144 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO.

2, JUNE 2017

Accurate Localization and Tracking of a Passive

RFID Reader Based on RSSI Measurements
Saurav Subedi, Eric Pauls, and Yimin D. Zhang , Senior Member, IEEE

Abstract—Real-time locating systems require high localization information is usually obtained using a directional antenna
accuracy in the order of a few centimeters. Conventional meth- or an antenna array [10], whereas the range informa-
ods for radio frequency identification (RFID) localization fail to tion can be obtained from the received signal strength
achieve such accuracy, particularly in complex radio frequency
propagation environments. In this paper, we propose a method indication (RSSI) [11], round trip-of-flight (TOF) [12], time-
for locating and tracking an RFID reader that can achieve such difference-of-arrival (TDOA) [13], and/or phase-difference-of-
accuracy in a complex propagation environment by exploiting arrival (PDOA) measurements [14]. Second, the location of
received signal strength indicator (RSSI) measurements as the the object is determined by utilizing such information in rela-
only form of observation obtained from multiple spatially dis- tion to multiple known reference points. These positioning
tributed passive tags. There are three key contributions of this
paper. First, we analyze the effect of propagation impairments, methods include multilateration, multiangulation, and hybrid
non-isotropic radiation pattern of the tag antennas and multipath direction/range methods [15]–[18]. Other approaches for the
propagation, on RSSI measurements and the overall localiza- localization of RFID-enabled objects include proximity and
tion and tracking performance. Next, we compensate for the radio map matching. A survey of the state-of-the-art RFID
artifacts of multipath propagation and non-isotropic antenna pat- localization techniques is presented in [2] and [19]. Selection
tern and obtain a maximum likelihood (ML) estimate of the
RFID reader in a 2-D Cartesian space. The ML estimates of of an appropriate localization algorithm is determined based
the reader position, together with its velocity, are then used as on the resource constraints (e.g., bandwidth and power lim-
inputs to the Kalman filter for dynamical estimation of its tra- itations), propagation conditions (e.g., indoor or outdoor),
jectory. Finally, we present experimental results to demonstrate choice of RFID tags (e.g., active or passive), and other system
that the proposed method substantially improves the localization limitations [20].
accuracy compared to other state-of-the-art methods for a given
tag density. In RFID systems, localization of tags is a more customary
problem (e.g., supply chain and transportation). In such appli-
Index Terms—Passive RFID, localization, tracking, multilater- cations, there are a large number of objects to be detected and
ation, RSSI, multipath propagation, antenna pattern.
localized. As such, each object is identified with a tag and
multiple spatially distributed readers at known locations serv-
I. I NTRODUCTION ing as reference points. On the other hand, when the number
ADIO frequency identification (RFID) is an important of objects of interest is limited as observed in, e.g., mobile
R technology for a diverse range of applications involv-
ing electronic identification, localization, and tracking [1]–[3].
robot localization [21] and path planning [22], the objects are
identified using RFID readers, whereas the tags serve as the
With the advent of real-time locating systems applications, reference points. RFID reader localization is a dual problem
research efforts have concentrated on improving the accu- commonly deployed for such applications. In this paper, we
racy of location estimation algorithms [4]–[7]. As such, in focus on the localization of an RFID reader by exploit-
the context of several current and emerging applications, such ing the RSSI measurements obtained from multiple spatially
as telemedicine [8] and gesture recognition [9], accuracy of distributed passive tags as the only form of observation.
localization has emerged as a key challenging issue in RFID In complex propagation conditions due to, e.g., multipath
localization and tracking systems. propagation or non-isotropic radiation of the reader and tag
Typically, localization of an RFID-enabled object is antennas, the one-to-one correspondence cannot be guar-
implemented as a two-step process [2]. First, the range anteed between the observed signal and the range infor-
and/or direction-of-arrival (DOA) information of the object mation, rendering compromised localization accuracy [20].
is estimated from the observed RFID signal. The DOA Different approaches have been proposed to mitigate the
effect of multipath propagation on range estimation and,
Manuscript received March 17, 2017; revised July 30, 2017; accepted
October 19, 2017. Date of publication October 23, 2017; date of cur- consequently, on RFID localization. For example, frequency
rent version December 28, 2017. This work was supported by the Office hopping method [23], averaging over multiple frequency pairs
of Naval Research under Grant N00014-16-1-2425. (Corresponding author: in phase-based methods, and ultra-wideband signaling [2] are
Yimin D. Zhang.)
S. Subedi and E. Pauls are with the Center for Advanced Communications, deployed for improved localization. In time-invariant propa-
Villanova University, Villanova, PA 19085 USA. gation environments, localization algorithms based on radio
Y. D. Zhang is with the Department of Electrical and Computer map matching or scene analysis, such as LANDMARC [24]
Engineering, Temple University, Philadelphia, PA 19122 USA (e-mail:
[email protected]). and its variants, are known to be less sensitive to such prop-
Digital Object Identifier 10.1109/JRFID.2017.2765618 agation impairments as these factors are incorporated into the
2469-7281 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
SUBEDI et al.: ACCURATE LOCALIZATION AND TRACKING OF PASSIVE RFID READER BASED ON RSSI MEASUREMENTS 145

