ch21 - Direct Position Estimation - 9781119458449

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

529

21

Direct Position Estimation


Pau Closas1 and Grace Gao2
1
Northeastern University, United States
2
Stanford University, United States

21.1 Introduction well as the basics of the signal processing that apply to
standard GNSS receivers. DPE is introduced in Section
This chapter presents an innovative concept that helps 21.3, where extensive details are provided regarding the sig-
overcome some of the critical challenges and limitations nal model and the direct-positioning concept. Additionally,
in standard global navigation satellite system (GNSS) posi- the section provides theoretical results demonstrating the
tion, navigation, and timing (PNT). The concept is known performance enhancement of DPE with respect to two-
as Direct Position Estimation (DPE) and was introduced in steps approaches. Implementation details are discussed in
[1]. At a glance, a receiver implementing DPE computes the Section 21.4, where we move from the purely theoretical
position directly from the sampled signal, unlike a conven- results in the previous section to the more practical aspects.
tional receiver in which range estimates (i.e. pseudoranges) Section 21.5 discusses some variants and alternative appli-
are required before solving for the user’s position [2]. cations of DPE. Finally, conclusions are provided in
Therefore, conventional receivers are generally referred Section 21.6.
to as two-step receivers and DPE as one-step receivers.
The direct search and joint optimization across multiple
satellites makes DPE a robust GNSS position and time esti- 21.2 Conventional GNSS Signal
mation technique [3]. In contrast, conventional techniques, Processing and Positioning
such as scalar tracking, first estimate intermediate pseudor-
ange and pseudorange rate measurements to each satellite in This section provides a common framework to mathemat-
view [4, 5], then they solve for the navigation solution, typ- ically describe the basic signal processing pipeline of a con-
ically via an iterative least-squares approach [6, 7]. Vector ventional two-step GNSS receiver. Then, in Section 21.3, we
tracking, a more robust approach than scalar tracking, will leverage these definitions to better understand DPE
jointly processes intermediate measurements across multi- processing. First, we discuss the transmitted signal and
ple satellites in view by mapping intermediate measurement channel models to characterize the signal received at the
residuals into shared navigation residuals [8, 9]. However, it antenna. Then, we describe the signal processing and nav-
is an indirect navigation estimation process based on inter- igation solution equations, which ultimately yield a posi-
mediate measurement residuals which are separately esti- tion, velocity, and time (PVT) estimate.
mated across multiple satellites [10, 11]. When the signals
are degraded, both scalar and vector tracking discard inter-
21.2.1 GNSS Signal Model
mediate measurements which are likely experiencing a fault.
[12, 13]. In contrast, DPE is able to preserve such informa- A general signal model for satellite navigation systems con-
tion [14] and provide improved performance under challeng- sists of a direct-sequence spread-spectrum (DS-SS) signal,
ing situations [2, 15]. A comparison of scalar tracking, vector synchronously transmitted by all the satellites in the con-
tracking‚ and DPE is given in Table 21.1. stellation. This type of signal enables code division multiple
The chapter is organized as follows. With the intention of access (CDMA) transmissions; that is, satellite signals are
facilitating the introduction of new material, Section 21.2 distinguished by orthogonal (or quasi-orthogonal) codes.
describes the signal model used through this chapter, as At a glance, these signals consist of two main components:

Position, Navigation, and Timing Technologies in the 21st Century: Integrated Satellite Navigation, Sensor Systems, and Civil Applications, Volume 1,
First Edition. Edited by Y. T. Jade Morton, Frank van Diggelen, James J. Spilker Jr., and Bradford W. Parkinson.
© 2021 The Institute of Electrical and Electronics Engineers, Inc. Published 2021 by John Wiley & Sons, Inc.
530 21 Direct Position Estimation

Table 21.1 Comparison between scalar tracking, vector tracking, and Direct Position Estimation (DPE)

Approach Remarks

Scalar tracking
• Estimates intermediate range measurements of each satellite
• Susceptible to intermediate range
estimation errors

• Does not account for inter-link


correlations

Vector tracking
•• Estimates intermediate measurement residuals
Couples signal tracking and PVT estimation, so that all
channels aid each other by sharing information
• Susceptible to intermediate residual
estimation errors

• Maps intermediate measurement residuals to navigation residuals

Direct Position
Estimation
• Maximizes cross-correlation of expected GPS signal reception with
received GPS signal at multiple navigation candidates •• Direct search
Joint optimization across satellites

a ranging code using a pseudorandom noise (PRN) spread-


ing sequence; and a low-rate data link to broadcast neces-
• PI,i is the transmitted power, considered equal for all
satellites, and is elevation-dependent due to the antenna
sary information for positioning such as satellite orbital pattern at the satellite [16].
parameters and corrections. The complex baseband model
of the signal transmitted by the i-th satellite is
• bI,i(t) {−1, 1} is the sequence of low-rate data bits, with
T bI being the bit period.

sT,i t = sI,i t + jsQ,i t , 21 1


• cI,i(t) {−1, 1} is the PRN spreading sequence, not to be
confused with c, the constant representing the speed of
light. The chip length of the sequence and the chip period
where its in-phase and quadrature components are are denoted by LcI and T cI , respectively. Therefore,
defined as T PRNI = LcI T cI is the codeword period. N cI denotes the
number of code epochs per data bit. Figure 21.1 aims to
clarify the relation between these bits/chips parameters.


sI,i t = 2PI,i qI,i t
∞ N cI LcI The energy-normalized chip shaping pulse is denoted
= 2PI,i bI,i mI cI,i kI gI t − mI T bI − uI T PRNI − kI T cI by gI(t).
mI = − ∞ uI = 1 k I = 1

21 2 Note that several carrier frequencies can be used at a time,


with the presented baseband structure being thus repli-
sQ,i t = 2PQ,i qQ,i t
N cQ LcQ
cated. Actually, GNSS systems typically transmit comple-

= 2PQ,i bQ,i mQ cQ,i k Q gQ t − mQ T bQ − uQ T PRNQ − k Q T cQ , mentary signals over several frequency bands. One of the
mQ = − ∞ uQ = 1 k Q = 1
main reasons for this is to combat the frequency-dependent
21 3 perturbations suffered by the signal when traveling through
the ionosphere and troposphere.
where the following definitions for the in-phase component
Satellite signals travel through a propagation channel
– analogous to the quadrature signal, defined in (21.3) with
which modifies its amplitude, phase‚ and delay. In fact,
subindex Q – hold:
multiple replicas of the same transmitted signal can reach

TbI

··· NcI 1 2 ··· NcI 1 ···

t
TPRNI = LcITcI

Figure 21.1 Relation among the parameters defining bits and spreading sequences in a generic navigation signal (in-phase component)
[17]. Source: Reproduced with permission of P. Closas.
21.2 Conventional GNSS Signal Processing and Positioning 531

a receiver’s antenna due to multipath propagation. In 21.2.2 Basics of GNSS Signal Processing
general, these replicas are caused by reflections of the
The primary goal of a GNSS receiver is to accurately esti-
line-of-sight (LOS) signal from surrounding obstacles (e.g.
mate its PVT, which is typically accomplished by first esti-
buildings, trees, ground). For the i-th satellite link, such a
mating the parameters in the signal model (Eq. (21.5)).
propagation channel is generically modeled by a linear
These estimates are used to construct a set of observables
time-varying impulse response with Ni propagation paths
per satellite. Observables are a set of ranges computed from
[18, 19]:
time-delay or phase-difference estimates, referred to as
Ni − 1
pseudoranges and carrier-phase measurements, respec-
hi t = αi,n t exp jϕi,n t δ t − τi,n t ,
n=0
tively. Once a set of valid observables is computed, a geo-
21 4 metrical problem is solved to compute the PVT solution
for the receiver.
where αi,n(t), ϕi,n(t), and τi,n(t) are the amplitude, phase‚ The received signal (Eq. (21.5)) is sampled at a suitable
and delay of the n-th propagation path for the i-th satellite. sampling rate fs = 1/Ts, resulting in a discrete-time signal
δ(t) is Dirac’s delta function [20]. Notice that subindex n = 0 x[n] = x(nTs) from which we aim to extract the observa-
denotes the LOS parameters. Typically, the amplitudes bles. The propagation time of a signal, the time between
are assumed to obey the Rayleigh or Rice distributions, its transmission from the i-th satellite to the user, is con-
depending on whether the LOS is blocked or not [21]. Time tinuously estimated by tracking algorithms. This time-
delays are assumed to be piecewise constant during the delay estimate (denoted by τi ) provides an estimation of
observation interval. In a first-order Taylor expansion, the distance between the i-th satellite and the user. Under
the time-varying phase ϕi,n(t) has a uniformly distributed the maximum likelihood (ML) principle, the unknown
independent term [22] and a linear term representing the parameters in Eq. (21.5) are obtained by maximizing the
effect of the Doppler spread (that is, a frequency shift cross-ambiguity function (CAF) of the received signal
caused by the relative motion between the satellite and and the PRN sequence of the i-th satellite. In general, the
the receiver). In general, we consider that paths are inde- CAF is computed over a coherent integration time, in
pendent, which is referred to as the Wide Sense Stationary which Nc PRN sequences are used to correlate with the
with Uncorrelated Scattering (WSSUS) [23] channel model. incoming signal. Additionally, Nnc CAFs can be averaged
Extensive measurement campaigns have been performed to over a non-coherent integration time to avoid bit transi-
characterize the GNSS propagation channel model, and tions. Here, we do not consider non-coherent integration
they support the WSSUS characterization; see, for instance, for the sake of simplicity, but some details can be found
the results reported in [24, 25]. in [2]. More formally,1 the i-th channel estimates [26] are
As a consequence, a GNSS receiver measures signals that defined as
are considered to be a superposition of plane waves cor-
2
rupted by noise and, possibly, interferences and multipath. Lc N c

