IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO.

7, 1 APRIL 2024 11271

Multistate Constraint Multipath-Assisted Positioning

and Mismatch Alleviation
Xueting Xu , Ao Peng , Member, IEEE, Xuemin Hong , Member, IEEE,
Yixiong Zhang , Member, IEEE, and Xiao-Ping Zhang , Fellow, IEEE

Abstract—Multipath propagation greatly affects the accuracy systems, such as Wi-Fi [2], Bluetooth [3], and ultra-
of Time-of-Arrival (ToA)-based indoor positioning when line-of- wideband [4], have been embraced. While these systems
sight (LOS) signals are only used. In this article, we present offer absolute positioning capabilities, they often grapple with
a novel real-time and low computation complexity multipath-
assisted ToA positioning method, namely, multistate constraint constraints like limited coverage, sparse infrastructure, and
(MSC)-MAP. The delays of reflected signals are taken as potential interference in the Industrial Scientific Medical band.
additional spatial observations to compensate for an insufficient Compared with short-range wireless communication systems,
number of physical transmitters to locate a moving user equip- commercial cellular networks, especially the fifth-generation
ment (UE). Virtual anchors are used to model the propagation (5G) new radio, stand out due to their widespread infras-
path of reflected signals, whose locations are obtained via an
MSC estimator, along with the trajectory of UE. In addition, tructure coverage and effective interference management, thus
we demonstrate the mismatch problem in data association and offering a promising solution for universal indoor positioning
its impact on positioning performance. To achieve real-time services [5].
processing, we propose two robust multipath-assisted position- Some angular-based indoor positioning methods are
ing methods with mismatch alleviation by randomly selecting proposed in [6] and [7], but the accuracy is greatly affected by
subset and constraint relaxation, respectively, to meet various
computational complexity requirements. Simulation results show the size of the antenna array. So, distance-based positioning
that, for the MSC-MAP method, the mean-square error of remains the most promising means to achieve high-accuracy
the position is generally less than 0.2 m in challenging indoor indoor positioning. However, the positioning accuracy is
environments. Among mismatch alleviation algorithms, posi- sensitive to the presence of multipaths [8]. Conventional
tioning error is reduced by 69% even when the percentage distance-based localization involves measuring the time delay
of mismatched measurement data is as high as 42%. The
proposed algorithms can also efficiently handle signals with of the line-of-sight (LOS) path [9], which can easily be
non-Gaussian impairments, a common characteristic in real- obscured by indoor obstacles. Moreover, non-LOS (NLOS)
world data. Moreover, these algorithms can substantially improve paths can compromise the delay measurement of an LOS
positioning performance while adding minimal computation time path due to limited bandwidth. The multipath phenomenon is
in the presence of measurement mismatches, outperforming state- thus frequently deemed a hindrance to high-accuracy indoor
of-the-art methods utilizing different data association techniques.
localization [10]. To eliminate errors induced by NLOS sig-
Index Terms—Data association, minimum observation nals in indoor positioning, a straightforward method entails
constraint (MOC), mismatch alleviation, multipath-assisted identifying the NLOS components in measurements to prevent
positioning, multistate constraint (MSC).
them from being used in positioning [11], [12], [13], [14].
However, the identification probability is often limited in
dense multipath scenarios. One misidentification may lead to
I. I NTRODUCTION severe positioning errors. Another classic method mitigates the
ARIOUS location-based applications supported by effects of NLOS measurements on positioning results [15],
V traditional global navigation satellite systems are mostly
restricted to outdoor scenarios [1]. In the context of
[16], [17], [18]. An estimator for the LOS Time of Arrival
(ToA) in multipath conditions is presented in [19] by obtaining
indoor positioning, many short-range wireless communication an approximate distribution for the received signals of all
the antennas. However, the accuracy deteriorates due to the
Manuscript received 26 September 2023; revised 24 October 2023; accepted multipath effect. Moreover, it also requires sufficient anchors
25 October 2023. Date of publication 1 November 2023; date of current to achieve trilateration positioning.
version 26 March 2024. This work was supported by the National Key Multipath-assisted positioning has emerged as a new
Research and Development Program of China under Grant 2018YFB0505200.
(Corresponding author: Ao Peng.) positioning paradigm and has garnered increasing research
Xueting Xu, Ao Peng, Xuemin Hong, and Yixiong Zhang are interest [20]. This paradigm treats multipaths as beneficial to
with the School of Informatics, Xiamen University, Xiamen 361005, positioning given their rich information about the environ-
Fujian, China (e-mail: [email protected]; [email protected];
[email protected]; [email protected]). ment [21]. In some typical indoor scenes, specular reflections
Xiao-Ping Zhang is with Tsinghua–Berkeley Shenzhen Institute, represent the main multipath source, where reflecting sur-
Shenzhen 518055, China, and also with the Department of Electrical, faces tend to be large smooth planes such as walls, floors,
Computer and Biomedical Engineering, Toronto Metropolitan University,
Toronto, ON M5B 2K3, Canada (e-mail: [email protected]). ceilings, and desktops [22]. The multipath signal caused
Digital Object Identifier 10.1109/JIOT.2023.3329147 by specular reflection can be modeled as a direct signal

c 2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
11272 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

between the receiver and a virtual anchor (VA), i.e., the Probability-based data association is proposed in [31], [35],
mirror of a transmitter in the environment. Then, the specular [38], [41], [42], and [43]. The authors seek to design a joint
reflection signal can be used as a new observation source estimation scheme, including data association, UE’s trajectory,
for positioning. Importantly, the geometric information of and VAs’ positions. In this case, the joint posterior PDF
VAs should be exploited before using NLOS observations is solved via recursive Bayesian filtering methods, such as
in positioning. The quality of VAs’ geometrical distribution belief propagation [31], [35], [41] or probability hypothesis
knowledge directly determines the contributions of NLOS density filtering [38]. Next, the posterior PDF of every required
observations in multipath-assisted positioning [23]. variable can be obtained by marginalizing the joint PDF.
A multipath-assisted positioning algorithm based on the In [35], [36], [37], and [44], the VAs’ states are updated by
least-square estimator has been proposed in [24], in which calculating expectations using the evaluated probabilities of all
the estimation of user equipment’s (UE’s) and VAs’ positions possible association hypotheses between VAs’ states and mea-
is formulated as an optimization problem. However, it can surements. Similarly, Kim et al. [38] used the measurements
only deal with stationary nodes. For moving target, methods to update the particle weights representing the UE’s state
based on alternating optimization [25], [26] and the Kalman by means of probabilities. Probabilistic methods come with
filter [27] or variants thereof [28], [29], [30], [31] have been several inherent limitations. First, they demand a high level
proposed. Given nonlinearity in the measurement model, some of spatial correlation, which can be challenging to maintain,
optimal Bayesian filter-based multipath-assisted positioning especially at low sampling rates. This is due to the frequent
algorithms [21], [32], [33], [34], [35], [36], [37], [38], [39], birth and death of multipath components, resulting in less
[40] that propagate the entire posterior distribution of state correlation between states at adjacent moments, and in some
vectors have been proposed to achieve better accuracy. The cases, even independence. Moreover, when dealing with low-
heavy computational load of the classical Bayesian filter quality measurements, probabilistic methods can suffer from
induced by Monte Carlo techniques for the sampling dis- inaccurate association probabilities. This discrepancy can lead
tribution function has further spawned many technologies, to a situation where the probability of an incorrect association
including particle filters [32], [33], message passing [21], surpasses that of a correct one, resulting in what is known
[34], [35], [36], [37], finite set statistics [38], [39], and as a probabilistic mismatch and subsequently deteriorating
variational Bayesian inference [40]. The methods presented performance after expectation. Additionally, in probabilistic
in [35], [36], and [37] realize SLAM based on the marginal data association methods, all available data from various VAs
probability density functions (PDFs) of the joint posterior PDF. is utilized. However, each VA may possess different estimation
The required marginal PDFs are approximated through the qualities, and considering all possible association hypotheses
message-passing algorithm on the factor graph. In [38], posi- of VAs when calculating expectations can lead to performance
tions are characterized by their first-order statistical moment, degradation, especially when the matching relationship is
and the conditional PHD is propagated rather than its PDF. unclear.
These Bayesian SLAM algorithms process the measurements Data association can also be solved indirectly with specially
discretely time step by time step and assume a Markovian designed baseband processing. The parameters of multipath
state dynamic of the moving agent. If the computational signals received at the UE are tracked in [45], [46], and [47]
resource is infinite, theoretical real-time positioning can be via baseband estimators. Then, based on the tracking results,
achieved. However, in practice, positioning using algorithms the posterior distribution of the UE’s and VAs’ positions
with high computation complexity will encounter estimation is estimated recursively using some positioning algorithms
lags owing to the limited computational resources of the (e.g., EKF [45], [46] and recursive Bayesian filtering [47]).
implementation platform. To our knowledge, no scholars have The random sample consensus is used in [45] and [46] to
discerned the minimum observation constraint (MOC) for initialize the set of measurement inliers, i.e., measurements
VA estimation—the minimum requisite observations needed agree with the initial estimate of the VA’s position. Then, the
to obtain a converged estimation of VA’s location. This EKF is used to track multipath parameters with inliers, and the
information will offer valuable guidance for designing real- UE’s position is estimated using tracked parameters. Accurate
time positioning methods. Real time here encompasses two positioning hinges on two key prerequisites: 1) maintaining
distinct aspects. One is theoretical real time, i.e., updating tracking continuity and 2) ensuring the correctness of the
past state without reliance on future information, and the initial inlier subset. However, maintaining tracking continuity
other is low computational complexity to facilitate real-time is a challenging task due to the inherently dynamic nature
processing in a limited hardware resource for implementation. of multipaths, which undergo frequent changes due to their
Because different paths cannot be distinguished through birth and death processes. Furthermore, in some scenarios, the
data or code modulated on the signal, a prime concern in real inliers within the tracking process can undergo significant
multipath-assisted positioning is the data association problem variations. For instance, when there is a long sampling interval,
(i.e., identifying the VA from which signal measurement substantial distances can separate two consecutive trajectory
originates). A direct way is to formulate the data association points, and environmental changes may occur. Consequently,
as a binary assignment problem to build the one-to-one measurements used in subsequent epochs may no longer align
correspondence between VAs and measurements [30]. The with the initial inliers associated with VAs. Additionally,
results of the simulation to quantify the data association mismatches can occur in subsequent epochs when two paths
show that data association mismatch can degrade positioning intersect or when tracking is interrupted, leading to substantial
performance. But the authors do not address it in [30]. positioning errors in the final positioning stage.
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11273