pre-stored radio map. Some recent works in through-the-wall

radar have coined ‘multipath exploitation’ (e.g., through-
the-wall target localization with a single sensor [25] and
through-the-wall radar imaging based on group sparse recon-
struction [26]), by using the prior information regarding the
multipath geometry to their advantage. Likewise, the com-
pensation for the non-isotropic radiation pattern is proposed
in [27], before the measurements are exploited for range and/or
DOA estimation. The need for compensating for the effects of
multipath propagation and non-isotropic radiation field pattern
is also established in [20] and [28], and the a priori knowledge
about the propagation environment is exploited to compensate
for their effects.
Fig. 1. Tag distribution and reader trajectory.
Recently, odometry and vehicle information are employed
as a supplemental source of information to further improve the
accuracy of localization [29]–[32]. For instance, the motion- Hadamard product. Rn×1 and Cn×1 , respectively, represent the
continuity property of differential-driving mobile robots is n-dimensional real and complex vectors.  · n denotes the ln -
utilized in [29]. In [30] and [31], the orientation of an RFID norm of a vector, and x ∼ N (a, b) denotes variable x to be
enabled vehicle is employed as an additional prior. Information Gaussian distributed with mean a and variance b.
from two heterogeneous sensors, namely, RFID sensors and The remainder of this paper is organized as follows.
wheel encoders, is fused in [32] to improve the accuracy of Section II describes the RFID reader-tag grid configuration
localization. and formulate the RSSI expressions with the consideration of
In this paper, we focus on the localization of a single non-isotropic radiation pattern of tag antennas, multipath prop-
RFID reader in a two-dimensional (2-D) space by exploiting agation, and measurement errors. Section III briefly introduces
the RSSI measurements obtained from multiple spatially dis- the conventional RSSI-based passive RFID reader localiza-
tributed passive tags. The key contributions of this work are tion technique through multilateration. Section IV models the
summarized in the following; problem as an ML estimation problem and proposes a method
1) We provide a comprehensive analysis of the individual to accurately estimate the location of a reader in a 2-D space by
and combined effects of multipath propagation and non- compensating for the effects of the propagation impairments.
isotropic radiation pattern of the reader antenna on the Section V presents the use of Kalman filter for dynamically
localization performance. estimating the trajectory of the reader by exploiting the instan-
2) We apply the maximum likelihood (ML) estimation taneous position estimates as the input. Section VI provides the
based method to obtain an optimum localization at a low experimental results and a comparison of the estimation accu-
computational cost. The ML estimation based method racy achieved by exploiting the proposed method against the
offers two distinct advantages compared to traditional state-of-the-art for similar applications. Finally, conclusions
approaches: (a) It does not require an intermediate step are drawn in Section VII.
of range estimation, which is the main source of error in
the traditional method; and (b) It accurately models the
propagation environment and effectively compensates II. S IGNAL M ODEL
for the effects of the propagation impairments. We consider a passive RFID system comprising K passive
3) Subsequently, the position estimates obtained in the form RFID tags distributed in a predefined arrangement over a rect-
of 2-D Cartesian coordinates of the reader position and angular search area and a reader as shown in Fig. 1. For the
its velocity are then used as inputs to a Kalman filter underlying problem of reader localization, tags are assumed
for dynamically estimating the trajectory of the reader to be located at known positions and serve as the reference
within the surveillance region. points. The position of the kth tag in a 3-D space is defined
4) We deploy the proposed algorithm in a realistic propaga- as tk  [txk , tyk , tzk ]T . At each tag position tk , we use an
tion environment. Experimental results corroborate that, orthogonal pair of tags to form a ‘super-tag’ arrangement. The
due to the ability to compensate for the channel impair- concept of super-tag is introduced in [34] where multiple tags
ments and incorporate supplemental vehicle information are deployed at the same location but with different orienta-
at a low computational cost, the proposed method signif- tions. Such orientation diversity ensures that at least one tag
icantly improves the accuracy, by an order of magnitude, is in an orientation of a high directive gain, thereby achieving
compared to the state-of-the-art techniques for similar a high likelihood of detecting tags for each tag location. Our
applications [29]–[33] under a comparable density of earlier empirical study [28] also implemented super-tags in the
tag deployment. form of a pair of orthogonal tags.
Notations: A lower (upper) case bold letter denotes a vector The reader is assumed to be moving along a trajectory
(matrix). Specifically, IN and 0N denote the N × N identity within the rectangular search area with a velocity vτ . The
and zero matrices, respectively. (.)T and (.)H , respectively, reader position at the τ th observation instant is defined as
denote transpose and hermitian operations, and ◦ denotes the qτ  [qx,τ , qy,τ , qz,τ ]T . The tags are assumed to be located
146 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO. 2, JUNE 2017