An antenna receives M scaled, time-delayed, and Doppler- τi , f di = arg max x n ci nT s − τi exp − j2π f di nT s ,
τi , f di
n=0
shifted signals with known signal structure. Each signal
21 6
corresponds to the LOS of one of the M visible satellites
after propagating through the WSSUS channel mode. The where we recall that Lc is the length of the PRN
received complex baseband signal can be modeled as the sequence ci( ). For convenience, the CAF is typically
convolutions of the channel (Eq. (21.4)) with the corre- parameterized as Λi τi , f di , showing the corresponding
sponding signal, resulting in: satellite indicator and the relevant synchronization para-
M meters. More compactly, in vector form, we can rewrite
x t = ai si t − τi exp j2π f di t + n t , 21 5 Eq. (21.6) as
i=1
2 2
τi , f di = arg max Λi τ i , f d i ≜ x H ci
where si(t) is the transmitted complex baseband low-rate τi , f di
navigation signal spread by the pseudorandom code of 21 7
the i-th satellite, which is considered to be known. ai is
its complex amplitude, τi is the time delay, f di is the Doppler
shift, and n(t) represents zero-mean additive noise with var-
iance N0/2 and other unmodeled terms besides LOS. N0 is
1 We use the convention to denote the Euclidean norm of
the one-sided power spectral density in dBW/Hz. Note that
n-dimensional vector spaces in n and to denote the absolute value,
subindex n is omitted since only LOS signals are explicitly the latter coinciding with the norm on the one-dimensional
modeled, in contrast to Eq. (21.4). vector space.
532 21 Direct Position Estimation

where fc
f di = − v i − v ui , 21 10
c
x0
where v = (vx, vy, vz) and v i = vxi , vyi , vzi are the velocity
x= and
vectors of the user and the i-th satellite, respectively. ui
x Lc N c represents the unitary direction vector of the i-th satellite
relative to the user, defined as

ci − τ i pi − p
ui = 21 11
pi − p
ci =
and fc represents the corresponding carrier frequency used
ci Lc N c T s − τi exp − j2π f di Lc N c T s
by the GNSS. Doppler (and phase) estimates are used to
21 8 compute carrier-phase observables much more accurately
than the code observable described in Eq. (21.9).
correspond to the LcNc snapshots and the local replica,
Summing up, the baseband operation of current GNSS
respectively.
receivers is briefly described. First, the receiver detects which
The time-delay estimate, τi , is used to construct the
satellites are visible and obtains rough estimates of the time
so-called pseudorange ρi = cτi Pseudoranges provide a
delays and Doppler shifts of those satellites, expressed as
nonlinear relation between the user’s position (p = [x, y,
z] ) and the time delay of each satellite according to τ1 f d1
the model
τ= and fd = 21 12
ρi = ϱi p + c δt − δt i + cΔTi + cΔIi + ϵi , 21 9 τM f dM
where c is the speed of light, satellites are indexed by
respectively. This initial operation is referred to as acquisi-
i {1, …, M}, and the following definitions apply:
tion, and it can also be seen as an open-loop processing of

• τi is the time-delay estimate at the receiver for the LOS


emitted by the i-th satellite.
the data. Once the rough estimates are obtained, the receiver
can start operating in tracking mode. The tracking mode is

• ϱi(p) = pi − p is the geometric distance between the


receiver and the i-th satellite. pi = (xi, yi, zi) are the
typically implemented in a closed-loop (CL) architecture –
that is, with phase lock loops (PLLs) and delay lock loops
coordinates of the i-th satellite in the Earth-Centered (DLLs) – or an open-loop (OL) architecture in a challenging
Earth-Fixed (ECEF) coordinate system, which can be environment, to obtain the accurate time delay, Doppler
computed from the ephemeris. shift, and carrier-phase estimations and thus ultimately

• δt is the bias of the receiver clock with respect to GNSS


time, which is unknown.
acquire accurate user position solutions. Ultimately, acqui-
sition and tracking operations aim at solving Eq. (21.7).

• δti is the clock bias of the i-th satellite with respect to


GNSS time, known from the ephemeris.
Regardless of the receiver operating architecture (OL or
CL), conventional GNSS receivers estimate user position, p,

• ΔTi and ΔIi are the non-dispersive tropospheric delay


and the frequency-dependent ionospheric delay terms.
from the computed observables. In this second step, the posi-
tion of the receiver is computed upon taking into account the

• The term ϵi includes errors from various sources such as


multipath biases, ephemeris errors‚ and relativistic
geometrical relation between the set of observables and the
user PVT. The resulting multilateration problem is typically
effects‚ among others. solved by a least-squares (LS) algorithm. Note that in this
process the receiver also needs to estimate the receiver clock
The other important effect experienced by the signal is a bias δt that represents the offset between the receiver time
frequency shift. The observed carrier frequency at the and the GNSS time.
receiver differs from its nominal frequency due to Doppler In its most simple form, that is, the single point solution
effect, which is caused by user-satellite relative motion. with code pseudoranges, the LS solution appears from a lin-
Additionally, frequency shifts due to clock drifts (δt ) and earized geometrical problem [6, 26, 27]. Such linearization
background errors are typically experienced on top of the Dop- makes perfect sense, given that the nonlinearity is, in
pler shifts. Accurate Doppler estimates yield precise velocity essence, a sphere whose radius is on the order of tenths
calculations that are useful in positioning and navigation of thousands of kilometers. Given that GNSS satellites orbit
applications with high user dynamics. The Doppler shift between 15 000 and 30 000 km, and given that a GNSS
due to the relative motion of the user and the i-th satellite is receiver tracking PRN code chip transitions must resolve
21.3 DPE 533

uncertainties on the order of 100s of meters, the geometric the de facto technique for GNSS receivers. This is due to its
localization problem may be treated as linear for all but the modularity, reuse of well-known receiver blocks, and,
most precise applications. The problem is to compute the importantly, notable performance over the years. The two-
user’s position and clock offset from a set of M estimated step technique will be referred to as the conventional
pseudoranges. Thus, from Eq. (21.9), we form the following approach throughout the remainder of the chapter.
system of equations:
ρi + cδt i − ϵi = pi − p + cδt
21 13 21.3 DPE
i 1, …, M M≥4 ,
which results in a nonlinear and possibly overdetermined
Proposals for new techniques are blooming due to the
system. The condition M ≥ 4 is due to the dimensionality
advances in digital signal processing devices, which allow
of the problem. The system is usually solved iteratively
increased computational complexity at faster rates [29]. Over
by linearizing each ϱi(p) with respect to an initial position
the last years, advanced receiver techniques have been pre-
estimate (po = (xo, yo, zo) ) and iterating until convergence:
sented which substitute well-established components of
xi − xo y − yo zi − zo two-step receivers by more sophisticated algorithms. These
ϱi p ϱoi + δx + i o δy + δz ,
ϱi
o ϱi ϱoi methods do not modify the architecture of the receiver, as
21 14 depicted in Figure 21.2(a), where acquisition and tracking
operations are as discussed in Section 21.2. For instance, sub-
where δx = x − x, δy = y − y, δz = z − z and
o o o
stitution of tracking loops by more versatile filtering solu-
ϱoi ≜ ϱi po = pi − po The Bancroft algorithm [28]
tions can be consulted in [30]. In this chapter, we are
provides an initial guess as to the position and the clock off-
interested in those approaches that not only substitute cer-
set of the user receiver without any prior knowledge.
tain parts of the receiver by more advanced algorithms,
Considering Eq. (21.14), the system in Eq. (21.13) can be
but those which entail an essential modification of the
formulated as the following LS problem:
receiver’s operation. Figure 21.2(b) shows an advanced
δ = arg min y − Tδ 2
, 21 15 architecture for a receiver which performs code/carrier
δ
estimation and position computation in a single step. This
where δ = δx , δy , δz , δt , is the basis of the so-called DPE concept in the context of
GNSS receivers [1]. It is worth noting the following here:
ρ1 + cδt 1 − ϵ1 − ϱo1
y=
ρM + cδt M − ϵM − ϱoM
,
• In DPE, an absolute initial position (or PVT, to be more
general) estimate has to be provided so that the method
can start exploring the space. This initialization can be
x1 − xo y1 − yo z1 − zo
1 either provided by an initial two-step process, as depicted
ϱo1 ϱo1 ϱo1 in Figure 21.2(b), where a coarse estimation of synchro-
T= , nization parameters is used to compute a rough naviga-
xM − x o
yM − y o
zM − z o tion solution that is used as the initial estimate in DPE’s
1
ϱoM ϱoM ϱoM algorithm, or provided by external means such as cellular
positioning or assistance GNSS data [31].
and the solution
−1
• Whereas tracking mode is typically implemented in CL
(through banks of parallel PLLs/DLLs) for two-step
δ = T† y ≜ TH T TH y 21 16 approaches, OL schemes are also explored in the context
of snapshot receivers. Conversely, DPE’s tracking mode
is straightforwardly given by the Moore–Penrose pseudoin- is in general implemented in OL, which involves the opti-
verse (T†). Therefore, we have that p = po + δ is the classi- mization of a cost function. This is the approach mainly
cal position estimation provided by GNSS receivers. Notice discussed in this chapter. However, CL schemes can also
that, once a new position is obtained, it can be used as ini- be considered within DPE. This is mentioned in
tialization point for linearization iteratively until conver- Section 21.4.3, where references are provided and a
gence. This is the simplest way of formulating the PVT PVT filtering described.
solution as a LS optimization problem, which can be
enhanced by incorporating side information, measurements DPE is a completely different approach to the receiver
weighting, and dynamics, for instance, using solutions based design, where the PVT solution is estimated in one step.
on weighted least-squares (WLS) or Kalman filtering techni- Rather than estimating a set of observables (through time
ques [17]. This positioning approach has established itself as delay and Doppler shifts) to infer the associated PVT, a
534 21 Direct Position Estimation

(a) Two-step approach

Acquisition υˆc Tracking υ̂f Navigation


γ̂
(OL) (OL/CL) Solution

x[n]

(b) DPE approach(self-initialized)

Position γ̂c DPE (OL/CL):


arg max { Λ (γ) } γˆ
initialization
γ

x[n]

Figure 21.2 Acquisition and Tracking schemes for (a) two-step and (b) DPE positioning approaches. Coarse estimates of synchronization
parameters are referred to as υc, which are typically the outcome of the acquisition stage, whereas fine estimates are provided by tracking
loops, υ f . Similarly, γ c represents coarse PVT estimates and γ fine PVT estimates.