In essence, mismatches are an inevitable aspect of the MSC estimator can maintain the constraint between sampling
generalized data association process when employing exist- points, thus keeping a stable positioning performance when the
ing multipath-assisted positioning methods. Consequently, the step size increases. Second, the hard association for choosing
positioning accuracy of prior methods remains highly sensi- VAs or assigning weights to VAs used in the methods can
tive to the correct match rate in data association. However, help to select beneficial VAs. Furthermore, the mismatch at
prior research has yet to delve into robust multipath-assisted each position update can be alleviated independently in our
positioning in the presence of data association mismatches. proposed methods without requiring the multipath to survive
In this article, we present a novel real-time multipath- long enough to maintain the tracking relationship. Simulation
assisted ToA positioning method based on linear expansion results show that, for the proposed MSC-MAP, the mean-
of the observation model, named multistate constraint (MSC)- square error (MSE) of the position is generally less than
MAP, to estimate the locations of a moving UE in complex 0.2 m in challenging indoor environments. Among mismatch
indoor environments in the 5G system. The delays of reflected alleviation algorithms, positioning error is reduced by 69%
signals are taken as additional spatial observations to compen- even when the percentage of mismatched measurement data
sate for an insufficient number of physical transmitters. VAs is as high as 42%, significantly improving the performance of
are used to model the propagation path of reflected signals, multipath-assisted positioning.
whose locations are obtained via an MSC estimator. In our Our main contributions are summarized as follows.
method, since we do not map the environment, the VAs’ 1) A novel real-time multipath-assisted ToA positioning
estimates are mainly used to assist UE’s positioning. Hence, method based on the MSC estimator is proposed
our proposed algorithm does not depend on the condition of to address estimation divergence due to insufficient
uniqueness mentioned in [25]. Differently, we use estimation observation.
quality control of the VA’s position to find some VAs in the 2) We use estimation quality control to find some VAs in
environment that can be used to assist positioning. Although it the environment that can be used to assist positioning
is challenging to use estimation quality control in our method without using strict specular reflection assumption.
to determine a VA’s position estimate when several solutions 3) We formulate the mismatch alleviation problem in data
exist, this does not affect our final localization results. We association, and propose two mismatch alleviation meth-
can exclude nonconverging VAs from the UE positioning, ods by RS-Subset and CR, respectively, to suit different
which just reduces the number of measurements. Similarly, computational resources.
estimate quality control also allows our method to avoid 4) The performances of the proposed multipath-assisted
the need for strict specular reflection assumption. We decide approaches are verified by Monte Carlo simulations.
whether a VA is available from the aspect of estimation quality. We compare the proposed algorithms against different
Even if a curved surface exists, we can also exclude the VA approaches in the literature in terms of positioning error
corresponding to it. Then, we formulate different measurement and average runtime.
models over prior estimations for cases where a VA’s position The remainder of this article is organized as follows.
is known and unknown. The low-complexity MSC estimator Section II formulates the multipath-assisted positioning
is applied to address estimation divergence due to insufficient problem and mismatch in data association. Section III
observation. The processing delay is only determined by the describes the linearized real-time multipath-assisted position-
bunch size of the multistate. Hence, the real-time positioning ing approach and analyzes the MOC. Section IV formulates
can be realized on both the theoretical and implementation the data association mismatch problem and outlines two
levels at each position update. Based on this, we theoretically mismatch alleviation algorithms. Numerical results and a
analyze the MOC for the proposed estimator, which serves as corresponding discussion are reported in Section V. Finally,
the basis for real-time multipath-assisted positioning. Section VI concludes this article. Main notations are summa-
Mismatch in data association is one of the key factors that rized in Table I.
cause the deterioration of multipath-assisted positioning accu-
racy. The solution of optimal mismatch alleviation can only II. P ROBLEM F ORMULATION
be obtained by brute-force search. To meet the requirement A. Multipath Propagation Model
of real-time processing, we introduce the mismatch matrix
To use multipath components for positioning, the rela-
and formulate the mismatch in data association. To solve the
tionship between the location of UE and the measurable
matrix, we propose two robust multipath-assisted positioning
parameters of the multipath signals should be established
methods with mismatch alleviation by randomly selecting sub-
in advance. We consider a mobile UE with unknown time-
set (RS-Subset) and constraint relaxation (CR), respectively,
varying position PUE,t at epoch t and R base stations (BSs)
to suit different computational resources. We select a subset
with known positions PBS,r , r = 1, . . . , R. The received signal
of VAs with good estimation quality and small association
at epoch t is modeled as
error in the random subset selection-based algorithm, which
provides a swift and suboptimal solution under light mismatch 
N(t)  
(i)
scenarios. In the CR-based algorithm, we set weights to VAs, srx (t) = Ai (t)stx t − τt + n(t) (1)
which offers an approximately optimal solution under severe i=1
mismatch scenarios. The proposed methods have the following where N(t) denotes the number of measurements (i.e., the
advantages over the probabilistic method. First, the proposed number of multipaths that can be observed at time t), Ai (t)
11274 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

TABLE I
N OTATION L IST

Fig. 1. Example of an environment map with VAs and reflected multipath.

The thick black line indicates walls, which we consider to be reflecting
surfaces. The BSs at fixed positions are indicated by, respectively, a red BS
graphic and a blue one within the floor plan. The UE moves along the thin
black line. The red circle and blue one indicate examples of the first-order
VAs associated with BS 1 and BS 2, respectively. The red and blue lines show
some examples of the first-order reflection paths between the moving UE and
two VAs. One orange ellipse indicates one position update using multiple
measurement epochs.

(r)
For LOS component, τt (for r ∈ {1, . . . , R}) provide
measurements to the distances between the UE and BSs

(r) 1 
PUE,t − PBS,r 
τt = (2)
c
where c is the speed of light. For NLOS components, the
propagation delays cannot be used directly because the prop-
agation path is not straight. To achieve multipath-assisted
positioning, we use the model of mirror point (i.e., VA) of a
transmitter against a reflecting surface [22], as illustrated in
Fig. 1. According to geometrical optics theory, the propagation
distance from the transmitter to the UE after reflection is equal
to the distance from the VA to the UE. With the help of VA,
the measurement model of the propagation delay for the ith
multipath component can be expressed as follows:

1

(i) 
τt(i) = PUE,t − PVA  (3)
c
(i)
where PVA , i = R + 1, . . . , N(t) denotes the position of the
ith VA corresponding to the ith multipath signal. The BSs can
also be regarded as special VAs for the LOS paths, which are
(1) (R)
recorded as {PVA , . . . , PVA }.
For irregular reflection surfaces, which are common in
typical indoor scenarios, we use small segments for approxi-
(i)
mation. In other words, we can increase the number of VAs
is the complex amplitude, and τt is the delay of the ith used in the model to achieve a piecewise approximation of the
path for i = 1, . . . , N(t). For simplicity, we always use i = reflection of irregular surfaces.
1, . . . , R to denote the LOS component. If LOS is blocked,
then Ai (t) = 0. stx (t) is the transmitted signal, and n(t)
indicates white Gaussian distributed receiver noise with mean B. Multipath-Assisted Positioning
0 and variance σn 2 . Note that a new reference signal called The UE receives several multipath signals at one moment.
positioning reference signal (i.e., PRS) is introduced by 3GPP To formularize the observation (3), the correspondence
in TS 38.211 (Release 16) [48] for UEs to perform downlink between VAs and multipath signals must be determined,
reference signal time difference (DL-RSTD) measurements at (i.e., data association). We use the classical Hungarian algo-
each BS, the delays τt(i) , i = 1, . . . , N(t) can be estimated with rithm [49] to achieve data association, then the observations
the help of PRS by some channel estimation methods. are arranged by order of associated VAs.
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11275

We record the distance vector between the UE and VAs mismatch alleviation is to estimate  by minimizing the error
(i)
as dt = [ dt(1) · · · dt(N(t)) ]T , where dt denotes the distance between P̂ and P, that is
between the UE and the ith VA, and T  2
  T   
 T   = arg min P̂() − P . (9)
(i)
dt =  1 · P UE,t − (1) (N(t))
PVA · · · PVA .
 (4)  t=1
i·
Mismatch alleviation algorithms are discussed in depth in
(·)i· represents the ith row of the matrix to consider the Section IV.
dimension of the position. Therefore, the residual vector of
UE and VAs position at epoch t is
III. R EAL -T IME M ULTIPATH -A SSISTED
 T
wt = dt − c τt(1) · · · τt(N(t)) . (5) P OSITIONING M ETHOD
In this part, we propose a real-time MSC multipath-assisted
Let P be the position matrix of UE and VAs positioning method, namely, MSC-MAP, to jointly estimate the
 T UE’s and VAs’ positions. We first give measurement models
(1) (M) . (6)
P = PUE,1 · · · PUE,T PVA · · · PVA based on Taylor expansion according to prior knowledge of a
The essential problem of multipath-assisted positioning is VA’s position. Then, we use the MSC to ensure convergence
to estimate the trajectory of UE PUE,t , t ∈ {1, . . . , T} and of the position estimates, followed by the description of the
(i) minimal requirement of observations for joint position estima-
positions of VAs PVA , i ∈ {1, . . . , M} simultaneously with
tion, which we call the MOC. Finally, we give the estimated
known τ t , t = 1, . . . , T, where τ t = [τt(1) · · · τt(N(t)) ]. Based positions of the UE and VAs via the joint measurement model
on the idea of the least-square method, we realize multipath- with MOC.
assisted positioning by minimizing the residual sum of squares
during the UE movement, that is
A. Linear Expansion of the Observation Model
T
P̂ = arg min wt 2 . (7) We adopt ToA-only positioning, in which positions are
P̂ t=1 calculated using distances between the UE and VAs estimated
We start by assuming that the DA algorithm used in this by the delay measurements of received signals. The distance
article can obtain ideal match results. In the next section, we measurement model at epoch t can be written as
 
propose a novel multipath-assisted positioning method based (i) (i)  (i) 
d̂t = cτ̂t = PUE,t − PVA  + n(t)
on the MOC with real-time capability and low computational  
complexity.  (i)
= f PUE,t , PVA + n(t) (10)