on the flat floor whereas the reader is placed at a known

and fixed height of h above the ground. As such, the reader
localization is reduced to a 2-D problem. We consider the
problem of tracking the reader over T observation instants,
i.e., τ = 1, . . . , T. At each observation instant, we discard the
tags that do not respond or report RSSI values below a certain
threshold. As such, for the τ th observation, the actual number
of tags contributing towards localization, Lτ , is much smaller
than K.
Since passive RFID tags are not equipped with built-in
energy sources, they rely on the radio frequency power trans-
mitted from the reader to activate their internal circuitry and
respond back to the reader. As such, the signal transmitted Fig. 2. Two-ray reflection model.
from a reader propagates twice along the propagation path.
According to the Friis transmission equation, for the τ th
observation instant, the signal power received at the reader ambiguity in range estimates calculated on the basis of RSSI
due to the signal backscattered from the kth tag is expressed measurements.
as [35] In the following, we separately analyze the effects of
 2η multipath propagation and directional pattern of the reader
λ antenna on range estimates, and consequently, on reader
pk,τ = pTx Gt,k,τ Gr,k,τ ζ
2 2
, (1)
4π dk,τ localization performance.

where pTx is the transmit power from the reader, Gt,k,τ and
Gr,k,τ are the directional gains of the kth tag, and reader A. Effect of Multipath Propagation
antenna, respectively, in the direction of propagation, ζ is the Multipath propagation, due to the presence of different
backscatter transmission efficiency of the passive tag, λ is the reflecting and scattering objects in the propagation scene, can
wavelength of the transmitted signal, and dk,τ is the range be a major source of error in RFID reader localization. In many
between the kth tag and the reader during the τ th time instant. practical applications, the geometry of the significant reflec-
The tag efficiency, ζ , is generally considered constant which tors is either known or can be pre-estimated. For example,
typically takes a value of −5 dB [35] but can be adjusted in a recent work [36], a low-complexity clustering algorithm
empirically based on the measured data [28]. The path loss is developed to detect multiple reflectors and estimate their
exponent, η, varies depending upon the propagation environ- geometry using time delay measurements. We consider a case
ment. Considering that the reader has line-of-sight (LOS) with where the reflector is a flat vertical surface with a known posi-
the tags, free space propagation is assumed and, as such, η is tion. Such situations arise, for example, when the reader is
assumed to be 2. However, other values empirically estimated mounted on a vehicle in a warehouse with a reflective sur-
for a given propagation environment can also be used. As such, face, whereas other reflectors are distant from the vehicle. In
for LOS propagation and known antenna gains of the reader such a situation, signal transmitted from a reader follows two
and the tag antennas in the direction of propagation, the range distinct paths: a direct path and a reflection path, as illus-
between the reader and the kth tag at the τ th observation can trated in Fig. 2. The analysis presented here can be easily
be estimated from the RSSI measurements as follows, extended to a multiple reflector scenario. Signals can also
propagate through multiple reflections but they are severely
  1
λ pTx G2t,k,τ G2r,k,τ 4 attenuated and, hence, become negligible in the RSSI mea-
d̆k,τ = . (2) surement. Along the direct path, the propagation distance
4π pk,τ
between the reader and the kth tag at the τ th observation
In practice, however, the actual directional gains of the equals dk,τ = ||qτ − tk ||. Similarly, along the reflection path,
reader and the tag antennas towards the respective propaga- the propagation distance equals d̃k,τ = ||q̃τ − tk ||, where q̃τ
tion directions depend on their relative positions and, hence, is the mirror image of the reader position about the reflecting
are unknown a priori. surface (see Fig. 2). Therefore, the path difference between
Since the RFID systems are typically operated over a short the direct path and the reflection path is
distance, the RSSI measurements are highly coupled with the