DPE-enabled receiver directly estimates PVT from the view, υi, individually. On the other hand, DPE defines a set
received signal x[n] (hence the name). The method was first of candidate positions, determines the time delays associ-
introduced in the context of localization of narrowband ated with the positions, and computes the energy found
radio frequency transmitters [32] and for multiple radio sig- at the different correlation outputs jointly. In a way, the
nals [33]. In the context of self-localization, [1] presented local replica is generated by jointly combining all signals.
the approach for GNSS receivers, and related works exist Then the tentative position that jointly maximizes correla-
in the collective detection literature [34–37]. tion with all considered satellites is selected through opti-
The key idea is to realize that time delays and Doppler mization of a cost function. The cost function appears as
shifts of all satellites are intimately related to one another a result of deriving the ML estimator of the PVT parameters
through the receiver PVT parameters (which we gather directly from the received signal (Eq. (21.5)), which maxi-
in a vector γ Rnγ ). Indeed, the fact that all those signals mizes the addition of autocorrelation functions over the
are received at the same location and at the same time M visible satellites [17, 38].
instant is crucial. Inspecting Eqs. (21.9) and (21.10), we
can readily identify that, actually, τ ≜ τ γ and f d ≜ f d γ ,
respectively. Therefore, a value for γ implies a value for 21.3.1 Qualitative Motivations for Direct-
all satellites’ time delays and Doppler shifts. One of the sim- Positioning
plest configurations for γ could be Although it is agreed that conventional two-step position-
ing has multiple benefits, some of its limitations have been
γ = p , δt, v , δt , 21 17
pointed out [6], which are precisely exploited by position-
encompassing PVT parameters. In contrast, the classical domain techniques [39]. As mentioned earlier, thanks to
solutions in Figure 21.2(a) aim to independently estimate the advances in computation and software-defined radio
the CAF parameters for each channel. That is, Eq. (21.6) [29], some of the technological limitations that prevented
technology from going beyond two-step positioning have
is optimized over τi , f di . We can define the joint vector
been overcome. Some of these qualitative drawbacks or
of synchronization parameters as
limiting factors are as follows:


υ ≜ υ1 , …, υM with υi = τi , f di 21 18
Estimates of the synchronization in Eq. (21.18), υ, are
Intuitively, one can see DPE as a direct approach for esti- obtained independently per satellite, omitting any potential
mating γ without individual channel estimates. On the one correlation among channels and propagation effects. The
hand, the conventional approach estimates the different errors are propagated to the LS solution, where a statistical
time delays by maximizing the correlation between the sig- characterization is deemed challenging due to nonlineari-
nal and the conditioned PRN sequence of each satellite in ties. Therefore, the two-step approach does not take into
21.3 DPE 535

consideration possible dependencies among channels, This section provides additional motivations, quantitative
which could improve the estimation performance. in this case, to support DPE’s potential. First, we discuss

• Related to the previous point, a conventional GNSS


receiver may require 10-20 parallel channels in order to
a result showing that DPE outperforms two-step positioning
in the mean square error (MSE) sense. Second, we discuss
process LOS signals and actively search for new signals some results on the theoretical lower bounds of positioning
previously obscured by structures or the horizon. In case accuracy, comparing both two-step and DPE approaches.
of multi-constellation receivers, the number of channels
is even larger. Since the ultimate parameter of interest is 21.3.2.1 DPE Outperforms Two-Step Positioning
the PVT solution (i.e. position, velocity‚ and time defined The following result provides a mathematical justification
in a three-dimensional space, γ), it seems that there exists for the fact that DPE provides asymptotically lower, or
a possible redundancy because the receiver performs esti- equal, MSE than two-step positioning. In other words,
mation in a higher-dimensional parameter space (delay the result proves that the covariance of the two-step
and Doppler for each channel, totaling 2M unknowns). approach cannot be smaller than the covariance of the
Thus, the conventional approach increases the dimen- one-step estimator [42, 43]. Thus, the estimation perfor-
sionality of the problem with respect to the amount mance of the conventional approach can only be, at most,
strictly (and theoretically) required.


equal to the one provided by the DPE approach. This result
The use of prior information is not straightforward when is heavily based on [42].
dealing with synchronization parameters. In the two-step Let υ υ Rnυ and γ γ Rnγ be two unknown
method, as mentioned in Section 21.2.2, side information parameters such that there exists an injective function
is typically introduced in the navigation solution once g γ ↦ υ,
observables have been computed. Using prior information
in the tracking loops is, in general, an involved task [40] υ=g γ , γ γ 21 19
and would require extensive test-field campaigns to pro- that relates them. In our case, function g( ) is given by the
duce the relevant data [25], and the algorithm would need PVT-observables relationship in Eqs. (21.9)–(21.10). Func-
to distinguish among a number of synchronization evolu- tion g( ) has a unique inverse mapping
tion models depending on the dynamics of the receiver.
This is much more difficult than considering prior infor- γ = g−1 υ , υ υ 21 20
mation when the parameter of interest is the user’s
under the subset υ = υ υ = g γ , γ γ υ.
position itself, γ, where the physical meaning of the param-
The K-sample estimators of γ based on the single-step and
eter aids the inclusion of side information.

• Although the cross-correlation properties of spreading


sequences used in GNSS signals provide a rather high
two-step approaches are denote by γ DPE and γ 2S , respec-
tively. Similarly, Σ γ DPE and Σ γ 2S represent the covari-
ance matrix of each estimator.
processing gain, there is indeed a remaining multiple
Then,
access interference (MAI) that is not combated in con-
ventional receivers. This issue becomes critical when C ≜ lim Σ γ 2S − Σ γ DPE 21 21
K ∞
the received power levels for the satellites are highly
unbalanced. To overcome this limitation, one could is a positive semi-definite matrix, which proves the main
incorporate multiple access techniques to GNSS recei- result. The complete proof can be consulted in [17, Appen-
vers, jointly processing signals from different satellites dix 4.A], where the details are provided.
[41], which is inherently achieved by DPE [2].
21.3.2.2 A Note on Position Estimation Bounds
Direct positioning aims at addressing these shortcomings
In the context of positioning (i.e. self-localization), results
by jointly processing signals from all satellites at the sample
exist where the Cramér–Rao lower bound (CRB) for PVT
level. The remainder of this section discusses quantitative
estimation is derived for both positioning approaches, high-
reasons for the superiority of DPE when compared to the
lighting the potential benefits of adopting DPE [44, 45].
two-step approach. Then, we will derive the optimal posi-
More recently, [46] provided additional results regarding
tion estimator, under the ML principle, which shares a sim-
MSE performance bounds that yield a better understanding
ilar structure as the two-step synchronization solution
of DPE and its potential. These bounds are based on the Ziv-
discussed in Eq. (21.6).
Zakai methodology. The so-called Ziv-Zakai bound (ZZB)
provides a bound on the MSE over the a priori probability
21.3.2 Quantitative Motivations for Direct-
density function. The generic derivation of the bound can
Positioning
be found in [47–49]. The ZZB allows us to determine, for
In previous section‚ we discussed some qualitative reasons instance, the signal-to-noise ratio at which both DPE and
for adopting DPE, as opposed to the two-step approach. two-step approaches are able to operate before breaking
536 21 Direct Position Estimation

down (that is, before entering the large-error region). It turns we have that
out from this analysis that, under certain scenarios, there is
an increase in sensitivity for DPE on the order of 10log10(M) M M
SNRi SNRi
dB, M being the total number of used satellites [46]. 2Q 0 and Γ3 2
i=1
2 i=1
4
Although this effect was already observed in earlier works,
derivation of the new bound allows analytic interpretation 1
of the enhanced performance of DPE. Particularly, the lower
MSE bound when estimating γ is 21 28

that causes the bound to reach, in the asymptotic region,


M M
SNRi −1 SNRi the FIM of γ
ZZB γ = Rγ 2Q + γ Γ3 2
2 4
i=1 i=1 −1
ZZB γ γ , 21 29
SNR ∞
21 22
which corresponds to the CRB expression. More precisely,
where Rγ is the a priori covariance matrix of γ, which models
γ = PT τ P with P = 1c u1 , …, uM being the deriva-
the degree of a priori uncertainty about the parameter; γ
is the Fisher Information Matrix (FIM) of γ; Q x = tive of τ with respect to γ. The FIM of τ can be written as

1 2π x exp − t 2 2 dt is the Q-function, expressed τ = β2s Γ, where Γ = diag(SNR1, …, SNRM) is a diagonal
M × M matrix with its diagonal elements given by the M × 1
in terms of the complementary error function as ∞
s t 2 dt
Q x = 1 2 erfc x 2 ; and Γa(x) is the incomplete SNR vector‚ and β2s = −∞∞ is the mean quadratic
2
gamma function given by − ∞ s t dt
bandwidth of s(t).
x
1
Γa x = e − v va − 1 dv 21 23 For the sake of completeness, we provide here some
Γa 0 insights on the ZZB of position estimates when two-steps
and Γ 3 2 = π 2 . The signal-to-noise ratio (SNR) per is considered. The ZZB for the time-delay estimation prob-
a2i lem was first given in [50], and then generalized for any
satellite is defined as SNRi =
N0 2
, assuming unitary a priori distribution of τi in [48]. The ZZB for estimating
energy for si(t). This result allows us to analyze the bound each individual time delay, τi, can be readily obtained fol-
in the high- and low-SNR regimes. At low SNRs, lowing that methodology and expressed in compact form
split in the low- and high-SNR terms:
M
SNRi 0, 21 24 2 SNRi −1 SNRi
E τi − τ i ≥ σ 2τi 2Q + τ i Γ3 2
i=1 2 4

we have that ≜ ZZB τi


21 30
M
SNRi Then, the position MSE bound conditioned to using a
2Q 1 and
2 −1
i=1
21 25 two-step scheme turns out to be PT ZZB − 1 τ P , where
M
SNRi ZZB τ = diag ZZB τ1 , …, ZZB τ M is a diagonal M × M
Γ3 2 0 matrix constructed with the individual bound expressions
4
i=1
for each satellite. The a priori variance of τi is represented
by σ 2τi and (τi) is the FIM of τi [45]. In specific situations,
such that, asymptotically, the bound is dominated by the a
priori covariance matrix for instance‚ when the satellites are received with the same
power, one can show analytically the log10(M) dB improve-
ZZB γ Rγ 21 26 ment of DPE with respect to the two-step approach. A more
SNR 0
general result can be obtained [46, 51] showing that
−1
Conversely, in the high-SNR regime, ZZB γ − PT ZZB − 1 τ P is a positive semi-definite
M
matrix, although the improvement cannot be computed
SNRi ∞, 21 27 in closed form in the general case and can only be shown
i=1 through Monte Carlo simulations.
21.3 DPE 537