C. Mismatch in Data Association (i) (i) (i)

where d̂t denotes the estimate of dt , and τ̂t denotes the
Mismatches in data association are unavoidable due to estimated propagation path delay from the ith VA.
measurement noise and the uncertainty of VAs’ geometric The visibility of VAs varies due to the complicated obstacles
information. According to (4) and (5), a mismatch will lead in an indoor environment. Visible VAs’ number in each epoch
to disorder between ToA measurements and VAs, resulting in is determined by the received multipath signals number. We
a disordered propagation model and even severe positioning use a set to record all the observed VAs. A new VA is added
error. Therefore, it is thus crucial to ensure the positioning to the set whenever a new multipath signal is received. We
(i)
performance of PUE,t and PVA , i ∈ {1, . . . , N(t)} even in the use the multipath delay difference between adjacent epochs to
presence of mismatches. determine whether a new multipath signal has been observed.
We define a data association alleviation sequence We classify VAs into known and unknown VAs based on the
{φi = j; 1 ≤ i ≤ N(t), 1 ≤ j ≤ N(t)}, meaning that the ith ToA convergence of the position estimates of VAs. We draw the
measurements are reassociated with the jth VA. As such, φi = i idea of receiver autonomous integrity monitoring (RAIM) to
indicates that the original allocation of ith measurement will judge the convergence. When the ToA residual obtained by
not be changed. The mismatch alleviation matrix is defined applying the estimated VA’s position is less than the threshold
as t ∈ RN(t)×N(t) , where each element (t )ij takes the value we set for RAIM, the estimate is converged, and this VA is
decided as a known VA.
1, φi = j
(t )ij = (8) When the position estimates of the VAs are unknown, given
0, φi = j (i)
an a priori estimate of the ith VA position PVA , we can linearize
where (·)ij represents the element of the ith row and jth (i) (i)
d̂t with respect to PUE,t and PVA using Taylor expansion,
column of the matrix. Only one nonzero element in each obtaining
row and column of t takes the value 1 and −1 t = Tt .  
When no mismatch is present, t = I. Let  be the diagonal (i) (i) ∂f
d̂t = f PUE,t , P̄VA + PUE,t
matrix consisting of mismatch alleviation matrix for T epochs, ∂PUE,t PUE,t
i.e.,  = diag[1 · · · T ].  
∂f (i) (i)
We assume that the estimated position matrix of UE + (i)
PVA + o PUE,t PVA (11)
and VAs is P̂ with mismatch. The goal of data association ∂PVA P(i)
VA
11276 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

(i) (i) (i)

where PVA = PVA − P̄VA is the location correction of the ith to observation noise. To rectify this problem, we consider
(i) (i) 2 an MSC [50], [51] (i.e., sequential observations) to jointly
VA. o(PUE,t PVA ) = o((PUE,t ) 2
+ (PVA ) ).
After the VAs’ position estimates have converged, we define estimate the locations of VAs and the UE. Because each
that VAs’ positions are known. We assume the position of the position update is independent and similar, we present the
(i) algorithm with a single position estimating step as an example.
ith VA is P̂VA and the a priori estimate of UE is PUE,t in
(i) (i) We use the ToA measurement model with T sequential
epoch t. In this instance, PVA = P̂VA . In linearizing the distance
epochs. Suppose there are M unknown VAs with M measure-
measurement model with respect to PUE,t , we can obtain
ment model for case 1 and K known VAs with K ones for
(i)
  ∂f   case 2 during T epochs. We assume that N is the total number
(i)
d̂t = f PUE,t , P̂VA + PUE,t + o PUE,t 2
∂PUE,t PUE,t of measurements for T epochs used for one position update.
(12) Accordingly, N = T(M + K). The joint measurement model
of the (M + K) VAs with T epochs can be written as
where PUE,t = PUE,t − PUE,t is the location correction of
 
the UE at epoch t. T T
Y = d̄1 T ··· d̄T
T
d1
T
··· dT
For convenience, we record the distance measurement resid- ⎡ ⎤
(i) (i) (i)
ual as d̂t = d̂t − f (PUE,t , P̂VA ) when a VA’s position is U1 ··· O V1
(i) (i) (i) ⎢ .. .. .. .. ⎥
known and as d̄t = d̂t − f (PUE,t , P̄VA ) when its position ⎢ . . . . ⎥
⎢ ⎥
is unknown. Let ⎢O ··· UT VT ⎥
=⎢
⎢ U1
⎥ X = HX (17)
∂f ∂f uit , for Case 1 ⎢ ··· O O⎥ ⎥
vit = , uit = = (13) ⎢ . .. .. ⎥
∂PVA ∂PUE,t uit , ⎣ .. O⎦
(i) for Case 2. .
PVA PUE,t .
With unknown VAs positions, the measurement model at O ··· UT O T(M+K)×2(M+T)
epoch t is
 T where
(1) (M)
d̄t · · · d̄t  T  T
⎡ 1 ⎤ Ut = u1t T ··· uM
T
, Ut = u1t
T
··· uK
T
(18)
ut v1t ··· o t t
⎢ .. ⎥ ⎡ 1 ⎤
= ⎣ ... ..
.
..
. . ⎦
vt ··· o
⎢ .. .. .. ⎥
uM o · · · vM Vt = ⎣ . . . ⎦ (19)
t t M×2(M+1)
 T o ··· vM
t
× PUE,t T PVA (1) T (M) T  T
··· PVA (14)
2(M+1)×1
X = PUE,1 T · · · PUE,T T (1) T (M) T .
PVA · · · PVA (20)
where M denotes the number of measurements corresponding
to VAs with unknown positions. X is the location corrections of VAs and UE.
With known VAs positions, the measurement model is The dimension of the observation matrix H of (17) is T(M+
described as K) × 2(M + T). Clearly, (17) has one and only one solution
 T  T
(1) (K) because the calculated UE’s and VAs’ positions are unique,
d̂t · · · d̂t = u1t t K×2 PUE,t
· · · uK
meaning that the observation matrix is full column rank
(15)
where K is the number of measurements corresponding to VAs T × (M + K) ≥ 2(M + T). (21)
with a posterior estimates.
For convenience, we let The above formula provides the MOC for solving (17). For
 T example, when K = 0 and M = 5, T ≥ 4 can guarantee
d̄t = d̄t(1) · · · d̄t(M) (16a) that (17) has a solution.
 T Equation (21) serves as a reference for determining the
dt = dt(1) · · · dt(K) . (16b) sequential epochs within the estimator for real-time position-
ing. The number of epochs T is the number of observation
B. Real-Time Multipath-Assisted Positioning Based on instances used for one position update, that is, a fraction of
Multistate Constraint discrete time instances. The selection of the number of epochs
can be tailored with flexibility based on the guidance provided
We consider positioning in a 2-D scene which can be by (21). The relationship between one position update and the
easily expanded to a 3-D scene. For 2-D positioning with number of measurement epochs is visually shown in Fig. 1.
an unknown VA, the dimension of the observation matrix When the number of observation epochs satisfies the
of (14) is M × 2(M + 1). Therefore, the observation matrix MOC (21), a solution of (17) is available based on the least-
is nonfull rank with a linear correlation column vector group, squares criterion
leading to insufficient observations for a single epoch. The
constraint of a single observation to the position estimation  T −1 T
is therefore weak and may easily induce divergence due X̂ = Ĥ Ĥ Ĥ Y (22)
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11277

where Ĥ is the estimated observation matrix. The positions of Q, assuming that the observation noise of different paths is
the UE and VAs estimated via joint position estimation with irrelevant. When no mismatch is present, Ĥ = H.
an MSC are then Based on the preceding section, Ĥ is full column rank.
 T T
Therefore, Ĥ Ĥ is a nonsingular matrix. The least-squares
(1) (M)
P̂ = X̂ + P̄UE,1 · · · P̄UE,T P̄VA · · · P̄VA .
estimation of X by applying the mismatch alleviation matrix
(23) to the observation model can be calculated as
 T  −1  T  T −1 T
For each estimation, when N measurements are received,
we obtain the linearized measurement model according X = Ĥ Ĥ Ĥ Y = Ĥ Ĥ Ĥ T Y.
to (17)–(20) and then apply (22) and (23) to estimate the (25)
positions of UE and VAs. In the MSC-MAP method, a position
update is performed every T measurement epoch. We do not Accordingly, an estimation error is generated due to the
need to update positions with future data, and the processing mismatch, namely
delay is determined by T only. In addition, the MSC-MAP  T −1 T  
method has low complexity. Therefore, the MSC-MAP method δ = X − X̂ = Ĥ Ĥ Ĥ T − I Y. (26)
can be considered real time on both the theoretical and
implementation levels at each position update. Naturally, δ includes a nonzero-mean bias; that is, severe
We can realize the estimate quality control of VA’s position positioning error exists in some parts of the UE’s trajectory
by judging the convergence of VA. When a VA’s position due to mismatch.
is still unknown after multiple position updates, the estimate The  is essentially an adjustment of the data association
quality of this VA’s position is poor and the estimate cannot results. We can get the correct association by finding the
converge, which may be caused by scatterers or nonspecular optimal  under any unordered association. However,  has
reflections. We can exclude these nonconverged VAs in the a total of N! possible values. The best solution of  can only
subsequent position updates. be obtained by a brute-force search. Therefore, it is necessary
It should be noted that the least square is used to explore to develop fast mismatch alleviation algorithms.
the distinct constraint of the observation epoch during real-
time state estimation. Upon adding the kinematic model of B. Randomly Selecting Subset-Based Mismatch Alleviation
the UE to the linearized observation model (17), joint position We propose a mismatch alleviation algorithm, namely, RS-
estimation can be easily extended to some filter methods; this Subset, by randomly evaluating some small-scale subsets and
topic is beyond the scope of this article. then finding a mismatch-free subset with the desired size to
obtain the final solution.
IV. DATA A SSOCIATION M ISMATCH We use the posterior residual to evaluate the mismatch
A LLEVIATION A PPROACH alleviation matrix  and estimated correction matrix X̂.
We first formulate the mismatch problem. Then, to meet the Specifically, the residual vector of all measurements is
requirement of real-time processing, two mismatch alleviation   T −1 T 
algorithms are proposed to fulfill distinct computational load w = Y − ĤX̂ = I − Ĥ Ĥ Ĥ Ĥ Y. (27)
and accuracy requirements.
When no mismatch occurs, the posterior residual vector
A. Mismatch Alleviation Problem follows a joint Gaussian distribution with zero mean. The
estimated bias induced by the mismatch can cause nonzero
As stated in Section II-C, to achieve joint position esti-
bias in the means of corresponding residuals. We can thus use
mation of VAs and the UE, it is essential to recover the
the residual sum to assess the quality of the final joint position
observation matrix H in the correct association order. When
estimates of both the UE and VAs.
the ideal data association can be obtained, the row of the
(i) We define a selection vector α = [α1 , . . . , αN ]T , whose
measurement matrix H (17) corresponding to every d̄t (or
(i) ith entry is αi = 0 if the ith measurement is excluded
dt ) should be uit and vit (or uit ). However, with a mismatch from the positioning solution and αi = 1 if it is selected in
in the data association, the ToA measurement corresponding the positioning solution. Therefore, the ideal way to achieve
to the ith VA is wrongly associated with the jth VA, turning mismatch alleviation is to find an optimal selection vector
(i) (i) j j
the row of H corresponding to d̄t (or dt ) into ut and vt and then use a subset of measurements to complete joint
j
(or ut ). This outcome will further lead to a mismatch of the position estimation. To avoid repeating the entire positioning
jth ToA measurements. procedure, we use the chosen residual sum p = α T w to
Considering the mismatch in data association, the measure- evaluate the quality of the final estimate.
ment model becomes Then, we can define an optimization problem as minimizing
the absolute value of the selected residual sum. To effectively
Y = ĤX + n (24)
solve this problem, we convert it to minimize the squared sum
where Ĥ denotes the estimated observation matrix derived of residuals, wherein
from data association results; and n denotes the Gaussian  2   
distributed noise vector with mean 0 and covariance matrix min p2 = α T w = α T w α T w . (28)
α
11278 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