reader and tag antenna patterns [27]. Therefore, the range dk,τ = d̃k,τ − dk,τ , (3)
estimates may become ambiguous in the presence of non-
isotropic radiation patterns. Multipath propagation is another and the corresponding phase difference is
major source of error in RSSI based localization. The signals 2π
propagating along different paths are superposed at the RFID αk,τ = dk,τ . (4)
λ
reader, constructively or destructively, depending on their rela-
tive phases. This causes significant variations in signal strength Thus, following the ground two-ray model [37], the signal
even within a fraction of a wavelength, which adds to the power backscattered to the reader from the kth tag can be
SUBEDI et al.: ACCURATE LOCALIZATION AND TRACKING OF PASSIVE RFID READER BASED ON RSSI MEASUREMENTS 147

expressed as RFID reader is expressed as

 4  4 gk,τ = pk,τ + ek,τ ,
exp(−jαk,τ ) 
(7)
λ  1
pk,τ = pTx  +  , (5)
4π  dk,τ d̃k,τ  where pk,τ is the noise free power of the backscatter signal
from the kth tag, and ek,τ represents a RSSI measurement
where is the reflection coefficient of the wall. The phase error modeled as an additive white Gaussian random variable
interaction between two multipath components creates ambi- ek,τ ∼ N (0, σek,τ
2 ).
guity in relation between the RSSI value observed at the In the following section, we review the conventional method
reader and distance between the corresponding reader-tag pair. of RSSI-based RFID reader localization through multilat-
Therefore, in order to unambiguously estimate the propaga- eration and evaluate its performance in the presence of
tion range by using the RSSI information available at the propagation impairments.
reader from multiple reference tag positions, it is impor-
tant to compensate for the effects of all significant multipath III. RSSI-BASED L OCALIZATION
components. VIA M ULTILATERATION
Following (2), at the τ th observation, the range between
B. Effect of Tag Antenna Directivity the reader and the kth tag can be estimated from the RSSI
Conventionally, RSSI-based localization algorithms assume measurements as follows,
isotropic radiation patterns for the reader and the tag anten-   1
λ pTx 4
nas [38]. However, in practice, the effective radiation pattern d̂k,τ = , (8)
is non-isotropic creating a bias towards the direction of high 4π gk,τ
gain in the radiation pattern. In this paper, we only consider where the RSSI measurement from the kth tag gk,τ is defined
the effect of tag antenna pattern because the antenna gain does in (7). In order to unambiguously localize a reader in an
not change significantly within the main lobe of the radiation n-dimensional space using multilateration, range information
pattern of the reader antenna. Also, the tags located away from from at least n+1 tags is required. For the underlying problem,
the main lobe of the reader antenna do not report RSSI mea- at the τ th observation, we consider estimating the position of
surements above the preset threshold and are not accounted a single RFID reader in a 2-D space, such that
for localization. For notational purpose, this point onwards,
2
2
qx,τ − txk + qy,τ − tyk + h2 = d̂k,τ 2
, k = 1, 2, . . . , Lτ ,
we use Gk,τ to represent the directional gain of the kth tag
antenna towards the direction of the reader antenna at the τ th (9)
observation instant. where Lτ ≥ 3 and d̂k,τ is the propagation range estimated
The effect of the tag antenna pattern is negligible in the using (8). Availability of more number of tags results in an
far-field. However, because RFID systems are often operated over-determined problem, which helps improving the accuracy
in the near field, the tag antenna gain depends on the rela- of estimation.
tive position between the reader and the tag and, as such, the After simple algebraic manipulations, equation (9) can be
spatial directivity has a pronounced effect on the RSSI mea- used to formulate the following linear relationship [2],
surements [27]. Therefore, it is required to compensate for
the effect of directional radiation pattern on RSSI measure- Aτ qτ = bτ , (10)
ments when estimating the range. Incorporating the effects of where qτ = [qx,τ , qy,τ ]T ,
tag antenna directivity in (5), the power backscattered to the ⎛ ⎞
reader from the kth tag can be modeled as tx1 − tx2 ty1 − ty2
⎜ .. .. ⎟
  4 Aτ = ⎝ . . ⎠ (11)
 4   G̃k,τ exp(−jαk,τ ) 

