = 3069.

APPENDIX
Wn = IK ⊗ wn , Cũn can be computed as a (1 × M) ×
The optimal solutions to (37) can be derived by setting
(M × KM) operation with K loops and a (K × M) ×
the partial derivatives of (38) with respect to wn , d0,n , and
(M × 1) operation with K loops. Hence, the complexity
a0,n to 0M × 1 , 0K × 1 , and 0M × 1 , respectively [58]. Denote
for the computation of Cũn can be reduced as
the optimal solutions by w̄n , d̄0,n , and ā0,n .
O((2M2 + 2M)K2 ) flops.
Differentiation of (38) with respect to wn gives
Fig. 8 shows how fast the computational complexity of
the tracking algorithms increases with increasing the Rn w̄n − Yn d∗0,n + λ1 a0,n = 0M×1 (52)
number of antennas for K = 3069. The EKFw/o BF and
PFw/o BF algorithms have exponential growths in the where Rn = Yn YHn . Hence,
complexity while the complexity of the proposed  
algorithm increases much more slowly than the other two w̄n = R−1n Yn d∗0,n − λ1 a0,n . (53)
algorithms. From the results, the computational efficiency By replacing wn with w̄n in the constraint in (27),
of the proposed algorithm becomes more obvious when
−1 −1 H −1 ∗
using a larger number of antennas. In general, Kalman λ1 = (aH
0,n Rn a0,n ) (a0,n Rn Yn d0,n − 1). (54)
filtering algorithms are known to be more efficient than
particle filtering algorithms. However, by using the The partial derivative of (38) with respect to d0,n is
reduced state signal model with beamforming, the state given by
space and the measurement size are significantly − YTn w∗n + d̄0,n + λ2 (d̄0,n − d̂0,n ) = 0K×1 . (55)
reduced. As a result, the proposed algorithm even with
10 000 particles is around 10 000 times faster than the Dividing (55) by λ2 , we get
EKFw/o BF algorithm for M = 50. The complexity of
λ2 + 1 1
the proposed algorithm can be further reduced with d̄0,n = d̂0,n + YTn w∗n
Rao-Blackwellization for the complex amplitude of the λ2 λ2
λ2 + 1 T ∗
LOS signal as addressed in [3]. = Yn wn + d̂0,n − YTn w∗n (56)
λ2
VI. CONCLUSIONS
Hence,
In this paper, we derived the reduced state signal
λ2
model with beamforming for the computational efficiency d̄0,n = YTn w∗n + (d̂0,n − YTn w∗n ). (57)
and the avoidance of the model mismatch from unknown λ2 + 1
interfering signals. In order to track the LOS signal based To derive λ2 that satisfies the inequality constraint (33), we
on the reduced state signal model, we proposed an arrange the difference between d̄0,n and d̂0,n as follows:
integrated robust hybrid beamformer and particle
filter-based algorithm. The proposed algorithm uses the λ2
d̄0,n − d̂0,n = −(d̂0,n − YTn w∗n ) + (d̂0,n − YTn w∗n )
particles to approximate the upper bounds of the λ2 + 1
uncertainties in the LOS signal estimates, and it derives 1
=− (d̂0,n − YTn w∗n ) (58)
the beamforming vector subject to the adaptive bounds in λ2 + 1

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1279
The inequality constraint can be expressed as [8] Van Nee, R. D. J., Siereveld, J., Fenton, P. C., and
 2 Townsend, B. R.
  1
d̄0,n − d̂0,n  = d̂0,n − YTn w∗n 2 ≤ εd,t . (59) The multipath estimating delay lock loop: Approaching
(λ2 + 1)2 theoretical accuracy limits.
In Proceedings of 1994 IEEE Position, Location and
Thus, λ2 ≥ 0 is given by Navigation Symposium, Las Vegas, NV, Apr. 11-15, 1994,
⎛  