Because α T w is a scalar, we have α T w = wT α. Let W = where

wwT ; the mismatch alleviation optimization problem with ∂diag(ν)
constraints can be written as = diag(e1 , e2 , . . . , en ) ∀i ∈ {1, . . . , N} (39)
∂νi
min α T Wα (29) 1, j = i
α ej = ∀j ∈ {1, . . . , N}. (40)
0, j = i
s.t. αi ∈ {0, 1} ∀i ∈ {1, . . . , N} (30)

αi ≥ 2(M + T) ∀i ∈ {1, . . . , N}. (31) Therefore, the optimality conditions of (37) obtained from the
i
gradient are
   
Constraint (31) represents that the number of selected mea- 1 ∂φA ν *
I νi =
*
+ = 0 ∀i ∈ {1, . . . , N}. (41)
surements needs to satisfy the MOC (21). εA ∂νi
As displayed in (29)–(31), the mismatch alleviation problem
is a 0-1 integer programming problem. Therefore, we use Let A = [W + diag(ν * )]−1 = [aij ]N×N ; then, we have
 
approximation algorithms to obtain a relatively good solution I v∗i = 1/εA − aii = 0 ∀i ∈ {1, . . . , N}. (42)
to this problem quickly.
The algorithm is split into two steps. In the first step, a εA For any εA > 0, the optimal solution ν ∗ of (37) can be
suboptimal lower bound of the original problem is obtained found from (42). As such, let εA an error threshold, we can
by solving the Lagrange dual problem, which can provide a acquire the εA suboptimal solution of (35). Additionally, the
reliable guideline for evaluating the subset in the subsequent εA suboptimal lower bound d∗ for the optimal solution p∗ of
step. In the second step, we use the lower bound to determine the original problem (32) can be obtained by d∗ = −1T ν ∗ .
whether mismatched data exist in a randomly selected subset Although there is a duality gap between p∗ and d∗ due to the
of measurements. Next, we merge several subsets without nonconvexity of (32) (i.e., p∗ > d∗ ), d∗ can still provide a
mismatches to obtain the final subset used for positioning. reliable guideline for randomly selecting a subset in the next
We first find a εA suboptimal lower bound of the original step.
problem. For the convenience of a symmetrical value domain Then, we use d∗ to evaluate the quality of selected subsets,
in the optimization problem, we define an intermediate vari- and then compose a final selection with a specified size.
able z = 2(α − [1/2]), then zi ∈ {−1, 1} ∀i ∈ {1, . . . , N}. The According to the union set, we can determine the mismatch
optimal problem (29) can be rewritten as alleviation matrix  = αα T via selection vector α. The
modified correction matrix is
min zT Wz (32)  T −1 T
z
s.t. zi ∈ (−1, 1) ∀i ∈ {1, . . . , N} (33) X̂ = Ĥ Ĥ Ĥ T Y. (43)

(zi + 1) ≥ 4(M + T) ∀i ∈ {1, . . . , N}. (34) Finally, we can obtain the estimated positions of UE and VAs
i (i.e., P̂) according to (23).
Although (32) is not convex, the dual function of (32) is The detailed process is presented in Algorithm 1. We set
concave because it is the pointwise infimum of a family of a mismatch judgment threshold η = d∗ (εA ) (line 4). When
affine functions. In this case, we can obtain a lower bound the cardinality of the correctly matched measurement set C
of (32) using the dual function. (denoted by |C|) is less than 2(T + M) (i.e., the set size under
According to [52], the Lagrange dual problem associated the MOC), m(m < n) measurements are randomly selected
with the problem (32) is from the set Y = [y1 , y2 , . . . , yn ]T of observations to form
subsets Bk , where k = {1, 2, . . .} (line 7). zi (i ∈ {1, 2, . . . , n})
min 1T ν (35) corresponding to the measurements in Bk are equal to 1;
ν
s.t. −W − diag(ν) 0 (36) otherwise, zi = −1 (lines 8–12). Afterward, we calculate the
average of the squared sum of the residual rBk of selected
where ν is Lagrange multiplier vector and diag(·) represents measurements according to the selection vector zBk of Bk
diagonal matrix. Problem (35) is a semidefinite program in (line 13). If rBk ≤ η, we assume that none of the measurements
inequality form, which can be solved via the barrier method. in the kth subset are mismatched and thus put the elements of
Using the generalized logarithm log det X, we have the Bk into C; if rBk > η, then Bk has mismatched measurements.
barrier function for problem φA (ν) = − log det |W + diag(ν)| When an identical measurement yi appears in l data mismatch
with domφA = {ν|W + diag(ν) 0}. Problem (35) can be subsets, yi is considered as a mismatched measurement, which
approximately formulated as is deleted from the Y to decrease the possibility of selecting a
1 T  mismatched measurement in a subsequent step (lines 17–20).
min 1 ν − log det|W + diag(ν)| (37)
ν εA When the number of elements in C exceeds 2(T + M), we use
C as the final selected subset to estimate the positions of UE
where εA > 0 is the parameter used to determine the quality
and VAs.
of the approximation, which improves as εA decreases. For a
In the RS-subset algorithm, a new measurement set is
strictly feasible ν, the gradient of φA (ν) is
 formed by a randomly selected subset of measurements with
∂φA (ν)  −1 ∂diag(ν) small residuals, whose elements are used to locate VAs and
= −tr W + diag(ν)
∂νi ∂νi the UE. Therefore, the RS-Subset algorithm can quickly find
∀i ∈ {1, . . . , N} (38) a suboptimal solution compared to the brute-force search.
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11279

Algorithm 1: RS-Subsets Algorithm among the total selection number becomes larger. In addition,
the accuracy of the solution to this algorithm is less than that
Input: Estimated measurement matrix Ĥ, measurement
of the global optimal solution due to the duality gap between
epoch T, number of unknown VA M,
the original problem and the dual problem. This algorithm is
measurements matrix Y, parameters εA , N, m, and
thus better suited to scenarios with constrained computational
l
resources.
Output: Mismatch alleviated set C and estimated
positions P̂ of UE and VAs
1 Calculate the residual vector w when the measurement C. Constraint Relaxation-Based Mismatch Alleviation
matrix is Ĥ according to (27); The RS-subset algorithm proposed in Section IV-B can
2 Get the residual matrix W = wwT ; improve the positioning accuracy. Because of the random
3 Obtain the suboptimal lower bound d∗ (εA ) = −1T ν ∗ with method used in the algorithm, the performance is not always
the error threshold εA according to (38) - (42); stable due to different ratios of mismatched pairs. Thus, in this
4 Set the mismatch judgment threshold η = d∗ ; section, we propose another mismatch alleviation algorithm
5 Initialize Bk = 0, C = 0, s = 0, k = 0 based on CR that can obtain a stable near-optimal solution.
6 while |C| < 2(T + M) do While retaining all measurements, we relax the range of values
7 Randomly select m(m < n) measurements yi from of all αi to [0, 1], which can then be used as weights for
measurement set Y to form a subset Bk ; the original measurements. To retain the subset size constraint
8 for i = 1 to N do from (31), in the relaxed problem, we constrain the sum of
9 if yi ∈ Bk then weights to be u.
10 zi = 1; Therefore, problem (29) is relaxed as
11 else min α T Wα (44)
12 zi = −1; % determine the selection vector of α
subset s.t. 0 ≤ αi ≤ 1 ∀i ∈ {1, . . . , N} (45)
N
13 rBk = (zBk T + 1)T W(zBk + 1)/(4|Bk |); % average of αi = u (46)
squared sum of residual of measurements in Bk i=1
14 if rBk ≤ η then
which is a convex quadratic program problem with inequality
15 C = C ∪ Bk ; % put the candidate measurement
constraints. The optimal solution to this problem exists at
into C
16 else the extreme point of an interval due to inequality con-
17 for each measurement in the subset Bk do straints; it is therefore challenging to provide the analytical
18 si ← si + 1; % the number of times solution directly. We solve problem (44) using the barrier
measurement yi is included in mismatch method combined with Newton’s method, which is suitable for
subsets plus one quadratic problem and has a feasible solution and polynomial
19 if si ≥ l then complexity.
20 delete measurement yi from Y; The barrier function for problem (44) is

n 
n
21 k ← k + 1; % start another round of selection until φB (α) = − log(1 − α i ) − log(α i ). (47)
|C| reaches threshold i=1 i=1
22 if yi ∈ C then Problem (44) with constraints is approximately formulated as
23 α i = 1;
else min h(α) = tB α T Wα + φB (α) (48)
24 α
25 α i = 0; 
N