21.3.3 ML Estimation of Position: Direct concatenation of the M local codes, defined in Eq. (21.8),
Positioning resulting in
Parameter estimation methods have been thoroughly cov- C γ = c1 , …, cM CK × M 21 34
ered in the literature [52]. The ML principle presents an
We first take into account that the MLE is equivalent to
optimal paradigm to obtain a parameter estimator that
the solution obtained by an LS criterion under the assump-
asymptotically attains its lower variance bound (i.e. the
tion of zero-mean AWGN. Neglecting additive and multi-
bounds discussed earlier), as the number of samples goes
plicative constants, maximizing the likelihood function
to infinity. ML is based on maximization of the likelihood
defined by Eq. (21.33) is equivalent to minimizing the fol-
function, constructed from the conditional probability of
lowing cost function:
measurements given a parameter value. In general, for a
set of K recorded samples gathered in x, the maximum like- 2
Λ a, γ = x − C γ a 21 35
lihood estimator (MLE) of a parameter γ is the solution to
with respect to a, γ. Expanding Eq. (21.35),
γ ML = arg max
γ
px γ 21 31
Λ a, γ = xH x − xH C γ a − aH CH γ x + aH CH γ C γ a
Under mild regularity conditions, the asymptotical distri- 21 36
bution (for large data sets) of the estimator satisfies
and taking the derivative with respect to aH, we obtain
−1
γ ML γ, γ , 21 32 ∂Λ a, γ
= − CH γ x + CH γ C γ a 21 37
∂aH
where γ is the FIM evaluated at the true value of the
parameter. which equated to zero gives us the desired estimation of
Thus, the claim is that the MLE is asymptotically effi- complex amplitudes:
cient; that is, it attains the lowest variance predicted by −1
a = CH γ C γ Cγ x 21 38
the CRB as K increases or the SNR is sufficiently high. γ=γ

The regularity conditions include the existence of the deri- The first term, the one in the inverse, can be simplified
vatives of the log-likelihood function and the FIM being taking into account the cross-correlation and autocorrela-
nonzero. For further details, refer to [52, Appendix 7.B]. tion properties of the PRN codes used in GNSS, although
For the purposes of this chapter, we are interested in deriv- that simplification might not be accurate in binary offset
ing the ML of position, γ , in a GNSS receiver due to the carrier (BOC) modulations, which feature (multiple) sec-
aforementioned asymptotic (and desirable) properties. ondary lobes. That is, we consider that the normalized
Interestingly, the MLE of position [53], which is derived CAF is such that
from the likelihood of x in Eq. (21.5), does not necessarily
coincide with the position estimation computed in a two- i c j ≈ 0 if i
cH j 21 39
step position. cH
i ci = 1 21 40
To derive the MLE of position from Eq. (21.5), we first
express the observations in vector form. Assuming that K to simplify
samples are processed and that the integration time is such
cH
1
that no bit transitions exist (or they can be removed by
other means), we can write x = C υ a + n, or equivalently CH γ C γ = c1 , , cM

x = C γ a + n, 21 33 cH
M
cH
1 c1 cH
1 c2 cH
1 cM
where x ℂK × 1 is the observed signal vector; a ℂM × 1 is a
cH
2 c1 cH
2 c2 cH
2 cM
vector whose elements are the complex amplitudes of the M
= ≈I RM × M
received signals a = (a1, …, aM) ; υ R2M × 1 gathers τ and
fd, that is, time delays and Doppler shifts of each visible sat-
ellite. However, in DPE we make use of the re- cH
M c1 cH
M c2 cH
M cM

parameterization υ ≜ υ γ ; n ℂK × 1 represents K snap- 21 41


shots of zero-mean additive white Gaussian noise and rewrite Eq. (21.38) as
(AWGN) with piecewise constant variance σ 2n = N 0 2 dur-
ing the observation interval; and C γ is composed of the a = C γ x γ=γ 21 42
538 21 Direct Position Estimation

Substitution of a in Eq. (21.36) yields a cost function that 21.3.4 Some Intuitions
depends on γ
We conclude this section providing some additional quali-
Λ a, γ = x x − x C γ C
H H H
γ x 21 43 tative intuition as to why DPE outperforms the two-step
approach. We saw that, theoretically it does, and that the
and the MLE of position is given by cost function is constructed such that the effective SNR is
augmented. In this subsection‚ we carry out a small exper-
γ = arg min Λ a, γ 21 44
γ iment comparing the cost functions of both two-step and
DPE approaches to observe how they are affected by a mul-
= arg min xH x − xH C γ CH γ x 21 45
γ tipath disturbance. For DPE, the cost function is plotted
2 with only x and y coordinates as variable terms and fixing
= arg max xH C γ CH γ x = arg max xH C γ ,
γ γ the rest of elements in γ to their true values. In particular, a
21 46 constellation of M = 6 GPS satellites minimizing the geo-
metrical dilution of precision is simulated, all received with
which can be further manipulated to obtain a more com- a carrier-to-noise density ratio (C/N0) of 45 dB-Hz and
pact expression that depends on the individual CAFs asso- transmitting the C/A code signal. At the receiver, the signal
ciated to each satellite is bandpass-filtered at 2 MHz, and an integration time of
M 1 ms is configured. Note that, for the sake of a better visu-
2
γ = arg max x H ci γ 21 47 alization, the cost function is inverted and normalized in
γ
i=1 the plots, turning the optimization into the minimization
in Eq. (21.45). The plots in Figure 21.4 represent the cost
To obtain that expression, we used
functions for both positioning approaches as a function
xH C γ = xH c1 , …, cM = xH c1 , …, xH cM of the error on the corresponding parameters. In benign
situations, both functions have a clear global optimum.
21 48
In contrast, when a multipath replica is present for one
Note that the resulting cost function in the DPE approach of the satellites the behavior of the cost function differs.
is the non-coherent addition of the CAFs from the M In particular, Figure 21.5 is generated by adding an echo
satellites, M ≜ M
2 2 for one satellite (3 dB lower than the LOS). Whereas DPE’s
i = 1 x ci γ i = 1 Λi γ = Λ γ , virtually
H

augmenting the effective SNR proportionally. Additional cost function remains virtually unaltered by this effect, two-
coherent/non-coherent integrations can be considered to step’s cost function exhibits a strong secondary optimum
extend Eq. (21.48), as described in [2]. An example of the due to the presence of a correlated signal. Furthermore,
shape of Λ γ is shown in Figure 21.3. Since the MLE esti- the two-step solution can be trapped in such a local opti-
mator is consistent, the cost function is optimized at the mum, yielding to potentially large biases in range and
true PVT solution. It was shown that DPE provides addi- velocity estimation. Considering that position is jointly esti-
tional robustness [54, 55] in challenging cases. mated with the information of all visible satellites, a sort of
diversity is introduced in this estimation as the propagation
path for each satellite link is independent.

highest vector correlation vector correlation


MLE solution amplitude
21.4 Implementation Aspects

We have seen in Section 21.3 that the DPE’s solution


expected signal reception at
involves an optimization problem that cannot be solved
a navigation candidate in closed form. In particular, each candidate solution γ
actual signal reception
at GPS antenna requires evaluation of a certain number of correlator out-
)
rth puts, which is known to be computationally intensive. This
(No
on
position
(E siti section discusses some approaches when it comes to imple-
ast) po
menting DPE on a software-defined radio (SDR) receiver.
Figure 21.3 The manifold of the vector correlation amplitudes Figure 21.6 depicts this situation. DPE performs a direct
Λ γ , shown in the east-north position domain [54]. Source: search, with inherent joint optimization across multiple
Reproduced with permission of Institute of Navigation. satellites, for the navigation solution on the signal
(a) The two-step approach (b) DPE

1 1

0.8 0.8

0.6 0.6

Λ (γ)
Λ (υ)

0.4 0.4

0.2 0.2

0 0
2 1000
1 4000 500 1000
0 2000 0 500
0 –500 0
–1 –2000 –500
τi – τˆi [Tc] –2 –4000 fdi – fˆdi [Hz] y – yˆ [m] –1000 –1000 x – xˆ [m]

Figure 21.4 Comparison of cost functions for two-step and DPE optimization problems [2]. In the two-step approach, it is a function of
the synchronization parameters of the locally generated code of the i-th satellite, υ. In DPE, the cost function is parametrized by the PVT
parameters gathered in γ. Source: Reproduced with permission of IEEE.

(a) The two-step approach (b) DPE

1 1

0.8 0.8

0.6 0.6
Λ (γ)
Λ (υ)

0.4 0.4

0.2 0.2

0 0
2 1000
1 4000 500 1000
0 2000 0 500
0 0
–1 –500
–2000 –500
τi – τˆi [Tc] –2 –4000 fdi – fˆdi [Hz] y – yˆ [m] –1000 –1000 x – xˆ [m]

Figure 21.5 Comparison of cost functions for two-steps and DPE optimization problems in the presence of a replica for one of the
satellites [2]. The echo has 3 dB less power than the corresponding LOS. In two-steps, a strong secondary optimum appears, whereas the
DPE cost function is virtually unaltered. Source: Reproduced with permission of IEEE.

DPE
solution

initialize multiple correlate expected signal


candidates in the receptions at candidate perform MLE
navigation domain states with received signal on correlations

Figure 21.6 Overview of the evaluation of candidate points of DPE. The grid points represent PVT candidates initialized in the navigation
domain; the color saturation of the grid points represents cost function magnitude [58].
540 21 Direct Position Estimation

correlations [56–58]. The basic steps, found in Figures 21.2 low duty-cycling, performing DPE updates only a fraction of
(b) and 21.6, to implement DPE on a receiver would be the time instead of calculating it continuously.