 ⎞ 246–251.
d̂0,n − YTn w∗n  [9] Hurskainen, H., Lohan, E. S., Hu, X., Raasakka, J., and Nurmi, J.
λ2 = max ⎝0, √ − 1⎠ . (60) Multiple gate delay tracking structures for GNSS signals and
εd,n their evaluation with Simulink, SystemC, and VHDL.
International Journal of Navigation and Observation, 2008
The optimal steering vector for the LOS signal ā0,n can (2008), Article ID 785695, doi:10.1155/2008/785695.
be obtained by differentiating (38) with respect to a0,n : [10] Bhuiyan, M. Z. H., and Lohan, E. S.
Advanced multipath mitigation techniques for satellite-based
λ1 wn + λ3 (ā0,n − â0,n ) = 0M . (61) positioning applications.
International Journal of Navigation and Observation, 2010
From (61), (2010), Article ID 412393, doi:10.1155/2010/412393.
[11] Lohan, E. S., Lakhzouri, A., and Renfors, M.
λ1
ā0,n = â0,n − wn . (62) Feedforward delay estimators in adverse multipath
λ3 propagation for Galileo and modernized GPS signals.
EURASIP Journal on Applied Signal Processing, 2006 (May
Substituting ā0,n into (34), the corresponding Lagrange
2006), Article ID 50971, doi:10.1155/ASP/2006/50971.
multiplier is given by [12] Izadpanah, A., O’Driscoll, C., and Lachapelle, G.
GPS multipath parameterization using the extended Kalman
|λ1 | filter and a dual LHCP/RHCP antenna.
λ3 = √ wn . (63) In Proceedings of the 21st International Technical Meeting of
εa,n the Satellite Division of the Institute of Navigation, Savannah,
GA, Sept. 16-19, 2008, 689–697.
[13] Closas, P., Fernández-Prades, C., Fernández-Rubio, J. A., and
REFERENCES Ramı́rez-González, A.
Multipath mitigation using particle filtering.
[1] Cahn, C. R., Leimer, D., Marsh, C., Huntowski, F., and LaRue, G. In Proceedings of the 19th International Technical Meeting of
Software implementation of a PN spread spectrum receiver to the Satellite Division of the Institute of Navigation, Fort
accommodate dynamics. Worth, TX, Sept. 26-29, 2006, 1733–1740.
IEEE Transactions on Communications, 25, 8 (Aug. 1977), [14] Closas, P., Fernández-Prades, C., and Fernández-Rubio, J. A.
832–840. Bayesian DLL for multipath mitigation in navigation systems
[2] Van Dierendonck, A. J., Fenton, P., and Ford, T. using particle filters.
Theory and performance of narrow correlator spacing in a In Proceedings of 2006 IEEE International Conference on
GPS receiver. Acoustics, Speech, and Signal Processing, Toulouse, France,
Journal of the Institute of Navigation, 39, 3 (Fall 1992), May 14-19, 2006, 129–132.
265–283. [15] Liu, X., Closas, P., Liu, J., and Hu, X.
[3] Closas, P., Fernández-Prades, C., and Fernández-Rubio, J. A. Particle filtering and its application for multipath mitigation
A Bayesian approach to multipath mitigation in GNSS with GNSS receivers.
receivers. In Proceedings of 2010 IEEE/ION Position, Location and
IEEE Journal of Selected Topics in Signal Processing, 3, 4 Navigation Symposium, Indian Wells, CA, May 4-6, 2010,
(Aug. 2009), 695–706. 1168–1173.
[4] McGraw, G. A., and Braasch, M. S. [16] Ristic, B., Arulampalam, S., and Gordon, N.
GNSS multipath mitigation using gated and high resolution Beyond the Kalman Filter: Particle Filters for Tracking
correlator concepts. Applications. London: Artech House, 2004.
In Proceedings of the 1999 National Technical Meeting of the [17] Lakhzouri, A., Lohan, E. S., and Renfors, M.
Institute of Navigation, San Diego, CA, Jan. 25-27, 1999, Estimation of closely-spaced paths via particle filters for
333–342. WCDMA positioning.
[5] Van Dierendonck, A. J., and Braasch, M. S. In Proceedings of 2004 First International Symposium on
Evaluation of GNSS receiver correlation processing Control, Communications and Signal Processing, Hammamet,
techniques for multipath and noise mitigation. Tunisia, Mar. 21-24, 2004, 791–794.
In Proceedings of the 1997 National Technical Meeting of the [18] Giremus, A., Tourneret, J.-Y., and Calmettes, V.
Institute of Navigation, Santa Monica, CA, Jan. 14-16, 1997, A particle filtering approach for joint detection/estimation of
207–215. multipath effects on GPS measurements.
[6] Irsigler, M., and Eissfeller, B. IEEE Transactions on Signal Processing, 55, 4 (Apr. 2007),
Comparison of multipath mitigation techniques with 1275–1285.
consideration of future signal structures. [19] Krach, B., Lentmaier, M., and Robertson, P.
In Proceedings of the 16th International Technical Meeting of Bayesian detection and tracking for joint positioning and
the Satellite Division of the Institute of Navigation, Portland, multipath mitigation in GNSS.
OR, Sept. 9-12, 2003, 2584–2592. In Proceedings of the 5th Workshop on Positioning,
[7] Bhuiyan, M. Z. H., Lohan, E. S., and Renfors, M. Navigation and Communication, Hannover, Germany, Mar.
A slope-based multipath estimation technique for mitigating 27, 2008, 173–180.
short-delay multipath in GNSS receivers. [20] Boulton, P., Borsato, R., Butler, B., and Judge, K.
In Proceedings of 2010 IEEE International Symposium on GPS interference testing.
Circuits and Systems, Paris, May 30-June 2, 2010, 3573–3576. Inside GNSS, (July/Aug. 2011), 32–45.