26 = αα T ; %mismatch alleviation matrix s.t. αi = u ∀i ∈ {1, . . . , N} (49)

T T i=1
27 X̂ = (Ĥ Ĥ)−1 Ĥ T Y;
%modified correction matrix
28 return mismatch alleviated set C and estimated positions where tB > 0.
We can obtain Newton’s descent direction α and
P̂ of UE and VAs.
decrement λ2 by solving the Karush–Kuhn–Tucker (KKT)
equations, which are described in the Appendix.
After determining α by solving the above optimization
The probability that a randomly selected subset contains no problem, we can similarly find the mismatch alleviation matrix
mismatch measurements is positively correlated to the ratio  and estimate the positions P̂ of UE and VAs according
of the mismatched pair in the data association results. Thus, to (43) and (23).
finding this subset has low computational complexity when the The detailed process appears in Algorithm 2. In the inner
number of mismatched pairs is small. On the contrary, when iterations, Newton’s method is used to obtain the central point
the mismatch problem is severe, the probability of obtaining vector α corresponding to each β (lines 5–9). The step length
a so-called clean subset in a single selection is slight, and the is obtained by backtracking the line search (line 7) [52]. The
time expectation of finding a clean subset at the required scale descent direction and decrement is computed based on the
11280 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

Algorithm 2: CR Algorithm
Input: Estimated measurement matrix Ĥ, number of
measurements N, constraint on the sum of
weights u, parameters εB , , and ρ
Output: Modified correction matrix X
1 Calculate the residual vector w when the measurement
matrix is Ĥ according to (27);
2 Get the residual matrix W = wwT
(0)
3 Initialize α i (0) = u/N, β = 1, α ∗ = 0, λ2 = 0, tB = 1
4 if 2N/tB ≥ εB then
5 repeat
6 α: = α + βα ∗ ;
7 update step length β of inner iteration by
backtracking line search;
8 Compute the Newton descent direction α ∗ and
Newton decrement λ2 according
to (50)–(51), (57) and (58);
9 until λ2 /2 ≤ ;
10 Update the vector α new = α;
11 tB : = ρtB ; Fig. 2. Example of the false alarm and miss-detection. Blue squares and pink
circles indicate VAs and measurements separately. The blue solid lines and
12  = α new α new T ; % determine the mismatch alleviation dashed lines indicate the correct data association and mismatch, respectively.
matrix  of mismatched observation matrix H The false alarm is equivalent to the mismatch.
T T
13 X̂ = (Ĥ Ĥ)−1 Ĥ Y; %modified correction matrix
14 Compute and return the estimated positions P̂ of UE and
VAs. measurements and does not affect the positioning results. In
addition, the false alarm in data association can be expressed
as the association of VA with noise that is mistaken as a
measurement. Therefore, the false alarm can also be equated to
first-order derivative vector and the second-order derivative
a mismatch, meaning that the proposed mismatch alleviation
matrix (line 8). At each outer iteration (lines 4–11), we
algorithms can also handle the false alarm. An example of
calculate a new α corresponding to the current tB starting from
the false alarm and miss-detection is depicted in Fig. 2. We
the previously computed α until tB , increased by the scale
assume the measurements are arranged by order of associated
factor ρ, surpasses the threshold. The final obtained α new is
VAs, i.e., the ith VA should be associated with the ith
the measurement weight that minimizes the weighted squared (i)
measurement τt at epoch t. Blue squares and pink circles
sum of the residual of all measurements.
indicate VAs and measurements separately. The blue solid
The CR algorithm can obtain the global near-optimal
lines and dashed lines indicate the correct data association and
solution of the original problem. However, the first- and
mismatch, respectively. A false alarm may be associated with
second-order derivatives must be calculated, as must the
a fixed VA (e.g., fixed VA 2 associated with false alarm 2f), or
inverse of the matrix. Subsequent positioning also needs
generate a new VA (e.g., new VA 5 generated by false alarm
to use all weighted measurements. Thus, this algorithm is
5f). The position estimate of the new VA generated by a false
computationally intensive and its complexity grows along with
alarm cannot converge (i.e., the latter case), so we can exclude
the number of measurements. Compared with the RS-subset
it by estimating quality control. For the former case, the false
algorithm proposed in the previous section, the computational
alarm is equivalent to the mismatch shown in the left picture
load of the CR-based algorithm is not sensitive to the ratio of
of Fig. 2.
mismatched pairs. Therefore, this algorithm is more suitable
Furthermore, for clutter measurement, the estimated posi-
for scenarios with sufficient computational resources and high
tion of the corresponding VA may not converge, and we
requirements for positioning performance.
can exclude the nonconverging VA from the UE positioning
through the estimate quality control of the VA’s position.
D. Processing of Non-Gaussian Impairments Therefore, the proposed algorithms are directly applicable to
Measurements provided by channel estimators not only real-world data with non-Gaussian impairments.
contain Gaussian measurement noise (related to system param-
eters and channel conditions) but also missed detection (a
V. P ERFORMANCE E VALUATION
measurement related to a VA is not detected) and false
alarms/clutter [44], [53], [54], [55]. A. Performance of MSC-MAP
For miss-detection, the unobserved measurements and their We assess the performance of the proposed MSC-MAP
corresponding VAs do not participate in the UE’s posi- algorithm via a 2-D scenario simulation (see Fig. 3), with a
tioning. So, the miss-detection only reduces the number of moving UE and two static BSs. The simulation parameters are
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11281

signal-to-noise ratio (SNR) at different bandwidths, respec-

tively. We define a VA as fixed when the posterior residual
of measurements corresponding to the VA is less than a
predefined threshold. In the simulation, the residual threshold
for a fixed VA is 0.1 m. Fig. 4(a) shows that the number
of fixed VAs increases first before becoming stable, as the
number of visible VAs is limited in the UE trajectory. The
number of estimated convergent VAs is about 11. We use
the median to evaluate the positioning error of fixed VAs to
exclude the effects of unfixed VAs. As presented in Fig. 4(b),
when the SNR is low, the number of VAs that are fixed is
small. The observation quality of such VAs is high, however,
leading the positioning accuracy of fixed VAs to be high as
well. While the quantity of fixed VAs rises with SNR, there
are instances of poor quality among them, thereby influencing
the precision of positioning results. The accuracy of ToA
measurements steadily enhances with increasing SNR, leading
Fig. 3. Simulated indoor scenario with two fixed BSs and a moving UE. to a gradual enhancement in the quality of fixed VAs and a
UE is moving on the track as indicated in the purple dashed line. The green reduction in positioning errors.
line indicates the UE position estimations of the proposed multipath-assisted The UE position can also be estimated well using the MSC-
positioning method. First- and second-order VAs are used in the simulations.
For ease of representation, we only mark the true positions of first-order VAs, MAP, as indicated in Fig. 4(c). The MSE of the UE position
which are indicated by red crosses. Orange circles represent estimated VAs. is below 0.2 m for most bandwidth settings when the SNR is
The proposed MSC-MAP method can estimate convergent VAs’ positions and greater than 16 dB. The MSE mostly remains stable due to the
UE’s trajectory with high accuracy.
estimated UE position residing on a circle, centered at a VA
with a radius defined by a ToA measurement. Consequently,
decided according to the typical indoor pedestrian scenario. the accuracy of VA positioning significantly impacts the UE’s
The UE is moving at a speed about 1 m/s along a polyline positioning outcome. The estimated VA positions used for UE
trajectory around the building, indicated by the green dotted positioning contain error, further leading to a lower bound for
line. The entire trajectory is about 85 m long. The sampling the UE-positioning MSE.
frequency of ToA measurements is 1 Hz, and the interval Fig. 5 plots the minimum observation epochs for a different
of measurement epoch is 1 s. At each epoch, the two BSs number of unknown VAs. Any integer epoch and number
transmit signals at a bandwidth of 100 MHz with a carrier greater than the minimum can satisfy the constraint. As
frequency of fc = 2.4 GHz, which is received and processed shown in Fig. 5, the number of minimum observation epochs
by the UE. For measured signals, the ToA estimation noises decreases with the number of known VAs. With insufficient
are generated according to [56] with subcarrier frequency fsc = known VA, the higher the number of unknown VA, the more
30 kHz. The performance is evaluated through Monte Carlo observation epoch is required.
simulations with 5000 independent runs.
As mentioned in Section II, the VA position of the multipath
is the mirror point of the source point (e.g., the BS) against the B. Performance of Proposed Mismatch Alleviation Methods
reflecting surface. Due to attenuation through every reflection This section first demonstrates the impact of a mismatch on
of the transmitted signal, we only consider at most second- positioning performance and next evaluates the proposed two
order VAs in the simulation. The number of the first- and mismatch alleviation algorithms (i.e., the RS-subset algorithm
second-order VAs are 26 and 396 separately. For convenience, and the CR algorithm). The simulation scenario and signal
we mark all first-order VAs with red crosses in Fig. 3. parameters are the same as in the prior section.
Fig. 3 depicts the estimated positions of the UE compared 1) Analysis of Mismatch Effect on Positioning Performance:
with the UE track and VAs compared with ideal VA positions. We first evaluate the percentage of mismatches using the data
The green line indicates the UE position estimations. The VAs association method in [30]. For the sake of comparison, we
are denoted by orange circles. For scatterers or nonspecular also consider the low sampling frequency scenario of 0.5 Hz.
reflections, the positions of corresponding VAs may vary with Thus, we use two different pace settings: 1) short pace (SP)
different UE locations. The estimated location of nonspecular and 2) long pace (LP), whose step sizes are 1 and 2 m, respec-
VAs may therefore not converge, and specular VAs can be tively. Fig. 6 illustrates the average mismatch rate for the two
identified through convergence evaluation methods. In this settings. In SP and LP, the average mismatch rates are 22.1%
scenario, results demonstrate that the proposed MSC-MAP can and 33.5% obtained by averaging over the 100 simulation runs.
estimate convergent VAs’ positions and UE’s trajectory with This result reveals that the mismatched association accounts
high accuracy. for a large proportion of all measurements. As the step size
The performance for positioning VAs is further pictured increases, the mismatch rate also rises.
in Fig. 4, displaying (a) the number of fixed VAs and To assess the effect of mismatch, Fig. 7(a) shows the
(b) the median positioning error of fixed VAs versus the cumulative distribution function (CDF) of positioning error
11282 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

