2.2 Bayesian Estimation Theory
2.2.2 Kalman Filter
So far we have reviewed different approaches for dealing with the estimation of fixed
parameters. In tracking, however, parameters are assumed to evolve in time. To clearly
distinguish between fixed parameters and those that evolve in time, the latter ones are
often referred to as a state. The task in tracking is then to infer the state s[n] from
the measurements y[n] taken at time-step n. Obviously, we could implement tracking
by estimating states at every time-step individually, using some of the earlier discussed
methods. However, if an estimate ˆs[n ≠ 1] is available at time-step n, this estimate can
often be used as prior information for the estimation of ˆs[n]. When tracking the location
of a pedestrian, as an example, the change in location is clearly determined by the
pedestrian’s velocity. Thus, if we have obtained an estimate for the pedestrian’s location
and velocity at time-step n ≠ 1, we can already estimate the pedestrian’s location at
time-step n without taking any measurements. In order to exploit such knowledge, we
obviously need a good model for the evolution of the state. In Kalman filtering this
model is often referred to as the state transition and is given by
s[n] = F s[n ≠ 1] + w[n]
(2.21)
The state transition (2.21) consists of the the state transition matrix F and the zero-
mean driving noise w[n] with covariance matrix E[w[n]w
T[m]] = ”
m≠n
R[n] where ”
i≠jis the Kronecker delta function. Note that (2.21) is a linear state transition as required
in conventional Kalman filtering. Extensions for nonlinear models exist [91, p. 407].
However, in this thesis all state transitions will be of the above form.
In order to infer the state from the measurements, we have to find a model for the
measurement equation
y[n] = h(s[n]) + u[n]
(2.22)
where u[n] is the zero-mean measurement noise with a covariance equal to E[u[n]u
T[m]] =
”
m≠nQ[n]. Note that we do not state the measurement equation in its most general
form (see [91, p. 407]). However, in contrast to the linear state transition we now
enable nonlinear relationships between the state and the measurements via the vector
function h. We allow this nonlinear relationship since all the measurements considered
in this thesis depend nonlinearly on the location that we track.
In cases where both the state transition as well as the measurement equations are
linear, i.e., y[n] = H[n]s[n] + u[n] it is possible to obtain state estimates using the
iterative Kalman filter. The Kalman filter is optimal in the MMSE sense if w[n] and
u[n] are jointly Gaussian. And if the Gaussian assumption does not hold, the Kalman
filter is still the optimal LMMSE estimator. However, as noted above the measurement
equations are often nonlinear. Thus, in practice we often use the so-called extended
Kalman filter (EKF) instead of the conventional Kalman filter. The main idea of the
EKF is to linearize the nonlinear parts of the models using the current state estimate.
As such, we generally cannot make any claims regarding optimality.
For each time-step n, the EKF consists of an a priori and an a posteriori estimation
stage. In the a priori stage, we estimate the state ˆs
≠[n] and covariance of the state
estimate P
≠[n] at time-step n using all measurements up to but excluding y[n]. The
estimation is given by
ˆs
≠[n] = Fˆs
+[n ≠ 1]
(2.23)
P
≠[n] = FP
+[n ≠ 1]F
T+ R[n].
(2.24)
In the a posteriori estimation stage, we estimate the state ˆs
+[n] and covariance of the
state estimate at time-step n using all measurements up to and including y[n]. The
2.2 Bayesian Estimation Theory
calculation in the a posteriori estimation stage is according to
K[n] = P
≠[n]H[n]
T(H[n]P
≠[n]H
T[n] + Q[n])
≠1(2.25)
ˆs
+[n] = ˆs
≠[n] + K[n]#y[n] ≠ h(ˆs
≠[n])$
(2.26)
P
+[n] = (I ≠ K[n]H[n])P
≠[n]
(2.27)
with the Jacobian matrix H[n] =
ˆh[n]ˆs[n]
evaluated at ˆs
≠[n]. Interestingly, the EKF
provides the possibility of state prediction. By executing the a priori estimation stage
N
times without the a posteriori stage, we are able to obtain a prediction ˆs
p[n + N]
using only measurements up to time-step n.