λ  Gk,τ tx(Lτ −1) − txLτ ty(Lτ −1) − tyLτ
pk,τ = pTx  +  , (6)
4π  dk,τ d̃k,τ

 is a (Lτ −1)×2 matrix, bτ  [b1 , . . . , bLτ −1 ]T is a (Lτ −1)×1
vector, and
1  2     
where G̃k,τ represents the tag antenna gain along the reflection
path. bk,τ = d̂k+1,τ − d̂k,τ
2
− tx(k+1)
2
− txk
2
− ty(k+1)
2
− tyk
2
,
2
(12)
C. Effect of RSSI Measurement Error
k = 1, . . . , Lτ − 1. Assuming that Lτ ≥ 3 and the tags are
From a practical standpoint, statistic and systematic errors not collinearly located, the least square (LS) estimate of the x-
that occur during the RSSI measurement introduce another and y-coordinates of the RFID reader position can be obtained
significant source of error in RSSI-based range estimation. In as [2]
most of the RFID localization applications, a fairly high sig-
−1 T
nal strength is maintained, and thus, these RSSI measurement q̂(LS)
τ = ATτ Aτ Aτ bτ , (13)
errors are more significant than the thermal noise. Considering
by minimizing the LS cost function
the effects of multipath propagation, tag antenna directivity,
and the RSSI measurement error, the received signal at the J(qτ ) = (bτ − Aτ qτ )T (bτ − Aτ qτ ). (14)
148 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO. 2, JUNE 2017

As such, the estimated reader position is corresponding phase difference. Equation (19) represents an
 T accurate model of the propagation environment, by exploiting
q̂(LS)
τ = q̂τ(LS)T , h . (15) the a priori knowledge of the radiation pattern of the reader
antenna and multipath due to the presence of the wall in the
As discussed in Section I, due to multipath propagation, tag
propagation scene. As such, we can reformulate the underlying
antenna directivity, and RSSI measurement error, the relation
estimation problem in (18) as
between the reader-tag distance and the corresponding RSSI  2
measurement is not consistent. As such, the accuracy of the q̂(ML)
τ = arg mingτ − ψ i (qi ) , (20)
qi
LS estimate of the reader position is adversely affected due
to these propagation impairments. Furthermore, it is not pos- where the Lτ × 1 vector ψ i  [p1,i , . . . , pk,i , . . . , pLτ ,i ]T .
sible to compensate for the effects of such impairments using In the following section, we feed the ML estimate of
the multilateration based method. In the following section, we the instantaneous reader position obtained from (20) to the
propose an ML estimation method, which does not involve an Kalman filter for estimating the trajectory of the reader.
intermediate range estimation process and allows for an effec-
tive compensation of the effect of the propagation impairments V. K ALMAN F ILTERING
to obtain an optimal solution.
For tracking the trajectory of the reader, we resort to the
Kalman filter. We define the state vector of the reader at the τ th
IV. P RECISE L OCALIZATION U SING ML E STIMATION
observation as a four-dimensional (4-D) vector xτ = [qTτ , vTτ ]T
From (7), we can infer that the Lτ × 1 received signal comprising its position qτ  [qx,τ , qy,τ ]T and velocity vτ 
vector gτ  [g1 , . . . , gLτ ]T follows a multivariate normal [vx,τ , vy,τ ]T in the two-dimensional (2-D) Cartesian coordinate
distribution, i.e., system. The reader dynamics is assumed to evolve according
gτ ∼ N (pτ , Cτ ), (16) to a constant velocity linear Gaussian model, such that