(a) (b) (c)

Fig. 4. Results for positioning VAs and UE using MOC positioning versus SNR with different bandwidths: (a) average number of estimated VAs; (b) median
positioning error of fixed VAs; and (c) MSE of estimated UE position. In (a) and (b), the number of estimated VAs grows and the median positioning error
of VAs declines with SNR. However, with small bandwidth, the positioning error increases with the number of VAs when VA’s number is less than 8. In (c),
the MSE of the UE position declines with SNR.

reaffirms the critical importance of addressing data association

mismatches in multipath-assisted positioning.
2) Performance of RS-Subset Algorithm: In the simulation
scenario, the number of first-order VAs is 26, and the number
of second-order VAs is 396. The number of visible VAs
fluctuates around 50 in each epoch. Due to the MOC, we set
the target size of the subset for positioning to 54. For each
choice, the size of the randomly selected subset is set to 4.
As illustrated in Fig. 7(b), the RS-subset algorithm can
significantly enhance the positioning accuracy compared to
scenarios without data mismatch alleviation. To explore this
improvement, we experimented with four distinct parameter
Fig. 5. Minimum observation epochs to achieve VA-assisted positioning settings for the mismatch proportion Rmm and the error
and make estimated positions of UE and unknown VAs converge for the threshold εA . Notably, when one-quarter of measurements are
different number of unknown VAs. M represents the number of unknown VAs mismatched, and εA is set at 1.0, the 1-sigma positioning
and the vertical coordinate represents the number of minimum observation
epochs. Different colors of curves represent the minimum observation epochs error remains below 0.8 m. Across all four cases, varying in
for different numbers of unknown VAs, which decreases with the number of mismatch proportions, the RS-subset algorithm consistently
known VAs. reduces the positioning error to one-third of that observed in
scenarios without mismatch alleviation.
Furthermore, within the RS-subset algorithm, we have the
flexibility to manage the upper limit of random selection
error by fine-tuning the parameter εA . Decreasing εA enhances
positioning precision, albeit at the cost of slightly increased
computational complexity. Additionally, it’s important to note
that the runtime of the RS-subset algorithm scales with the
number of measurements m contained within each subset.
3) Performance of CR Algorithm: To confirm the elimi-
nation performance of the CR algorithm, we use empirical
values as α = 0.01, β = 0.5,  = 10−5 , and ρ = 15.
Fig. 7(c) presents the CDF of the positioning error for the
CR algorithm versus the original method without mismatch
Fig. 6. Average mismatch rate for two different pace settings. The elimination. The CR algorithm’s positioning error falls to
mismatched association accounts for a large proportion of all measurements, nearly one-quarter of that of the original method, revealing the
and the mismatch rate increases with the step size.
effectiveness of this algorithm in eliminating the impact of a
mismatch.
when mismatched measurements account for 16%, 25%, 33%, The CR algorithm also performs similarly to the RS-subset
and 42%. Positioning performance deteriorates rapidly as algorithm for the error threshold εA = 1.0. While the RS-
the percentage of the mismatch increases. Specifically, when subset algorithm may not attain the optimal solution, it excels
42% of the associated measurements are mismatched, the in striking a balance between positioning performance and
1-sigma (68.27%) positioning errors surpass 3 m and may runtime through the flexible adjustment of εA . In contrast, the
extend to as much as 6.5 m. Therefore, as described in CR algorithm distinguishes itself by excelling in mismatch
Section IV, the data association mismatch can significantly elimination, resulting in more accurate positioning as it
impede positioning performance. Consequently, our simulation approaches a near-optimal solution.
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11283

(a) (b) (c)

Fig. 7. Results for mismatch analysis and alleviation: CDF of positioning error with different mismatch proportions Rmm (a) of measurement in original data
association; (b) for random selection in RS-subset algorithm with different error thresholds εA ; and (c) for CR algorithm (indicated by solid lines) compared
with the original algorithm without mismatch alleviation (indicated by dashed lines). In (a), positioning performance deteriorates rapidly as the percentage
of the mismatch increases. In (b) and (c), both mismatch alleviation algorithms can substantially reduce the positioning error. The RS-subset algorithm can
control the positioning accuracy by adjusting εA . The CR algorithm has better alleviation performance.

Fig. 8. Result of performance under non-Gaussian impairments. Two

different parameter settings are considered: 1) H: high SNR with false alarm Fig. 9. Result of performance comparison with state-of-the-art methods: CDF
probability 0.1 and miss-detection probability 0.2 and 2) L: low SNR with of positioning error of RS-subset, CR, STAMP, and BP-SLAM. The solid and
false alarm probability 0.2 and miss-detection probability 0.3. The proposed dashed lines represent SP and LP, respectively. The RS-subset and CR can
two algorithms are both effective in improving the positioning performance both substantially improve positioning performance by solving the mismatch
in a scenario where false alarms and miss-detection exist. problem in both settings. The positioning performance of BP-SLAM decreases
sharply when the step size increases.
4) Performance Under Non-Gaussian Impairments: Then,
we verify the performance of the proposed mismatch alle- performance compared to the CR algorithm. Fig. 9 shows the
viation algorithms in the scenario with false alarms and CDF of the positioning error of the BP-SLAM, STAMP, and
miss-detection. We consider two different parameter settings. the proposed two mismatch alleviation algorithms. The RS-
One is high SNR with false alarm probability 0.1 and miss- subset and CR algorithms reduce the positioning error of the
detection probability 0.2. The other is low SNR with false STAMP algorithm by 69.0% and 77.3% for the SP, and the
alarm probability 0.2 and miss-detection probability 0.3. The reductions are 23.8% and 44.8% for the LP. For the BP-SLAM,
high SNR and low SNR are denoted as H and L, respec- the positioning performance is stable when the step size is
tively, in Fig. 8. As demonstrated in Fig. 8, the RS-subset small. The proposed two algorithms both slightly improve
algorithm and CR algorithm are both effective in improving the positioning performance, reducing the 1-sigma positioning
the positioning performance in a scenario where false alarms error from 0.74 m to 0.42 and 0.57 m, respectively. However,
and miss-detection exist. The result coincides with our analysis when the step size increases, the correlation between adjacent
of treating non-Gaussian impairments as mismatches. sampling points weakens, resulting in a sharp decline in the
5) Comparison With State-of-the-Art Methods: Next, we positioning performance of BP-SLAM, with a positioning
compare the CDF of the positioning error of our proposed error as high as 2.33 m. The two proposed algorithms improve
algorithm with those of STAMP [30] and BP-SLAM [35], in the positioning performance by 27.5% and 47.4%.
which the hard data association method and probabilistic data Therefore, the following further conclusions can be drawn:
association method are used, respectively. We also consider 1) the proposed two mismatch alleviation algorithms can
two different pace settings: dubbed SP and LP. We set εA = 2.0 substantially improve positioning performance by solving the
for the RS-subset algorithm with a slightly weaker alleviation mismatch problem in both settings and 2) BP-SLAM requires
11284 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

TABLE II
AVERAGE RUNTIME P ER S IMULATION RUNS can be achieved using two BSs in a complex environment
for an SNR greater than 16 dB with 20-MHz bandwidth. In
addition, we validate that mismatched measurements indeed
constitute a significant portion and evaluate their impact.
Mismatch alleviation approaches can reduce the positioning
a strong spatial correlation of adjacent sampling points. The error by 69% even when the percentage of mismatched
correlation weakens when the step size increases. Hence, BP- measurement data is as high as 42%, significantly improving
SLAM and other probabilistic data association methods are the performance of multipath-assisted positioning. Moreover,
unsuitable for scenarios with sparse sampling points. However, our proposed algorithms can substantially improve positioning
the proposed MSC estimator can maintain the constraint performance when confronted with measurement mismatches
between sampling points, thus keeping a stable positioning while incurring minimal additional computation time com-
performance when the step size increases. pared to existing state-of-the-art methods employing different
In addition, we compare the complexity of the four methods data association techniques.
using the average time per simulation run (average over In this article, we have not used angular measurements,
100 simulation runs), shown in Table II. The simulations are which can also be accurately resolved by 5G MIMO equip-
implemented by MATLAB R2017b on an Intel Core i5-8265U ment. Our future research will explore the incorporation of
@1.60 GHz with 8-GB RAM. It is seen from Table II that angle information to enhance positioning accuracy further.
the complexity of STAMP, RS-subset, and CR is similar. The Additionally, it is important to note that our performance
addition of mismatch alleviation adds little to no computation evaluations have been based on synthetic measurements, and
time. In addition, due to complicated probability calculation, further investigations are warranted using real-world data.
the complexity of BP-SLAM is very high, and the runtime is Furthermore, an exciting avenue for future research lies in
about 15 times of the other three algorithms. Therefore, BP- exploring RIS-assisted positioning.
SLAM is challenging to achieve real-time positioning in terms
of implementation. A PPENDIX
N EWTON ’ S D ESCENT D IRECTION AND D ECREMENT
VI. C ONCLUSION Here, we introduce the detailed derivation of Newton’s
We propose a novel ToA-based multipath-assisted real- descent direction and decrement.
time and low computation complexity positioning algorithm, The first-order derivative and the Hessian matrix of h(α)
namely, MSC-MAP, taking advantage of delay measurements are needed. By taking partial derivatives of αi , the first-order
of reflected signals as additional spatial observations. We model derivative vector of h(α) is