Note that the EKF is not the only approach for nonlinear filtering. An overview
of other popular approaches such as unscented Kalman filters or particle filters can be
found, for example, in [91]. Note, moreover, that the CRB discussed in Section 2.1.2 does
not lower-bound the performance of Bayesian estimators such as the EKF. Sometimes,
we can learn certain general properties of the estimation problem by deriving the classical
CRB and benefit from the knowledge of these properties also in Bayesian estimation
(see Section 5.1 for an example). However, if we are interested in a lower performance
bound for, e.g., an EKF we have to refer to the so-called posterior CRB [101].
CHAPTER 3
Localization and Tracking in
Wireless Networks
T
he aim of localization or positioning is to estimate the location ¸ = [x, y]
Tof a
TN. In this thesis, we generally assume that this task is achieved by a set of K
collaborating ONs in a wireless network. In accordance with the discussion in Section 1.1,
we moreover assume that each of the ONs is equipped with some sort of a directional
antenna as depicted in Figure 3.1. In order to localize the TN, each ON k, k = 1 . . . K
makes a measurement
kbased on the TN signal and communicates this measurement
to the fusion center, where the measurements from all K ONs are combined into a TN
location estimate. In contrast to localization, we talk about location tracking when we
estimate the evolution of the TN location over time. In the following discussion, we
will mostly discuss localization. However, the majority of the topics discussed in this
chapter also apply to location tracking.
This chapter is organized as follows. In Section 3.1 we discuss important properties
of wireless networks, the ON devices, and the TN as well as their implications for
localization and location tracking. Thereafter, in Section 3.2 we discuss measurements
commonly used for localization in detail. Finally, in Section 3.3 we summarize the
Stansfield fusion algorithm.
3.1 Properties of Wireless Networks and Implications
for Localization and Location Tracking
In a practical wireless network, the choice of measurements for localization and location
tracking as well as the fusion mechanisms are confined by the network properties, the
ON devices and the nature of the TN. For time-based measurements the synchronization
Fig. 3.1. Considered localization system where the observing nodes (ONs) collaborate to estimate or
track the location of the target node (TN). ONs are assumed to be equipped with directional antennas.
of the network is of particular importance. In the best-case scenario, the clocks of the
ONs are perfectly synchronized with the TN clock. In that case the distance d
kcan be
extracted directly from the ToA estimates (see Section 3.2.3). However, already very
small synchronization errors significantly increase the ToA estimation error, which is
related to the relative clock offsets via the speed of light, c ¥ 3 ◊ 10
8m/s. In many
networks, a sufficiently accurate synchronization with the TN is not assumed to be
given. Thus, location estimates are often obtained by fusing time difference of arrival
(TDoA) estimates instead [5, 46]. However, utilizing TDoA estimates for localization
still requires synchronization of the ONs. Without any synchronization in the network,
the propagation time can be exploited using the so-called round-trip time of arrival
(RToA) [29,117]. In order to measure the RToA, an ON sends a packet to the TN. Once
the TN receives the packet, it responds by sending the packet back to the ON [117]. The
time between transmission and reception at the ON is then approximately equal to 2 d
k/c.
However, RToA estimation obviously requires dedicated two-way localization signaling,
which may not be feasible in networks that are primarily designed for communication
purposes, such as the 5G network considered in [57] that also forms the basis for the
localization system proposed in [P7].
Apart from synchronization, the nature of the TN has important implications for
the applicability of time-based measurements. In PU localization, spectrum policy
enforcement, or military applications, the TN is generally non-cooperative [55,79,105–
107,109]. This means that both ToA and RToA estimates are not available for localization.
If, in addition, the network is unable to synchronize the ONs, TDoA estimates are also
unavailable. In these cases, localization is generally based on RSS [64, 109] or DoA
estimates [79, 105]. A problem arises when multiple non-cooperative TNs have to be
3.1 Properties of Wireless Networks and Implications for Localization and Location Tracking