1) To provide some sort of initialization, which can be pro-


21.4.1 Grid Optimization by Divide and Conquer
vided either by using two-step schemes or by exter-
nal means. Divide and conquer [66] is an algorithmic approach
2) The second step is optimization of the cost function in adopted in many signal processing applications, which
Eq. (21.48). This process involves the selection of candi- breaks down a complex problem into two or more subpro-
date points for γ, which results in some τi , f di candi- blems that are easier to solve. In addition, this methodology
date pairs for each visible satellite. The τi , f di allows the exploitation of parallel structures in hardware
candidate pairs are then used to generate the expected architectures, and makes use of concurrency and multi-
signal reception at candidate states. Next, correlations threading scheduling when designing under the SDR para-
are computed between the local replicas at candidate digm. This is one of the reasons for adopting two-step
states and the received signal, and Λ( ) is evaluated. architectures in GNSS receivers: independent per-satellite
3) The third step is to perform MLE through optimizing the channel track delay-Doppler maps, which are then fused
cost function. These steps are performed each time new in an LS solution as detailed in Section 21.2.
measurements are available. In the context of DPE positioning, a similar approach can be
adopted in selecting the candidate points. It might be desira-
The main challenge of implementing DPE is the compu- ble to split γ into sets of correlated parameters. One
tational load. Related work on improving DPE’s computa- way to reduce the number of candidates is to group the
tional efficiency includes use of sparsely located navigation navigation parameters into two subsets: position/clock bias
candidates [59], reducing the number of required naviga-
γ 1 = p ,δt and velocity/clock drift γ 2 = v ,δt [3, 67].
tion candidates [3, 53, 60], efficient techniques for calculat-
ing multiple correlations [61], and efficient techniques for Co-estimation of γ 1 and γ 2 reduces the number of
estimating the navigation solution given the vector correla- navigation candidates from N 8c to 2N 4c, in the case of Nc can-
tion distribution [59]. To improve the efficiency of initiali- didate points for each of the eight navigation search
zation, coarse-grid search [59] and initialization using dimensions.
Assisted-GNSS (A-GNSS) [62] were proposed. Estimation of γ 1 is relatively insensitive to departure from
To improve the efficiency after initialization near the the γ 2 search center and vice versa [3]. This is because the
main vector correlation peak, optimization over navigation correlation amplitude with respect to γ 1 is maintained over
subsets was proposed. One such implementation used a relatively broad range; the same is true for the correlation
parameter grouping by position/clock bias and velocity/ amplitude with respect to γ 2. Thus, the estimation of γ 1 and
clock drift [3, 59]. It has been shown that, according to this γ 2 parameters can be safely decoupled
decoupling, parameters within subsets are strongly corre- γ 1 = arg max Λ γ 1 , γ 2 21 49
γ1
lated‚ while parameters across subsets are weakly corre-
lated. The subsets can then be separately optimized. γ 2 = arg max Λ γ 1 , γ 2 21 50
γ2
Other subset optimization implementations used parame-
ter grouping where each parameter formed its own subset When decoupled, the CAF’s amplitude Λi p, δt, v, δt
[63, 64], which can be solved algorithmically through for varying γ 1 candidates with respect to a γ 2 fixed near the
space-alternating generalized expectation-maximization γ 2 search center is given in Eq. (21.51). The normalized
(SAGE) [53]. These subset optimization methods reduce
the number of required navigation candidates and thus correlation amplitude Λi p, δt, v, δt for varying γ 2 can-
the overall computational load. However, they are sensitive didates with respect to a γ 1 fixed near the γ 1 search center is
to initialization [53], which can be addressed using given in Eq. (21.52).
stochastic optimization methods such as accelerated ran-
1− Δτi , 0 ≤ Δτi ≤ 1
dom search (ARS) [60]. To increase the chances of initial- Λi p, δt, v, δt ≈
0, otherwise
ization near the main navigation correlation peak for
subsequent DPE measurement updates, motion model fil- 21 51
tering can be performed between DPE measurement Λi p, δt, v, δt ≈ sinc π Δ f di ΔT , 21 52
updates [3, 65].
In this section, we present some DPE implementation where Δτi is the delay error between the candidate delay
techniques in detail. The remainder of this section discusses τi γ 1 and the search center. Similarly, Δ f di represents
(i) subset optimization strategies, (ii) faster local replica gen- the frequency error between candidate f di γ 2 and the
eration and correlation through fast Fourier transforms search center for γ 2 . Note that these ideas resemble those
(FFTs), (iii) CL tracking schemes, and (iv) how to implement of DLLs and PLLs in standard receiver architectures.
21.4 Implementation Aspects 541

ˆ
ˆ δt ˆ δt
correlations across candidates p, ˆ
correlations across candidates p,

normalized amplitude

normalized amplitude
width = 1 code chip width = 1/ΔT

Δτi = 0 Δfd = 0
i

Figure 21.7 Shape and width of (left) correlation amplitude Λi p, δt, v, δt for varying position/clock-bias candidates with respect to a

fixed velocity/clock-drift; (right) correlation amplitude Λi p, δt, v, δt for varying velocity/clock-drift candidates with respect to a fixed
position/clock-bias [58].

The expression for the CAF in Eq. (21.51) assumes suc- Λi p, cδt, v, cδt = −1 ∗
x ci p, cδt, v, cδt
cessful navigation bit and carrier wipe-off of the received
signal near the γ 2 search center. Likewise, the expression 21 53
in Eq. (21.52) assumes successful navigation bit and code
wipe-off near the γ 1 search center. The shape and width Batch pre-processing using FFTs more efficiently calcu-
of the two CAFs, where one of the sets of parameters is lates multiple correlations. The batch pre-process is a result
fixed, is illustrated in Figure 21.7. of Eq. (21.53). To form the vector correlation, we first calcu-
late the τ γ associated with each candidate p, cδt. We then
assign the FFT results to the navigation candidates, using
21.4.2 Batch Pre-Processing of Correlations linear interpolation for τ γ values between FFT points.
Using Fast Fourier Transforms In addition, these operations can be performed in parallel
As in two-step positioning, one can exploit the circular cor- for all navigation candidates. Similarly, the correlation
relation properties of the PRN codes to speed up correlation Λi p, δt, v, δt for varying velocity/clock-drift candidates
through the FFT [27]. The correlation Λi p, cδt, v, c δt for
with respect to a fixed position/clock bias is approximated
varying position/clock-bias candidates with respect to a ∗
as the Fourier transform x ci p, cδt, v, c δt .
fixed velocity/clock drift in Eq. (21.51) can be implemented
in the frequency domain. Let us use ( ) to represent the The above computations are carried out in two parallel
Fourier transform operator and −1( ) the inverse Fourier threads as in Figure 21.8, one for correlations and the other
transform. Then, for the Fourier transforms.
Normalized correlation amplitude

Normalized spectrum magnitude

1.0 1.0

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
–300 –200 –100 0 100 200 300 –300 –200 –100 0 100 200 300
code phase residual Δϕ (m) carrier frequency residual Δf (m/sec)

Across N satellites for the kth receiver

Figure 21.8 Correlation amplitude with respect to code residuals (left) and magnitude of the spectral density with respect to carrier
residuals across satellites (right) [68]. Source: Reproduced with permission of Institute of Navigation.
542 21 Direct Position Estimation

21.4.3 Filtering of DPE’s PVT Solution matrix, which incorporates possible knowledge regarding
the motion dynamics of the receiver. For instance, given
The implementations considered so far operate in OL. To
that ΔT is the time in seconds between two consecutive
improve the accuracy of DPE, CL tracking of the navigation
instants k and k + 1, we can use
solution can be implemented, as suggested in [3, 17, 64,
69–72]. We mentioned earlier that DPE is a suitable frame- I4 ΔTI4
Fk + 1 = 21 60
work to introduce a priori information about γ that 04 I4
one might have [73]. In the context of CL schemes, this
0 0 0 0 0 0 0 0
is particularly relevant since dynamics can be exploited.
0 0 0 0 0 0 0 0
A methodology for how this can be actually implemented
using nonlinear filtering is described in [69]. 0 0 0 0 0 0 0 0
A simpler, yet effective, alternative is to smooth DPE 0 0 0 0 0 0 0 0
Qk = F k + 1 FTk + 1
results using a low-pass filter. For instance, implementa- 0 0 0 0 σ 2v 0 0 0
tion with an optimal Kalman filter (KF) allows for easy 0 0 0 0 0 σ 2v 0 0
incorporation of motion models and integration of informa- 0 0 0 0 0 0 σ 2v 0
tion from prior measurements into the estimation process 0 0 0 0 0 0 0 σ2
[73]. The KF has two main steps: a measurement update δt

step and a time update step [74, 75]. 21 61


Recall that, at the discrete time instant k, the observations where I4 and 04 represent the identity and all-zero square
of this KF are the estimated PVT parameters from DPE’s matrices of dimension 4, respectively.
optimization process, which we denote by yk in the follow- The dynamic process noise model captures the intuition
ing discussion. These are noisy measurements of the true that it is more difficult for a vehicle’s velocity to change
PVT, whose likelihood is assumed normally distributed as when its speed is large as compared to when its speed is
small. Thus, vehicle acceleration, modeled as noise, is lower
yk γk γ k , Rk , 21 54 when the vehicle’s speed is large and higher when the vehi-
cle’s speed is small.
where, according to Eq. (21.54), we know that Rk asymptot-
ically reaches the FIM of the parameters of interest, − 1 γ .
Therefore, the measurement update at time k is given by
21.4.4 Low Duty-Cycling
Low duty-cycling performs DPE for a fraction of the time
γ k k = I − Kk γ k k − 1 + Ky k 21 55
instead of calculating constantly. Duty-cycling DPE is par-
Σ k k = I − Kk Σ k k − 1 21 56 ticularly promising to mitigate the high computation cost of
DPE. Instead of computing DPE-based solutions for all
−1
Kk = Σ k k − 1 Σ k k − 1 + R k , 21 57 integration times, some are skipped for the sake of saving
computations. In other words, a DPE-based receiver can
where the state γ k k is estimated along with its covariance
operate intermittently, which may not work so well for
Σk k taking into account the new observations yk, and tracking loop-dependent conventional receivers. Research
the predicted stated and covariance, γ k k and Σk k, respec- has demonstrated viable DPE operation even at quite low
tively. In these equations‚ I represents the identity matrix. duty cycle rates [80]. The duty-cycled DPE receiver archi-
Those predictions incorporate historical information regard- tecture is shown in Figure 21.9. The measurement update
ing previous observations. This scheme integrates prior nav- uses optimization over navigation parameter subsets,
igation solutions into the estimation process [70]. combines and computes batch signal replica generation
In the time update step, the next receiver state γ k + 1 k is and correlation using FFTs, and estimates the navigation
predicted using, for instance, a constant acceleration solution. The exact computation time depends on signal
motion model [3, 65, 76]. This approach incorporates vehi- sampling parameters and the number of satellites in view.
cle accelerations using a dynamic process noise model on During the time required to compute the DPE measure-
the velocities [77–79]. Then, ment update, we iteratively predict satellite movements,
using the satellite broadcast ephemerides, and update
γ k + 1 k = Fk + 1 γ k k 21 58
the signal code phase and carrier Doppler frequency
Σk + 1 k = Fk + 1 Σk k FTk + 1 + Qk , 21 59 parameters. This DPE time update reduces the accumula-
tion of signal tracking errors during the time in which
where Σk + 1 k is the predicted state error at time k + 1, and the DPE receiver is not performing any DPE measurement
Qk is the dynamic process noise. Fk + 1 is the propagation update.
21.5 DPE Variants 543

Measurement update process Time update process

Initialize over yes no


navigation subsets Switch on?