where pτ  [p1 , . . . , pLτ ]T is the Lτ × 1 mean vector, and Cτ xτ = Fxτ −1 + wτ −1 , (21)

is the corresponding Lτ × Lτ covariance matrix. The 2 × 1 where F is the state transition matrix defining the linear
unknown vector qτ = [qx,τ , qy,τ ]T structures the mean value dynamics, expressed as
vector pτ (qτ ) according to a known function defined in (6).  
The covariance matrix, on the other hand, represents the effect I2 I2
F= , (22)
of the RSSI measurement error modeled as an additive white 02 I2
Gaussian variable.  is the sampling interval, and wτ ∼ N (0, Q) is the process
The log-likelihood function of the unknown vector qτ noise modeled as additive white Gaussian. The process noise
structuring the received signal vector gτ modeled as (16) is covariance matrix can be expressed as
given as  3 
 2
I 2 I 2
1 Q = σw 32
2 2 , (23)
L(qτ |gτ ) = − ln(det(Cτ )) I2
2 2 I2
1
− (gτ − pτ (qτ ))T C−1
τ (gτ − pτ (qτ )), (17) where σw2 is the variance of the process noise.
2 For each observation instant, τ , we feed the ML esti-
and the corresponding ML estimate qˆτ (ML) is given as mates of the instantaneous reader position as a measurement
to the Kalman filter. We define the measurement vector as
q̂(ML)
τ = arg max L(qτ |gτ ), (18) zτ = [q̂τ
(ML)T T T (ML)
, vτ ] , where q̂τ is the ML estimate of the
qτ ∈
instantaneous reader position obtained from (20). Since this
where  represents all the points within the search area.
is a self-localization problem, we assume the prior knowl-
For a hypothetical reader position qi = [qxi , qyi ]T within the
edge of the instantaneous velocity. As such, we can define the
2-D search area, the error-free power measurement from the
measurement vector as
kth tag located at tk is modeled, following (6), as
  4 zτ = Hτ xτ + ητ , (24)
 4   )

λ  G G̃k,i exp(−jα k,i 
pk,i = pTx  k,i
+  , (19) where Hτ = I4 is the matrix mapping the measurement to

4π  dk,i 
d̃k,i  the target state space and ητ ∼ N (0, Rτ ) is the measure-
ment noise modeled as an additive white Gaussian vector.
where G̃k represents the reader antenna gain along the reflec- As such, following the conventional Kalman filter formula-
tion path. where Gk,i and G̃k,i , respectively, represent the tions [39], we obtain the predictions for the state vector and
directional gains of the reader antenna, hypothetically located the corresponding covariance matrix, respectively, as
at qi = [qTi , h]T , towards the tag located at tk along the direct
path and the reflection path. Likewise, dk,i = ||qi − tk || and x̂τ |τ −1 = Fx̂τ −1|τ −1 , (25)
d̃k,i = ||q̃i − tk ||, respectively, represent the distance between and
the hypothetical reader position and the kth tag along the direct
path and the reflection path, and αk,i = 2π λ (d̃k,i − dk,i ), is the Pτ |τ −1 = Qτ −1 + FPτ −1|τ −1 FT . (26)
SUBEDI et al.: ACCURATE LOCALIZATION AND TRACKING OF PASSIVE RFID READER BASED ON RSSI MEASUREMENTS 149

Fig. 3. Structure hosting the RFID reader.