the propagation paths of reflected signals as VAs and estimate the ∂h ∂h
positions of UE and VAs via an MSC estimator to deal with ∇h(α) = ,...,
∂α1 ∂αn
the insufficient observation. We use estimation quality control
∂h n
1 1
to find some VAs in the environment that can be used to assist = 2t wij αj − − ∀i ∈ {1, . . . , N}. (50)
positioning without using strict specular reflection assumption. ∂αi αi − 1 αi
j=1
Based on the observation model, we analyze the MOC for the
proposed estimator. Once the number of observation epochs The second-order derivative is then given by
used in a single estimating step satisfies the MOC, the converged ∂ 2h 1 1
estimations of UE’s positions can be obtained. = 2twii + + 2 ∀i ∈ {1, . . . , N} (51a)
∂αi 2
(αi − 1) 2 αi
We then propose two mismatch alleviation approximation
∂ 2h
algorithms to handle mismatches in data association. Finding = 2twij ∀i ∈ {1, . . . , N}. (51b)
the optimal mismatch alleviation matrix needs a brute-force ∂αi ∂αj
search. The proposed algorithms meet the requirement of real- Thus, the Hessian matrix of h(·) is ∇ 2 h(α) ∈ RN×N with
time processing. Specifically, the first strategy is based on elements {∇ 2 h(α)}i,j = ∂ 2 h/∂αi 2 if i = j and {∇ 2 h(α)}i,j =
random selection. It can obtain a suboptimal solution with light
∂ 2 h/∂αi ∂αj if i = j.
computational complexity and achieve a tradeoff between time
According to [52], we replace problem (45) with its second-
and accuracy. The second method based on CR can obtain
order Taylor approximation near α and get
a near-optimal solution, but it requires greater computational
resources. For practical use, we can choose a mismatch allevia- min h(α + α)
α
tion algorithm in accordance with computational resources and
1
levels of mismatch. The proposed algorithms are also efficient = h(α) + ∇h(α)T α + α T ∇ 2 h(α)α (52)
for signal with non-Gaussian impairments that are common in 2
real-world data. s.t. 1T (α + α) = u. (53)
Simulation results using synthetic data show that the Next, we have the following KKT conditions:
proposed algorithms estimate the trajectory of a moving UE ⎧
with high accuracy and robustness, even when encountering a ⎨ ∇α L(α + α, ν)
high proportion of mismatched measurements in data associa- = ∇ 2 h(α) · α + ∇h(α) + ν · 1T = 0 (54)
⎩ T
tion. An accuracy rate below 0.2 m in the simulated scenario 1 · α = 1T (α + α) − 1T · α = 0 (55)
XU et al.: MULTISTATE CONSTRAINT MULTIPATH-ASSISTED POSITIONING AND MISMATCH ALLEVIATION 11285

where ν is the Lagrange multiplier associated with [16] D. Jin, F. Yin, A. M. Zoubir, and H. C. So, “Exploiting sparsity of
problem (53). ranging biases for NLOS mitigation,” IEEE Trans. Signal Process.,
vol. 69, pp. 3782–3795, 2021.
By solving (54), we have [17] G. Dumphart, R. Kramer, R. Heyn, M. Kuhn, and A. Wittneben,
  “Pairwise node localization from differences in their UWB channels to
α ∗ = −∇ 2 h(α)−1 ∇h(α) + ν ∗ · 1T . (56) observer nodes,” IEEE Trans. Signal Process., vol. 70, pp. 1576–1592,
2022.
We next substitute (56) into (55) to obtain the Lagrange [18] Z. Wu, Y. Li, X. Meng, X. Lv, and Q. Guo, “A minimum joint
multiplier error entropy-based Localization method in mixed LOS/NLOS environ-
ments,” IEEE Internet Things J., vol. 10, no. 22, pp. 19913–19924,
1T · ∇ 2 h(α)−1 · ∇h(α) Nov. 2023.
ν∗ = − . (57)
1T · ∇ 2 h(α)−1 · 1
[19] N. Rogel, D. Raphaeli, and O. Bialer, “Time of arrival and angle of
arrival estimation algorithm in dense multipath,” IEEE Trans. Signal
Then, α ∗ can be calculated by substituting (57) into (56). Process., vol. 69, pp. 5907–5919, 2021.
[20] K. Witrisal et al., “High-accuracy localization for assisted living: 5G
The Newton decrement is [52] systems will turn multipath channels from foe to friend,” IEEE Signal
Process. Mag., vol. 33, no. 2, pp. 59–70, Mar. 2016.
λ2 = α T · ∇ 2 h(α) · α
T  T  [21] R. Mendrzik, H. Wymeersch, G. Bauch, and Z. Abu-Shaban,
 
= ∇h(α) + ν ∗ · 1T · ∇h(α)−1 · ∇h(α) + ν ∗ · 1T .
“Harnessing NLOS components for position and orientation estimation
in 5G millimeter wave MIMO,” IEEE Trans. Wireless Commun., vol. 18,
(58) no. 1, pp. 93–107, Jan. 2019.
[22] J. Borish, “Extension of the image model to arbitrary polyhedra,” J.
Acoust. Soc. Amer., vol. 75, no. 6, pp. 1827–1836, 1984.
R EFERENCES [23] M. Z. Win, Y. Shen, and W. Dai, “A theoretical foundation of network
localization and navigation,” Proc. IEEE, vol. 106, no. 7, pp. 1136–1165,
[1] H. Chen, H. Sarieddeen, T. Ballal, H. Wymeersch, M.-S. Alouini, and Jul. 2018.
T. Y. Al-Naffouri, “A tutorial on terahertz-band localization for 6G [24] H. Naseri and V. Koivunen, “Cooperative simultaneous localization
communication systems,” IEEE Commun. Surveys Tuts., vol. 24, no. 3, and mapping by exploiting multipath propagation,” IEEE Trans. Signal
pp. 1780–1815, 3rd Quart., 2022. Process., vol. 65, no. 1, pp. 200–211, Jan. 2017.
[2] W. Shao, H. Luo, F. Zhao, H. Tian, S. Yan, and A. Crivello, “Accurate [25] M. Kreković, I. Dokmanić, and M. Vetterli, “Shapes from echoes:
indoor positioning using temporal-spatial constraints based on Wi-Fi Uniqueness from point-to-plane distance matrices,” IEEE Trans. Signal
fine time measurements,” IEEE Internet Things J., vol. 7, no. 11, Process., vol. 68, pp. 2480–2498, 2020.
pp. 11006–11019, Nov. 2020. [26] E. Leitinger and F. Meyer, “Data fusion for multipath-based SLAM,” in
[3] Y. Yu et al., “A novel 3-D indoor localization algorithm based on Proc. 54th Asilomar Conf. Signals Syst. Comput, 2020, pp. 934–939.
BLE and multiple sensors,” IEEE Internet Things J., vol. 8, no. 11,
[27] J. Kulmer, E. Leitinger, S. Grebien, and K. Witrisal, “Anchorless
pp. 9359–9372, Jun. 2021.
cooperative tracking using multipath channel information,” IEEE Trans.
[4] S. Zhao, X.-P. Zhang, X. Cui, and M. Lu, “A new TOA localiza-
Wireless Commun., vol. 17, no. 4, pp. 2262–2275, Apr. 2018.
tion and synchronization system with virtually synchronized periodic
[28] Y. Wang, W. Ren, L. Cheng, and J. Zou, “A grey model and mixture
asymmetric ranging network,” IEEE Internet Things J., vol. 8, no. 11,
Gaussian residual analysis-based position estimator in an indoor envi-
pp. 9030–9044, Jun. 2021.
ronment,” Sensors, vol. 20, no. 14, p. 3941, 2020.
[5] J. A. del Peral-Rosado, R. Raulefs, J. A. López-Salcedo, and
G. Seco-Granados, “Survey of cellular mobile radio localization meth- [29] C.-H. Park and J.-H. Chang, “Robust localization based on ML-type,
ods: From 1G to 5G,” IEEE Commun. Surveys Tuts., vol. 20, no. 2, multi-stage ML-type, and extrapolated single propagation UKF methods
pp. 1124–1148, 2nd Quart., 2018. under mixed LOS/NLOS conditions,” IEEE Trans. Wireless Commun.,
[6] H.-J. Shao, X.-P. Zhang, and Z. Wang, “Efficient closed-form algorithms vol. 19, no. 9, pp. 5819–5832, Sep. 2020.
for AOA based self-localization of sensor nodes using auxiliary vari- [30] L. Li and J. L. Krolik, “Simultaneous target and multipath position-
ables,” IEEE Trans. Signal Process., vol. 62, no. 10, pp. 2580–2594, ing,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 1, pp. 153–165,
May 2014. Feb. 2014.
[7] Z. Wang, J.-A. Luo, and X.-P. Zhang, “A novel location-penalized max- [31] R. Amiri et al., “Indoor environment learning via RF-mapping,” IEEE
imum likelihood estimator for bearing-only target localization,” IEEE J. Sel. Areas Commun., vol. 41, no. 6, pp. 1859–1872, Jun. 2023.
Trans. Signal Process., vol. 60, no. 12, pp. 6166–6181, Dec. 2012. [32] X. Chu, Z. Lu, D. Gesbert, L. Wang, and X. Wen, “Vehicle localization
[8] A. Elzanaty, A. Guerra, F. Guidi, and M.-S. Alouini, “Reconfigurable via cooperative channel mapping,” IEEE Trans. Veh. Technol., vol. 70,
intelligent surfaces for localization: Position and orientation error no. 6, pp. 5719–5733, Jun. 2021.
bounds,” IEEE Trans. Signal Process., vol. 69, pp. 5386–5402, 2021. [33] J. Xiong, J. W. Cheong, Y. Ding, Z. Xiong, and A. G. Dempster,
[9] S. Zhao, X.-P. Zhang, X. Cui, and M. Lu, “Optimal two-way TOA “Efficient distributed particle filter for robust range-only SLAM,” IEEE
localization and synchronization for moving user devices with clock Internet Things J., vol. 9, no. 21, pp. 21932–21945, Nov. 2022.
drift,” IEEE Trans. Veh. Technol., vol. 70, no. 8, pp. 7778–7789, [34] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, and
Aug. 2021. H. Wymeersch, “Position and orientation estimation through millimeter-
[10] Y. Han, Y. Shen, X.-P. Zhang, M. Z. Win, and H. Meng, “Performance wave MIMO in 5G systems,” IEEE Trans. Wireless Commun., vol. 17,
limits and geometric properties of array localization,” IEEE Trans. Inf. no. 3, pp. 1822–1835, Mar. 2018.
Theory, vol. 62, no. 2, pp. 1054–1075, Feb. 2016. [35] E. Leitinger, F. Meyer, F. Hlawatsch, K. Witrisal, F. Tufvesson,
[11] M. Compagnoni et al., “A geometrical–statistical approach to outlier and M. Z. Win, “A belief propagation algorithm for multipath-
removal for TDOA measurements,” IEEE Trans. Signal Process., vol. 65, based SLAM,” IEEE Trans. Wireless Commun., vol. 18, no. 12,
no. 15, pp. 3960–3975, Aug. 2017. pp. 5613–5629, Dec. 2019.
[12] Z. Li, Z. Tian, M. Zhou, Z. Zhang, and Y. Jin, “Awareness of line-of-sight [36] R. Mendrzik, F. Meyer, G. Bauch, and M. Z. Win, “Enabling situational
propagation for indoor localization using Hopkins statistic,” Sensors, awareness in millimeter wave massive MIMO systems,” IEEE J. Sel.
vol. 18, no. 9, pp. 3864–3874, May 2018. Topics Signal Process., vol. 13, no. 5, pp. 1196–1211, Sep. 2019.
[13] I. Podkurkov, G. Seidl, L. Khamidullina, A. Nadeev, and M. Haardt, [37] E. Leitinger, S. Grebien, and K. Witrisal, “Multipath-based SLAM
“Tensor-based near-field localization using massive antenna exploiting AoA and amplitude information,” in Proc. IEEE Int. Conf.
arrays,” IEEE Trans. Signal Process., vol. 69, pp. 5830–5845, 2021. Commun. Workshops, 2019, pp. 1–7.
[14] D. Feng et al., “An adaptive IMU/UWB fusion method for NLOS indoor [38] H. Kim, K. Granström, L. Gao, G. Battistelli, S. Kim, and
positioning and navigation,” IEEE Internet Things J., vol. 10, no. 13, H. Wymeersch, “5G mmWave cooperative positioning and mapping
pp. 11414–11428, Jul. 2023. using multi-model PHD filter and map fusion,” IEEE Trans. Wireless
[15] R. K. Martin, A. S. King, J. R. Pennington, R. W. Thomas, R. Lenahan, Commun., vol. 19, no. 6, pp. 3782–3795, Jun. 2020.
and C. Lawyer, “Modeling and mitigating noise and nuisance parameters [39] H. Zhang and S. Y. Tan, “TOA based indoor localization and tracking
in received signal strength positioning,” IEEE Trans. Signal Process., via single-cluster PHD filtering,” in Proc. IEEE Global Commun. Conf.,
vol. 60, no. 10, pp. 5451–5463, Oct. 2012. 2017, pp. 1–6.
11286 IEEE INTERNET OF THINGS JOURNAL, VOL. 11, NO. 7, 1 APRIL 2024