Batch replica generation Correct signal state


20 ms and correlation using
snippet FFT and linear
interpolation Get satellite PVT

MLE Get satellite


PVT
using weighted ave. transmit time

Correct receiver state Increment receiver state

Correct signal state Increment signal state

Figure 21.9 Block diagram of duty-cycled DPE receiver architecture [79]. Source: Reproduced with permission of Institute of Navigation.

20 ms 1s
... on DPE time update on DPE time update on ...

DPE measurement update DPE measurement update DPE ...

PVT PVT PVT PVT

Figure 21.10 Overview of duty-cycled DPE receiver architecture with DPE measurement update and DPE time update [80].
Source: Reproduced with permission of Institute of Navigation.

An implementation with a duty-cycling factor of 2% has 21.5.1 Direct Timing Estimation


succeeded in providing positioning updates at 1 Hz [80].
Many GNSS-based applications do not need a full PVT
The implementation has the following parameters, as
solution. For example, GNSS provides timing and syn-
shown in Figure 21.10:
chronization to critical infrastructures [81–84] such as
1) Perform 1 DPE measurement update every second on a cellular communication network, financial transactions‚
20 ms GNSS raw signal snippet during which the GNSS and the power grid. Such applications only require the
receiver is “on.” time (i.e. clock and clock-drift) element‚ and the position
2) Perform 49 DPE time updates during the 980 ms in information of the GNSS receivers are often known. By
which the GNSS receiver is “off.” using the prior information available to the timing-
focused receiver, the DPE search domain can be reduced
to the two clock parameters in a Direct Time Estimation
(DTE) problem. DTE naturally reduces the computation
load challenge, while maintaining DPE’s advantage in
21.5 DPE Variants terms of robustness [85].
The architecture of DTE is comprised of two main stages.
So far, we have analyzed DPE in the context of PVT estima- First, a two-dimensional search space is generated with
tion. This section discusses alternative uses of the DPE con- clock bias and clock-drift parameters δt and δt, respectively.
cept, which sometimes require additional tuning but are A combined satellite signal replica is generated for each
conceptually similar to what we have seen in previous sec- grid point by utilizing the known receiver and satellite
tions. We intentionally wanted to keep this material at a three-dimensional position and velocity vectors. Then,
basic level just to give the flavor, avoiding deep theoretical non-coherent vector correlation is performed to obtain
derivations and providing the main concept and references weights that are derived proportional to the likelihood of
instead. the point in the grid, as described in Eq. (21.47). After
544 21 Direct Position Estimation

Clock Bias
Receiver 3D All satellites
position and 3D position
Across the
velocity and velocity
candidates in
search space
Clock Drift

Combined satellite Maximum


signal replica likelihood

Correlation valve
clock state

Vector
Correlation
Incoming raw

t

GPS signal

)
cδt

(m
t

cδt

as
cδt

bi
Cloc


k dr

k
oc
ift (m cδt

t

/s)

Cl
Figure 21.11 Diagram of the first stage of DTE, where the MLE cost function is the computed CAFs across satellites [85].
Source: Reproduced with permission of IEEE.

optimizing the ML cost function, an MLE of the clock para- Yaw

meters is computed tML = δt, δt at that instant.

Figure 21.11 shows the first stage of DTE. Based on the pre-
dicted clock estimate, the error measurement values are
computed and sent to the second stage, where a KF is in
charge of computing the corrected clock bias and clock-
drift parameters. This KF acts similarly as the low-pass fil-
tering solution presented in Section 21.4.3.

21.5.2 Multi-Receiver DPE and DTE Four manifolds are projected to


and overlap at centroid
Multi-receiver-DPE (MR-DPE) extends the DPE concept to
the situation where a network of DPE-enabled receivers is Figure 21.12 Example of aggregation of MLE cost functions.
available. Such multi-receiver architecture can provide per- They are projected to the centroid of the network, with knowledge
of antenna baselines and aircraft yaw, pitch, and roll (the latter two
formance benefits, as additional information redundancy is
are not shown) [86]. Source: Reproduced with permission of
exploited. The receiver elements form a network to generate Institute of Navigation.
a unified, network-wise PVT solution. MR-DPE improves
receiver robustness of the following aspects: (i) on the
receivers at its vertices. MR-DPE operates iteratively to esti-
receiver level, MR-DPE inherits robustness from individual
mate the network’s PVT coordinate [86]. Similarly, the
DPE receivers; (ii) on the network level, since the baselines
multi-receiver concept has been applied to DTE as multi-
among the antennas are fixed and assumed known, one can
receiver DTE [68, 87].
relate receivers’ measurements among them through linear
transformations, providing measurement redundancy; and
21.5.3 Joint GNSS-Vision DPE
(iii) while each of these antennas associated with the GNSS
receiver elements may have a different, incomplete sky view, Here we briefly comment on the fusion of GNSS and cam-
they together create a more complete observation by aggre- era measurements through DPE’s principle. The direct-
gating the information these antennas respectively capture, positioning concept can be applied similarly to both GNSS
providing geometric redundancy. and vision-based positioning. It is possible to formulate the
As an example, Figure 21.12 depicts a potential use in the MLE for both cases. In other words, one searches for the
context of aircraft positioning. The measurement aggrega- underlying position parameters that maximize the correla-
tion process takes place by fusing the different MLE cost tion with the GNSS raw signals and camera images. In
functions of the form in Eq. (21.35) by projecting toward vision-based positioning, the image replicas are selected
the centroid. In the figure, a rigid platform represented from a database of geotagged reference images. An example
by a fixed-wing aircraft is shown with four DPE-enabled of a suitable database is Google Street View [89, 90], where
21.5 DPE Variants 545

Deep Coupling

LOS Projections Initialization

Feature Matching
Vector Correlation x, y, z, cdt, Filter
GPS DP x, y, z, cdt,
x, y, z, cdt
x, y, z, cdt
Time
Coor. Transform
Update

Feature Matching
Measurement
Feature-based Update PVT
x, y, z
Homography
Vision DP

Tracking

Figure 21.13 Block diagram describing joint GNSS-vision DPE [88]. Source: Reproduced with permission of IEEE.

the images are indexed by latitude, longitude, heading‚ and parameters. Navigation candidates may be initialized around
tilt. Image features, such as those obtained from the the current navigation prediction. Constraints are used to elim-
Oriented FAST and Rotated BRIEF feature extractor, can inate candidates and reduce the computation load, for instance,
be used to match consecutive frames. By estimating the by requiring that a road user has to be on the road.
camera's translation through traditional computer vision Initialization of the navigation guesses is followed by fea-
approaches, navigation-domain measurements can be gen- ture matching, as shown in Figure 21.14. GNSS DPE uses
erated. These navigation-domain vision measurements are vector correlation to get the likelihood distribution across
easily fused with the navigation-domain likelihood mani- the navigation guesses, estimating the mean and variance.
fold of DPE, providing deep integration of the sensors Vision DPE is unable to get a similar continuous likelihood
[91]. The direct measurements from both GNSS and vision distribution. The replica sampling in Vision DPE is discrete
are then combined in a navigation filter. and depends on the reference database. In addition, vision
An example block diagram is shown in Figure 21.13. Each feature matching is susceptible to spurious results. A robust
iteration of the joint GNSS-vision DPE begins from a prediction, Vision DPE algorithm is inspired by a combination of prior
provided by the navigation filter, of the current navigation work by other researchers [92, 93]. Two rounds are involved.

Guessed candidates
Perform feature matching

Vector Correlation Feature Matching


Homography

Figure 21.14 GNSS feature matching using the vector correlator (left). Vision feature matching using vision features and homography
analysis (right) [88]. Source: Reproduced with permission of IEEE.
546 21 Direct Position Estimation

In the first round, 2D features are extracted from the reference Sriramya Bhamidipati for reviewing the chapter. Pau Closas
images [94], after which feature matching is performed has been partially supported by the National Science Foun-
between the reference images and the observed image. The dation under Awards CNS-1815349 and ECCS-1845833.
reference images are then ranked according to their overall
feature distance to the observed image. The lower the overall
feature distance, the better the match. A threshold on the References
overall feature distance is used to select potential reference
images. In the second round, homography analysis is used 1 P. Closas, C. Fernández-Prades, and J. Fernández-Rubio,
to verify that the reference image and observed image are gen- “Maximum likelihood estimation of position in GNSS,”
erated from a similar camera view [95–97]. IEEE Signal Processing Letters, vol. 14, no. 5, pp.
359–362, 2007.
2 P. Closas and A. Gusi-Amigó, “Direct Position Estimation
21.6 Conclusions of GNSS Receivers: Analyzing main results, architectures,
enhancements, and challenges,” IEEE Signal Processing
DPE is an emerging concept in the design of advanced Magazine, vol. 34, no. 5, pp. 72–84, 2017.
GNSS receivers. At a glance, DPE estimates a receiver’s 3 L.R. Weill, “A high performance code and carrier tracking
PVT using signals from all satellites jointly at the sample architecture for ground-based mobile GNSS receivers,” in
level, thus avoiding the intermediate step of computing Proceedings of the 23rd International Technical Meeting of
observables. DPE-enabled receivers are known to improve The Satellite Division of the Institute of Navigation (ION
metrics such as accuracy (reducing the positioning error, GNSS 2010), Portland, OR, 2010, pp. 3054–3068.
particularly in scenarios where conventional solutions fail, 4 P. Ward, “The natural measurements of a GPS receiver,” in
such as in urban canyons), reliability (allowing for nearly Proceedings of the 51st Annual Meeting of the Institute of
optimal operation in cases where the received signal is Navigation, 1995, pp. 67–85.
severely degraded), and availability (increasing the receiver 5 P.W. Ward, J.W. Betz, and C.J. Hegarty, ch. Satellite signal
sensitivity). The main drawback of DPE with respect to acquisition, tracking, and data demodulation,
two-step positioning is its increased computational burden. Understanding GPS: Principles and Applications Second
However, researchers have begun to explore approxima- Edition, Artech House, Boston, 2006, pp. 153–240.
tions, assumptions, and computing methods that reduce 6 B.W. Parkinson and J.J. Spilker, Progress in Astronautics
the limitations of this load. and Aeronautics: Global Positioning System, American
This chapter introduced standard GNSS signal processing Institute of Aeronautics and Astronautics, 1996.
and describes how DPE modifies existing architectures. 7 P. Misra and P. Enge, Global Positioning System, Ganga-
Both qualitative and quantitative arguments were given Jamuna Press, 2006.
in support of DPE, even providing theoretical estimation 8 J.J. Spilker Jr, “Vector delay lock loop processing of
bounds. We derived the MLE of position, resulting in a cost radiolocation transmitter signals,” March 14 1995, US
function that shares many similarities with the popular Patent 5,398,034.
CAF. The chapter also provided a deep discussion of prac- 9 J.J. Spilker and B.W. Parkinson, Fundamentals of signal
tical implementation aspects as well as variants and appli- tracking theory, Progress in Astronautics and Aeronautics:
cations in several fields. Future work includes designing Global Positioning System, American Institute of
more efficient methodologies for implementing DPE with Aeronautics and Astronautics, 1996, pp. 245–328.
less computational cost as well as principled methods to 10 M. Lashley and D.M. Bevly, “Vector delay/frequency lock
fuse heterogeneous information from several sources. loop implementation and analysis,” in Proceedings of the 2009
International Technical Meeting of The Institute of Navigation,
Anaheim, CA, January 2009, 2009, pp. 1073–1086.
21.7 Acknowledgments 11 S. Bhattacharyya, “Performance and integrity analysis of
the vector tracking architecture of GNSS receivers,” PhD
Pau Closas would like to highlight that the part of this work dissertation, University of Minnesota, 2012.
related to the estimation bounds is the fruit of joint works 12 P.W. Ward, “Performance comparisons between FLL, PLL
with his colleague Adrià Gusi-Amigó. Grace X. Gao would and a novel FLL-assisted-PLL carrier tracking loop under
like to acknowledge that this chapter is based on a number RF interference conditions,” in Proceedings of the 11th
of students’ theses, papers‚ and presentations in her research International Technical Meeting of the Satellite Division of
group, including by Yuting Ng, Arthur Chu, and Sriramya The Institute of Navigation (ION GPS 1998), Nashville, TN,
Bhamidipati. She is thankful to Matthew Peretic and 1998, pp. 783–795.
References 547