1280 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
[21] Wildemeersch, M., Pons, E. C., Rabbachin, A., and Guasch, J. F. Proceedings of the IEEE, 57, 8 (Aug. 1969), 1408–1418.
JRC scientific and techincal reports: Impact study of [35] Jahn, A., Bischl, H., and Heiss, G.
unintentional interference on GNSS receivers. Channel characterisation for spread spectrum satellite
European Commission Joint Research Centre, Tech. Rep., communications.
2010. In Proceedings of 1996 IEEE 4th International Symposium on
[22] Balaei, A. T., Motella, B., and Dempster, A. G. Spread Spectrum Techniques and Applications, Mainz,
GPS interference detected in Sydney-Australia. Germany, Sept. 22-25, 1996, 1221–1226.
In Proceedings of 2007 International Global Navigation [36] Lee, J.-H., and Cheng, C.-C.
Satellite Systems Symposium on GPS/GNSS, Sydney, Spatial correlation of multiple antenna arrays in wireless
Australia, Dec. 4-6, 2007, 74–76. communication systems.
[23] Kim, S.-J., and Iltis, R. A. Progress In Electromagnetics Research, 132 (2012), 347–368,
Performance comparison of particle and extended Kalman doi:10.2528/PIER12080604.
filter algorithms for GPS C/A code tracking and interference [37] Kong, S.-H.
rejection. A deterministic compressed GNSS acquisition technique.
In Proceedings of the 36th Annual Conference on Information IEEE Transactions on Vehicular Technology, 62, 2 (Feb.
Sciences and Systems, Princeton, NJ, Mar. 20-22, 2002. 2013), 511–521.
[24] Kim, S.-J., and Iltis, R. A. [38] Borio, D.
A mean-value theorem-based particle filter for nonlinear GNSS acquisition in the presence of continuous wave
estimation. interference.
In Proceedings of the 36th Annual Conference on Information IEEE Transactions on Aerospace and Electronic Systems, 46,
Sciences and Systems, Princeton, NJ, Mar. 20-22, 2002. 1 (Jan. 2010), 47–60.
[25] Vaswani, N. [39] Kim, S.-J., and Iltis, R. A.
Bound on errors in particle filtering with incorrect model STAP for GPS receiver synchronization.
assumptions and its implication for change detection. IEEE Transactions on Aerospace and Electronic Systems, 40,
In Proceedings of 2004 IEEE International Conference on 1 (Jan. 2004), 132–144.
Acoustics, Speech and Signal Processing, Montreal, Canada, [40] Iltis, R. A.
May 17-21, 2004, 729–732. An EKF-based joint estimator for interference, multipath, and
[26] Vaswani, N. code delay in a DS spread-spectrum receiver.
Particle filtering for large-dimensional state spaces with IEEE Transactions on Communications, 42, 234 (Feb. 1994),
multimodal observation likelihoods. 1288–1299.
IEEE Transactions on Signal Processing, 56, 10 (Oct. 2008), [41] Tsui, J. B.-Y.
4583–4597. Fundamentals of Global Positioning System Receivers: A
[27] He, X., Chen, Y., and Iz, H. B. Software Approach. New York: Wiley, 2000.
A reduced-order model for integrated GPS/INS. [42] Karlsson, R., Schon, T., and Gustafsson, F.
IEEE Aerospace and Electronic Systems Magazine, 13, 3 Complexity analysis of the marginalized particle filter.
(Mar. 1998), 40–45. IEEE Transactions on Signal Processing, 53, 11 (Nov. 2005),
[28] Luo, Y., Wu, W., Babu, R., Tang, K., and Luo, B. 4408–4411.
A simplified baseband prefilter model with adaptive Kalman [43] Petovello, M.
filter for ultra-tight COMPASS/INS integration. GNSS solutions: Clock offsets in GNSS receivers.
Sensors, 12, 7 (July 2012), 9666–9686. Inside GNSS, (Mar./Apr. 2011), 23–25.
[29] Seco-Granados, G., Fernández-Rubio, J. A., and [44] Vorobyov, S. A., Gershman, A. B., and Luo, Z.-Q.
Fernández-Prades, C. Robust adaptive beamforming using worst-case performance
ML estimator and hybrid beamformer for multipath and optimization: A solution to the signal mismatch problem.
interference mitigation in GNSS receivers. IEEE Transactions on Signal Processing, 51, 2 (Feb. 2003),
IEEE Transactions on Signal Processing, 53, 3 (Mar. 2005), 313–324.
1194–1208. [45] Manosas-Caballu, M., Seco-Granados, G., and Swindlehurst,
[30] Arribas, J., Fernández-Prades, C., and Closas, P. A. L.
Multi-antenna techniques for interference mitigation in GNSS Robust beamforming via FIR filtering for GNSS multipath
signal acquisition. mitigation.
EURASIP Journal on Advances in Signal Processing, 2013 In Proceedings of 2013 IEEE International Conference on
(Sept. 2013), Article ID 143, Acoustics, Speech, and Signal Processing, Vancouver, Canada,
doi:10.1186/1687-6180-2013-143. May 26-31, 2013, 4173–4177.
[31] Arribas, J., Fernández-Prades, C., and Closas, P. [46] Liao, B., Chan, S.-C., and Tsui, K.-M.
Antenna array based GNSS signal acquisition for interference Recursive steering vector estimation and adaptive
mitigation. beamforming under uncertainties.
IEEE Transactions on Aerospace and Electronic Systems, 49, IEEE Transactions on Aerospace and Electronic Systems, 49,
1 (Jan. 2013), 223–243. 1 (Jan. 2013), 489–501.
[32] Lee, S., Choi, J., Park, C., Lee, M. J., and Kim, S. [47] Bergman, N.
Joint channel tracking and beamforming via particle filtering Recursive Bayesian estimation: Navigation and tracking
in GNSS receivers. applications.
In Proceedings of 2013 International Symposium on Global Ph.D. dissertation, Linköpings University, Sweden, 1999.
Navigation Satellite System, Istanbul, Turkey, Oct. 22-25, [48] Hur, S., Kim, T., Love, D. J., Krogmeier, J. V., Thomas, T. A.,
2013. and Ghosh, A.
[33] Li, J., Stoica, P., and Wang, Z. Millimeter wave beamforming for wireless backhaul and
On robust Capon beamforming and diagonal loading. access in small cell networks.
IEEE Transactions on Signal Processing, 51, 7 (July 2003), IEEE Transactions on Communications, 61, 10 (Oct. 2013),
1702–1715. 4391–4403.
[34] Capon, J. [49] Robert, C. P., and Casella, G.
High resolution frequency-wavenumber spectrum analysis. Monte Carlo Statistical Methods. New York: Springer, 2004.