[40] L. Lian, A. Liu, and V. K. N. Lau, “User location tracking in massive Ao Peng (Member, IEEE) received the M.Sc.
MIMO systems via dynamic variational Bayesian inference,” IEEE degree and the Ph.D. degree in communication
Trans. Signal Process., vol. 67, no. 21, pp. 5628–5642, Nov. 2019. and information system from Xiamen University,
[41] R. Mendrzik, H. Wymeersch, and G. Bauch, “Joint localization and Xiamen, Fujian, China, in 2011 and 2014,
mapping through millimeter wave MIMO in 5G systems,” in Proc. IEEE respectively.
Global Commun. Conf., 2018, pp. 1–6. In 2015, he joined the School of Informatics,
[42] W. Zhang, J. Zhang, M. Bao, X.-P. Zhang, and X. Li, “Multitarget Xiamen University, where he is currently an
tracking based on dynamic Bayesian network with reparameterized Assistant Professor. His research interests include
approximate variational inference,” IEEE Internet Things J., vol. 9, satellite navigation, indoor navigation, and multi-
no. 13, pp. 11542–11559, Jul. 2022. source positioning and navigation.
[43] Y. Tian, M. Liu, S. Zhang, R. Zheng, and Z. Fan, “Feature-aided passive
tracking of noncooperative multiple targets based on the underwater sen-
sor networks,” IEEE Internet Things J., vol. 10, no. 5, pp. 4579–4591,
Mar. 2023. Xuemin Hong (Member, IEEE) received the Ph.D.
[44] X. Li, E. Leitinger, A. Venus, and F. Tufvesson, “Sequential detection degree from Heriot–Watt University, Edinburgh,
and estimation of multipath channel parameters using belief propaga- U.K., in 2008.
tion,” IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8385–8402, He is currently a Professor with the School of
Oct. 2022. Informatics, Xiamen University, Xiamen, China. He
[45] X. Li, E. Leitinger, M. Oskarsson, K. Åström, and F. Tufvesson, has published one book chapter and over 60 articles
“Massive MIMO-based localization and mapping exploiting phase in refereed journals and conference proceedings. His
information of multipath components,” IEEE Trans. Wireless Commun., research interests include wireless communication
vol. 18, no. 9, pp. 4254–4267, Sep. 2019. networks and localization.
[46] X. Li, K. Batstone, K. Åstrom, M. Oskarsson, C. Gustafson, and
F. Tufvesson, “Robust phase-based positioning using massive MIMO
with limited bandwidth,” in Proc. IEEE 28th Annu. Int. Symp. Pers.,
Indoor, Mobile Radio Commun. (PIMRC), 2017, pp. 1–7. Yixiong Zhang (Member, IEEE) was born in Fujian,
[47] C. Gentner, M. Ulmschneider, R. Karásek, and A. Dammann, China, in 1981. He received the B.S. degree in
“Simultaneous localization of a receiver and mapping of multipath information engineering and the Ph.D. degree in
generating geometry in indoor environments,” in Proc. IEEE Radar information and communication engineering from
Conf., 2021, pp. 1–6. Zhejiang University, Hangzhou, China, in 2003 and
[48] Physical Channels and Modulation (Release 16), 3GPP Standard TS 2009, respectively.
38.211, V16.1.0, 2020. In 2009, he joined the School of Information
[49] B. Sahbani and W. Adiprawita, “Kalman filter and iterative-Hungarian Science and Engineering, Xiamen University,
algorithm implementation for low complexity point tracking as part of Xiamen, China, where he is currently an Associate
fast multiple object tracking system,” in Proc. Int. Conf. Syst. Eng. Professor. His current research interests include
Technol., 2016, pp. 109–115. signal detection and parameter estimation, radar
[50] A. I. Mourikis and S. I. Roumeliotis, “A multi-state constraint Kalman imaging, and hardware/software co-design of embedded systems.
filter for vision-aided inertial navigation,” in Proc. IEEE Int. Conf.
Robot. Autom., 2007, pp. 3565–3572.
[51] Y. Tian, A. Peng, X. Xu, and W. Zhang, “A heading estimation Xiao-Ping Zhang (Fellow, IEEE) received the B.S.
algorithm for wrist device assisted by sequential geomagnetic observa- and Ph.D. degrees in electronic engineering from
tions,” Sensors, vol. 22, no. 6, pp. 5309–5317, Mar. 2022. Tsinghua University, Beijing, China, in 1992 and
[52] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex Optimization. 1996, respectively, and the M.B.A. degree (with
Cambridge, U.K.: Cambridge Univ. Press, 2004. Hons.) in finance, economics, and entrepreneurship
[53] T. L. Hansen, B. H. Fleury, and B. D. Rao, “Superfast line spectral esti- from the University of Chicago Booth School of
mation,” IEEE Trans. Signal Process., vol. 66, no. 10, pp. 2511–2526, Business, Chicago, IL, USA, in 2008.
May 2018. He is a Chair Professor with Tsinghua–Berkeley
[54] D. Shutin, W. Wang, and T. Jost, “Incremental sparse Bayesian learning Shenzhen Institute, Shenzhen, China. He has also
for parameter estimation of superimposed signals,” in Proc. 10th Int. been with the Department of Electrical, Computer
Conf. Sampling Theory Appl., 2013, pp. 6–9. and Biomedical Engineering, Toronto Metropolitan
[55] B. H. Fleury, M. Tschudin, R. Heddergott, D. Dahlhaus, and University (formerly, Ryerson University), Toronto, ON, Canada, as a
K. I. Pedersen, “Channel parameter estimation in mobile radio envi- Professor and the Director of the Communication and Signal Processing
ronments using the SAGE algorithm,” IEEE J. Sel. Areas Commun., Applications Laboratory, and has served as the Program Director of Graduate
vol. 17, no. 3, pp. 434–450, Mar. 1999. Studies. He is cross-appointed to the Finance Department with the Ted
[56] W. Xu, M. Huang, C. Zhu, and A. Dammann, “Maximum likelihood Rogers School of Management, Toronto Metropolitan University. His research
TOA and OTDOA estimation with first arriving path detection for 3GPP interests include image and multimedia content analysis, sensor networks and
LTE system,” Trans. Emerg. Telecommun. Technol., vol. 27, no. 3, IoT, machine learning, statistical signal processing, and applications in big
pp. 339–356, 2016. data, finance, and marketing.
Dr. Zhang is the General Co-Chair for the IEEE International Conference
on Acoustics, Speech, and Signal Processing 2021, the GlobalSIP Symposium
on Signal and Information Processing for Finance and Business 2017,
and the GlobalSIP Symposium on Signal, Information Processing and AI
for Finance and Business 2019. He is the Editor-in-Chief of the IEEE
J OURNAL OF S ELECTED T OPICS IN S IGNAL P ROCESSING. He is a Senior
Area Editor of the IEEE T RANSACTIONS ON I MAGE P ROCESSING. He
served as a Senior Area Editor for the IEEE T RANSACTIONS ON S IGNAL
P ROCESSING and an Associate Editor for the IEEE T RANSACTIONS
Xueting Xu received the B.S. degree from Nanjing
ON I MAGE P ROCESSING , the IEEE T RANSACTIONS ON M ULTIMEDIA ,
University of Science and Technology, Nanjing,
the IEEE T RANSACTIONS ON C IRCUITS AND S YSTEMS FOR V IDEO
China. She is currently pursuing the Ph.D. degree
T ECHNOLOGY, the IEEE T RANSACTIONS ON S IGNAL P ROCESSING, and
in communication and information system with
the IEEE S IGNAL P ROCESSING L ETTERS. He was selected as an IEEE
Xiamen University, Xiamen, China.
Distinguished Lecturer by the IEEE Signal Processing Society and by
Her research interests include wireless communi-
the IEEE Circuits and Systems Society. He is a Fellow of the Canadian
cations, multipath-assisted positioning, and optimal
Academy of Engineering and the Engineering Institute of Canada, a registered
state estimation.
Professional Engineer in Ontario, Canada, and a member of Beta Gamma
Sigma Honor Society. He was an Elected Member of the ICME Steering
Committee.