13 P.W. Ward and T.D. Fuchser, “Stability criteria for GNSS A Single-Frequency Approach, Springer Science & Business
receiver tracking loops,” Navigation, Journal of The Media, 2007.
Institute of Navigation, vol. 61, no. 4, pp. 293–309, 29 S. Bancroft, “An algebraic solution of the GPS equations,”
2014. IEEE Transactions on Aerospace and Electronic Systems,
14 P. Axelrad, J. Donna, and M. Mitchell, “Enhancing GNSS vol. 21, no. 7, pp. 56–69, January 1985.
acquisition by combining signals from multiple channels 30 J. Dampf, T. Pany, W. Bär, J. Winkel, C. Stöber, K.
and satellites,” in Proceedings of the 22nd International Fürlinger, P. Closas, and J.A. Garcia-Molina, “More than
Technical Meeting of The Satellite Division of the Institute of we ever dreamed possible: Processor technology for GNSS
Navigation (ION GNSS 2009), 2009, pp. 2617–2628. software receivers in the year 2015,” Inside GNSS, vol. 10,
15 A.M. Reuven and A.J. Weiss, “Direct position no. 4, pp. 62–72, 2015.
determination of cyclostationary signals,” Signal 31 J. Vilà-Valls, P. Closas, M. Navarro, and C. Fernández-
Processing, vol. 89, no. 12, pp. 2448 – 2464, 2009. [Online]. Prades, “Are PLLs dead? A tutorial on Kalman filter-based
Available: http://www.sciencedirect.com/science/article/ techniques for digital carrier synchronization,” IEEE
pii/S0165168409001650 Aerospace and Electronic Systems Magazine, vol. 32, no. 7,
16 J. B.-Y. Tsui, Fundamentals of Global Positioning System pp. 28–45, 2017.
Receivers: A Software Approach, vol. 173, John Wiley & 32 F.S.T. Van Diggelen, A-GPS: Assisted GPS, GNSS, and
Sons, 2005. SBAS, Artech House, 2009.
17 P. Closas, “Bayesian Signal Processing Techniques for 33 A.J. Weiss, “Direct position determination of narrowband
GNSS Receivers: From multipath mitigation to radio frequency transmitters,” IEEE Signal Processing
positioning,” PhD dissertation, Dept. of Signal Theory and Letters, vol. 11, no. 5, pp. 513–516, 2004.
Communications, Universitat Politècnica de Catalunya 34 A.J. Weiss and A. Amar, “Direct position determination of
(UPC), Barcelona, Spain, June 2009. multiple radio signals,” EURASIP Journal on Advances in
18 T.S. Rappaport, Wireless Communications: Principles and Signal Processing, vol. 2005, no. 1, p. 653549, 2005.
Practice, Prentice-Hall, 1996. 35 P. Axelrad, B.K. Bradley, J. Donna, M. Mitchell, and S.
19 F. Fontan, M. Vazquez-Castro, C. Cabado, J. Garcia, and E. Mohiuddin, “Collective detection and direct positioning
Kubista, “Statistical modeling of the LMS channel,” IEEE using multiple GNSS satellites,” Navigation, Journal of The
Transactions on Vehicular Technology, vol. 50, no. 6, pp. Institute of Navigation, vol. 58, no. 4, pp. 305–321, 2011.
1549–1567, November 2001. 36 .A. García-Molina and J.A. Fernández-Rubio, “Collective
20 A.V. Oppenheim, A.S. Willsky, and S.H. Nawab, Signals unambiguous positioning with high-order BOC signals,”
and Systems, vol. 2, Prentice-Hall Englewood Cliffs, NJ, IEEE Transactions on Aerospace and Electronic Systems,
vol. 6, no. 7, p. 10, 1983. vol. 55, no. 3, pp. 1461–1473, 2018.
21 M. Irsigler, J.A. Ávila Rodríguez, and G.W. Hein, “Criteria 37 L. Li, J.W. Cheong, J. Wu, and A.G. Dempster,“
for GNSS multipath performance assessment,” in Improvement to multi-resolution collective detection in
Proceedings of the ION GPS/GNSS 2005, Long Beach, CA, GNSS receivers,” The Journal of Navigation, vol. 67, no. 2,
September 2005. p. 277, 2014.
22 W.C. Jakes, Microwave Mobile Communications, Wiley, 38 L. Narula, K.P. Singh, and M.G. Petovello, “ Accelerated
New York, 1974. collective detection technique for weak GNSS signal
23 J. G. Proakis, Digital Communications, 3rd ed. New York: environment,” In 2014 Ubiquitous Positioning Indoor
McGraw Hill, 1995. Navigation and Location Based Service (UPINLBS), IEEE,
24 A. Steingass and A. Lehner, “Measuring Galileo’s 2014, November, pp. 81–89
multipath channel,” in Proceedings of the European 39 F. Vincent, E. Chaumette, C. Charbonnieras, J. Israel, M.
Navigation Conference, ENC-GNSS 2003, Graz, Austria, Aubault, and F. Barbiero, “Asymptotically efficient
April 2003. GNSS trilateration,” Signal Processing, vol. 133, pp.
25 ——, “Measuring the navigation multipath channel – 270–277, 2017.
A statistical analysis,” in Proceedings of the ION GPS/GNSS 40 T. Pany, Navigation Signal Processing for GNSS Software
2004, Long Beach, CA, September 2004. Receivers, Artech House, 2010.
26 E. Kaplan and C. Hegarty, Understanding GPS: Principles 41 P. Closas, C. Fernandez-Prades, and J.A. Fernandez-Rubio,
and Applications. Artech House, 2005. “A Bayesian approach to multipath mitigation in GNSS
27 E.D. Kaplan (ed.), Understanding GPS: Principles and receivers,” IEEE Journal of Selected Topics in Signal
Applications, 2nd Ed., Artech House, 2006. Processing, vol. 3, no. 4, pp. 695–706, Aug 2009.
28 K. Borre, D.M. Akos, N. Bertelsen, P. Rinder, and S.H. 42 R. Iltis and L. Mailaender, “An adaptive multiuser detector
Jensen, A Software-Defined GPS and Galileo Receiver: with joint amplitude and delay estimation,” IEEE Journal
548 21 Direct Position Estimation