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1281
[50] Chong, E. K. P., and Zak, S. H. Interference mitigation in GNSS receivers by a
An Introduction to Optimization, time-frequency approach.
4th ed. New York: Wiley, 2013. IEEE Transactions on Aerospace and Electronic Systems, 49,
[51] Wan, E. A., and van der Merwe, R. 1 (Jan. 2013), 415–438.
The unscented Kalman filter for nonlinear estimation. [55] Qu, B., Wei, J., Tang, Z., Yan, T., and Zhou, Z.
In Proceedings of the IEEE 2000 Adaptive Systems for Signal Analysis of combined effects of multipath and CW
Processing, Communications, and Control Symposium, Lake interference on coherent delay lock loop.
Louise, Canada, Oct. 1-4, 2000, 153–158. Wireless Perssonal Communications, 77, 3 (Feb. 2014),
[52] van der Merwe, R., and Wan, E. A. 2213–2233.
Sigma-point Kalman filters for integrated navigation. [56] McAssey, M. P.
In Proceedings of the 60th Annual Meeting of the Institute of An empirical goodness-of-fit test for multivariate distributions.
Navigation, Dayton, OH, June 7-9, 2004, 641–654. Journal of Applied Statistics, 40, 5 (Mar. 2013), 1120–1131.
[53] Propagation data required for the design of Earth-space land [57] MATLAB, version 8.4.0.150421 (R2014b). The MathWorks Inc.,
mobile telecommunication systems. 2014.
International Telecommunication Union Recommendation [58] Petersen, K. B., and Pedersen, M. S.
P.681-7, Oct. 2009. The Matrix Cookbook, Nov. 2012. [Online]. Available:
[54] Savasta, S., Lo Presti, L., and Rao, M. http://matrixcookbook.com.