Fig. 5. RSSI map illustrating the effect of tag antenna orientation.

consistently similar RSSI measurements in a controlled test

environment [28] are used for the experiment. The wooden
structure hosting the RFID reader antenna moves along a
trajectory comprising 46 points within a designated surveil-
lance area where the tags are arranged in a fixed pattern. As
discussed in Section II, for each observation instant τ , we
discard the tags that report RSSI values below −104 dBm.
It was observed empirically that the RSSI values below this
threshold suffer from a high variance, rendering the measure-
ments unreliable for localization. Similar observations were
Fig. 4. Experiment setup. reported in [28]. As such, the actual number of tags con-
tributing towards localization is different for each observation
instant τ . The experiment is conducted in a large patio devoid
Likewise, the updates are obtained as of any significant reflecting objects in the close vicinity of the

experimental setup.
x̂τ |τ = x̂τ |τ −1 + Kτ zτ − Hτ x̂τ |τ −1 , (27)
and A. Effect of Tag Antenna Directivity
Pτ |τ = Pτ |τ −1 − Kτ Sτ KTτ , (28) An example of the RSSI map observed at the reader is
presented in Fig. 5. This figure clearly demonstrates the effect
where Sτ is the covariance of the innovation term sτ =
of the radiation pattern of the tag antenna on the RSSI mea-
zτ −Hτ x̂τ |τ −1 , and Kτ = Pτ |τ −1 HTτ S−1
τ represents the Kalman surements. The squares and diamonds respectively show RSSI
gain.
results corresponding to the two orthogonally oriented tags. It
is evident that the two tags arranged as an orthogonal pair
VI. E XPERIMENTAL R ESULTS
report significantly different RSSI values depending on their
In this section, we present the experimental study results to respective orientation even though they are located exactly
validate the proposed algorithm. For the experiments, we use at the same distance from the reader. Also, the RSSI val-
an Alien ALR-9680 RFID reader with transmit power of 1W ues reported from equidistant tags located towards different
(30 dBm) communicating in the ultra high frequency (UHF) directions from the reader are different. This empirical obser-
frequency band between 902.75 − 927.25 MHz. The reader vation corroborates the analysis in Section II and illustrates the
is connected to a Poynting PatchA0026 right-hand circularly need for compensating for the effect of non-isotropic radiation
polarized antenna with a nominal gain of 6.5 dBi (±0.5 dB) pattern of the tag antenna.
and 3 dB beamwidth of 60◦ (±5◦ ) and 74◦ (±5◦ ) in elevation In order to accurately compensate for the effect of the non-
and azimuth, respectively. Alien Squiggle Higgs3 passive UHF isotropic radiation pattern of the tag antenna, we empirically
RFID tags are used for the experiments. estimate its radiation pattern. For the experiment, we place an
The RFID reader antenna is mounted at a fixed height of RFID tag directly beneath the RFID reader antenna at a fixed
approximately 0.88 m onto a movable wooden structure, as distance of 0.76 m and rotate the tag in the range [ − 90◦ , 90◦ ]
shown in Fig. 3, and the tags are distributed at 57 fixed at an increment of 10◦ as shown in Fig. 6. The angle of rotation
locations on a carpet over a concrete surface at a minimum θ is the angle between the reader-tag line of sight vector a and
tag spacing of 0.762 m in an arrangement comprising con- the tag axis vector b, as shown in Fig. 6(b), and is obtained as
tiguous equilateral triangles, as depicted in Fig. 4. In order  
to account for the non-linear artifacts, including tag effi- a·b
θ = arccos . (29)
ciency, we prescreen tags such that only those that generate a2 b2
150 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO. 2, JUNE 2017

Fig. 6. Experiment for estimating the radiation pattern of the tag antenna
(a) Experimental setup. (b) Illustration of tag orientation angle.