on Selected Areas in Communications, vol. 12, no. 5, pp. Realistic Multipath Channels,” in in Proceedings
774–785, June 1994. of the ION GNSS 2015, Tampa, Florida (USA),
43 A. Amar and A.J. Weiss, “New asymptotic results on two September 2015.
fundamental approaches to Mobile Terminal Location,” in 56 P. Closas, A. Gusi-Amigó, and J. Blanch, “Integrity
Proceedings of the 3rd International Symposium on measures in direct-positioning,” in Proceedings of
Communications, Control and Signal Processing, IEEE, the ION GNSS+ 2017, Portland, Oregon,
March 2008, pp. 1320–1323. September 2017.
44 P. Closas, C. Fernández-Prades, and J.A. Fernández-Rubio, 57 P. Closas, C. Fernández-Prades, and J. Fernández-
“Direct position estimation approach outperforms Rubio, “On the maximum likelihood estimation of
conventional two-steps positioning,” Proceedings of position,” in Proceedings of the 19th International
EUSIPCO ’09, August 2009. Technical Meeting of the Satellite Division of The
45 P. Closas, C. Fernandez-Prades, and J.A. Fernandez- Institute of Navigation (ION GNSS 2006), Fort Worth,
Rubio, “Cramér–Rao bound analysis of positioning TX, 2006, pp. 1800–1810.
approaches in GNSS receivers,” IEEE Transactions on 58 A. Amar and A. J. Weiss, “Localization of narrowband
Signal Processing, vol. 57, no. 10, pp. 3775–3786, radio emitters based on Doppler frequency shifts,” IEEE
October 2009. Transactions on Signal Processing, vol. 56, no. 11, pp. 5500–
46 A. Gusi-Amigó, P. Closas, A. Mallat, and L. Vandendorpe, 5508, November 2008.
“Cramér-Rao bound analysis of UWB based localization 59 Y. Ng, “Improving the robustness of GPS direct position
approaches,” in Ultra-Wideband (ICUWB), 2014 IEEE estimation,” Master’s thesis, University of Illinois at
International Conference on, 2014. Urbana-Champaign, 2016.
47 ——, “Ziv-Zakai bound for direct position estimation,” 60 B. Bradley, P. Axelrad, J. Donna, and S. Mohiuddin,
Navigation, vol. 65, no. 3, pp. 463–475, Dec 2018. “Performance analysis of collective detection of weak GPS
48 D. Chazan, M. Zakai, and J. Ziv, “Improved lower signals,” in Proceedings of the 23rd International
bounds on signal parameter estimation,” IEEE Technical Meeting of The Satellite Division of the Institute of
Transactions on Information Theory, vol. 21, no. 1, Navigation (ION GNSS 2010), Portland, OR, 2010,
pp. 90–93, 1975. pp. 3041–3053.
49 K. Bell, Y. Steinberg, Y. Ephraim, and H. Van Trees, 61 P.C. Gómez, “Bayesian signal processing techniques for
“Extended Ziv-Zakai lower bound for vector parameter GNSS receivers,” PhD dissertation, Universitat Politecnica
estimation,” Information Theory, IEEE Transactions on, de Catalunya, 2009.
vol. 43, no. 2, pp. 624 –637, March 1997. 62 J.W. Cheong, J. Wu, A.G. Dempster, and C. Rizos,
50 R.D.J. V. Nee, J. Siereveld, P.C. Fenton, and B.R. “Efficient implementation of collective detection,” in
Townsend, “Synchronization over rapidly time-varying IGNSS symposium, 2011, pp. 15–17.
multi-path channels for cdma downlink receiver in time- 63 ——, “Assisted-GPS based snap-shot GPS receiver with
division mode,” IEEE Trans. on Vehicular Technology, FFT-accelerated collective detection: Time synchronisation
vol. 56, no. 4, pp. 2216–2225, July 2007. and search space analysis,” in Proceeding of the 2012
51 J. Ziv and M. Zakai, “Some lower bounds on signal International Technical Meeting of the Satellite Division of
parameter estimation,” IEEE Transactions on Information the Institute of Navigation, 2012.
Theory, vol. 15, no. 3, pp. 386–391, 1969. 64 J. Liu, X. Cui, M. Lu, and Z. Feng, “A direct position
52 A. Gusi-Amigó, “Bounds on the accuracy of ultra- tracking loop for GNSS receivers,” in Proceedings of the 24th
wideband based positioning,” PhD dissertation, Université International Technical Meeting of The Satellite Division of
Catholique de Louvain, Louvain-la-Neuve, Belgium, the Institute of Navigation (ION GNSS 2011), Portland, OR,
July 2015. 2011, pp. 3634–3643.
53 S. M. Kay, Fundamentals of Statistical Signal Processing: 65 ——, “Direct position tracking loop based on linearised
Estimation Theory, Prentice-Hall, Englewood Cliffs, New signal model for global navigation satellite system
Jersey, USA, 1993. receivers,” IET Radar, Sonar & Navigation, vol. 7, no. 7,
54 P. Closas, C. Fernández-Prades, and J. Fernández-Rubio, pp. 789–799, 2013.
“Maximum likelihood estimation of position in GNSS,” 66 F. D. Nunes, J. M. S. Margal, and F. M. G. Sousa, “Low-
IEEE Signal Processing Letters, vol. 14, no. 5, pp. complexity VDLL receiver for multi-GNSS constellations,”
359–362, 2007. in Proceedings of 5th ESA Workshop on Satellite Navigation
55 P. Closas, C. Fernández-Prades, A. J. Fernández, M. Wis, Technologies and European Workshop on GNSS Signals and
G. Veccione, F. Zanier, J. Garcia-Molina, and M. Crisci, Signal Processing (NAVITEC 2010), December 2010,
“Evaluation of GNSS Direct Position Estimation in pp. 1–8.
References 549

67 T.H. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein, Journal of The Institute of Navigation, vol. 50, no. 2,
Introduction to Algorithms, MIT press, Cambridge pp. 79–93, 2003.
MA, 1990. 80 Mathworks. (2015, jun) State estimation using time-
68 P. Closas, C. Fernández-Prades, and J. Fernández- varying kalman filter. [Online]. Available: http://
Rubio, “ML estimation of position in a GNSS receiver www.mathworks.com/help/control/getstart/estimating-
using the SAGE algorithm,” in Proceedings of IEEE states-of-time-varying-systems-using-kalman-
International Conference on Acoustics, Speech and filters.html
Signal Processing–ICASSP ’07, vol. 3, April 2007, pp. 81 Y. Ng and G.X. Gao, “Computationally efficient direct
III–1045–III–1048. position estimation via low duty-cycling,” in Proceedings of
69 JS. Bhamidipati and G.X. Gao, “ Multi-receiver GPS-based the Institute of Navigation GNSS+ conference (ION GNSS+
direct time estimation for PMUs,” in ION GNSS+ 2016, 2016), Portland OR, September 2016.
Portland, OR, 2016. 82 J. Merrill, “Patriot watch vigilance safeguarding America,”
70 P. Closas and C. Fernández-Prades, “Bayesian nonlinear in Presentation Telcordia-NIST-ATIS Workshop
filters for direct position estimation,” in Proceedings of Synchronization Telecommun. Syst. (WSTS’12), 2012.
IEEE Aerospace Conference, Big Sky, MT,IEEE, 2010, 83 M.G. Amin, P. Closas, A. Broumandan, and J.L. Volakis,
pp. 1–12. “Vulnerabilities, threats, and authentication in satellite-
71 Y. Ng and G.X. Gao, “Mitigating jamming and meaconing based navigation systems [scanning the issue],”
attacks using direct GPS positioning,” in Proceedings of Proceedings of the IEEE, vol. 104, no. 6, pp. 1169–1173,
IEEE/ION PLANS 2016, Savannah, GA, IEEE/ION, 2016, June 2016.
pp. 1021–1026. 84 R.T. Ioannides, T. Pany, and G. Gibbons, “Known
72 J. Dampf, K. Frankl, and T. Pany, “Optimal particle filter vulnerabilities of global navigation satellite systems, status,
weight for Bayesian direct position estimation in a GNSS and potential mitigation techniques,” Proceedings of the
receiver,” Sensors, vol. 18, no. 8, p. 2736, 2018. IEEE, vol. 104, no. 6, pp. 1174–1194, 2016.
73 K. Strandjord, P. Axelrad, D. Akos, S. Mohiuddin, 85 E.G. Agency, “GNSS Market Report,” GSA, Prague, Czech
“Improved urban navigation with direct positioning and Republic, Tech. Rep., May 2017.
specular matching,” Proceedings of the 2020 International 86 Y. Ng and G.X. Gao, “Robust GPS-based direct timing
Technical Meeting of The Institute of Navigation, San Diego, estimation for PMUs,” in Proceedings of IEEE/ION
California,.2020/1, pp. 787–800 PLANS 2016, Savannah, GA, IEEE/ION, 2016, pp.
74 T. Lin, J.T. Curran, C. O’Driscoll, and G. Lachapelle, 472–476.
“Implementation of a navigation domain GNSS signal 87 A. Chu and G.X. Gao, “ GPS multi-receiver direct position
tracking loop,” in Proceedings of the 24th International estimation for aerial applications,” IEEE Transactions on
Technical Meeting of The Satellite Division of the Institute of Aerospace and Electronic Systems. vol. 56, no. 1, pp.
Navigation (ION GNSS 2011), Portland, OR, 2011, pp. 249–262, February 2020.
3644–3651. 88 S. Bhamidipati and G.X. Gao, “GPS spoofer localization for
75 R.E. Kalman, “A new approach to linear filtering and PMUs using multi-receiver direct time estimation,” in ION
prediction problems,” Journal of Basic Engineering, vol. 82, GNSS+ 2017, Portland, OR, 2017.
no. 1, pp. 35–45, 1960. 89 Y. Ng and G.X. Gao, “ Joint GPS and vision direct position
76 R. E. Kalman et al., “Contributions to the theory of optimal estimation,” in Proceedings of IEEE/ION PLANS 2016,
control,” Boletín de la Sociedad Matemática Mexicana, Savannah, GA, IEEE/ION, 2016, pp. 380–385.
vol. 5, no. 2, pp. 102–119, 1960. 90 Google Inc. (2016, January) Google Maps. [Online],
77 P. Closas, C. Fernández-Prades, and J. Fernández-Rubio, Available: https://www.google.com/maps/
“Bayesian direct position estimation,” in Proceedings of the 91 Google Developers. (2016, January) Google Street View
21st International Technical Meeting of the Satellite Division Image API. [Online],Available: https://developers.google.
of The Institute of Navigation (ION GNSS 2008), Savannah, com/maps/documentation/streetview/
GA, 2008, pp. 183–190. 92 M.A. Fischler and R.C. Bolles, “Random sample consensus:
78 D. Simon and T.L. Chia, “Kalman filtering with state a paradigm for model fitting with applications to
equality constraints,” IEEE Transactions on Aerospace image analysis and automated cartography,”
and Electronic Systems, vol. 38, no. 1, pp. 128–136, Communications of the ACM, vol. 24, no. 6,
Jan 2002. pp. 381–395, 1981. [Online]. Available: http://doi.acm.org/
79 F.D. Busse, J.P. How, and J. Simpson, “Demonstration of 10.1145/358669.358692
adaptive extended Kalman filter for low-earth-orbit 93 M. Salarian, A. Manavella, and R. Ansari, “Accurate
formation estimation using CDGPS,” Navigation, localization in dense urban area using Google Street View
550 21 Direct Position Estimation

images,” in SAI Intelligent Systems Conference (IntelliSys), 96 R. Hartley and A. Zisserman, Multiple View Geometry in
2015, IEEE, 2015, pp. 485–490. Computer Vision, vol. 23, no. 2, Cambridge University
94 A.R. Zamir and M. Shah, “Accurate image localization Press, 2005.
based on Google Maps Street View,” in Proceedings 97 T. Sattler, B. Leibe, and L. Kobbelt, “Fast image-based
of 11th European Conference on Computer Vision localization using direct 2D-to-3D matching,” in 2011 IEEE
(ECCV 2010). Springer Berlin Heidelberg, 2010, pp. International Conference on Computer Vision (ICCV),
255–268. November 2011, pp. 667–674.
95 E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, “ ORB: 98 Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli,
An efficient alternative to SIFT or SURF,” in 2011 IEEE “Image quality assessment: From error visibility to
International Conference on Computer Vision (ICCV), structural similarity,” IEEE Transactions on Image
November 2011, pp. 2564–2571. Processing, vol. 13, no. 4, pp. 600–612, 2004.

You might also like