Sangwoo Lee received the B.S. and M.S. degrees both in electrical engineering from
Ajou University, Suwon, Korea in 2009 and 2011, respectively.
He is currently pursuing the Ph.D. degree in electronics and computer engineering
at Hanyang University, Seoul, Korea. His research interests include
localization/positioning systems, wireless sensor networks, and channel
estimation/tracking.

Elena Simona Lohan (SM’13) received an M.Sc. degree in electrical engineering from
Polytechnics University of Bucharest, Romania, in 1997, a D.E.A. degree (French
equivalent of master) in econometrics, at Ecole Polytechnique, Paris, France, in 1998,
and a Ph.D. degree in telecommunications from Tampere University of Technology
(TUT), Finland, in 2003.
She is now an Associate Professor and an Academy Research Fellow at the
Department of Electronics and Communication Engineering (ELT) at TUT and she is
the group leader for the mobile and satellite-based positioning activities (signal
processing part) at ELT. She is also a Visiting Professor at Universitat Autonoma de
Barcelona, Spain. She is currently participating in the Marie Curie ITN network
MULTI-POS on wireless positioning as scientist in charge and equality officer. Her
current research interests include wireless location techniques based on signals of
opportunity, wireless navigation receiver architectures and multipath mitigation, and
cognitive spectrum sensing for positioning purposes.

1282 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 52, NO. 3 JUNE 2016
Sunwoo Kim (M’05) received the B.S. in electrical engineering from Hanyang
University, Seoul, Korea in 1999, and the M.S. and the Ph.D. degrees in electrical and
computer engineering from the University of California, Santa Barbara, in 2002 and
2005. Since 2005 he has been with the Department of Electronics Engineering,
Hanyang University, Seoul, Korea, and is currently an associate professor.
His research interests include wireless communications, wireless localization, and
signal processing, with current emphasis on location-aware communications.

LEE ET AL.: ARRAY-BASED GNSS SIGNAL TRACKING WITH A REDUCED STATE SIGNAL MODEL 1283

You might also like