Fig. 9. RSSI map for a fixed reader position. (a) Without metallic wall.
Fig. 7. Radiation pattern of the tag antenna. (b) With metallic wall.

aluminum plate of height 0.91 m between the points (3, 1.694)

m and (5.3, 0.366) m as illustrated in Fig. 8. The wall is
located parallel to the reader trajectory and is approximately
at a vertical distance of 0.55 m from the reader trajectory. We
compare the RSSI map observed at the reader at a fixed posi-
tion with and without the metallic wall in Fig. 9. The presence
of the reflecting surface within the read range of the reader
creates a bias in the RSSI map. For example, the two tags
(circled in red in Fig. 9(b)) located at an equal distance from
the reader, report significantly different RSSI values due to the
strong reflection from the metallic surface.

Fig. 8. Location of the metallic wall.

C. Comparison of Multilateration and the ML
Based Methods
The empirically estimated radiation pattern of the tag First, we compare the estimated reader trajectories using
antenna is shown in Fig. 7, where it is evident that the the multilateration based method and the ML estimation
directional gain is approximately constant over the range based method without any compensation. The results in
of [−40◦ , 40◦ ] and rapidly decreases outside this range. Figs. 10 and 11 show that the multilateration based loca-
Empirical results verify that the radiation pattern of the Alien tion estimation process suffers significantly more than the ML
Squiggle Higgs3 passive UHF RFID tags [40] is toroidal and a estimation based method in the presence of the propagation
higher gain is experienced along the bore-sight direction. Also, impairments, i.e., non-isotropic radiation pattern and multipath
we can observe that the radiation pattern of the tag antenna is propagation. This is primarily due to the intermediate range
largely symmetric about the axis of rotation. estimation process and a subsequent non-linear relation-
ship between the estimated ranges and the reader position
estimate.
B. Effect of Metallic Wall While the ML estimation based method has a lower error
In order to investigate the effect of a reflecting surface inside compared to the multilateration based approach even with-
the region of interest, a reflecting surface in the form of an out compensation as illustrated in Fig. 11. However, there
SUBEDI et al.: ACCURATE LOCALIZATION AND TRACKING OF PASSIVE RFID READER BASED ON RSSI MEASUREMENTS 151

Fig. 10. Estimated reader trajectories without a metallic wall.

(a) Multilateration based method. (b) ML estimation based method without
compensation. Fig. 12. Estimated reader trajectories using the ML method after compen-
sation. (a) Without metallic wall. (b) With metallic wall.

Fig. 13. Consistency of location estimation.

properly modeling and compensating for the effects multipath

propagation.
Next, we compare the reliability of location estimation for
each point in the reader trajectory using the conventional
multilateration based method and the proposed ML estima-
tion based method in Fig. 13. Experimental results show that
the proposed method achieves a significantly lower variance.
As such, it not only improves the accuracy but also notably
increases the consistency of location estimation by compen-
Fig. 11. Estimated reader trajectories with the metallic wall. sating for the effects of non-isotropic radiation pattern and
(a) Multilateration based method. (b) ML estimation based method without multipath propagation.
compensation.

is clearly a bias in location estimates towards the metallic D. Reader Position Tracking Using Kalman Filter
wall because the RSSI measurement is sensitive to reflection. As discussed in Section V, we can further improve the
Fig. 12 shows that this error can be significantly reduced by estimation accuracy of the position of the RFID reader by
152 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO. 2, JUNE 2017

TABLE I
AVERAGE RMSE OF L OCALIZATION ( IN m)

Fig. 14. Estimated reader trajectory after applying Kalman filter on ML Fig. 15. Estimated reader trajectory after applying Kalman filter on LS
estimates. (a) Without metallic wall. (b) With metallic wall. estimates. (a) Without metallic wall. (b) With metallic wall.

multilateration based method is its inability to correct for

exploiting the dynamic model assumed to be known a priori.
the artifacts of multipath propagation and antenna directivity.
The estimated trajectory after Kalman filtering is shown in
These propagation impairments manifest themselves as esti-
Fig. 14. It can be observed that the Kalman filter prevents sud-
mation bias and larger estimation variance, which are more
den changes in the estimated trajectory by suppressing the ML
difficult to mitigate by exploiting the Kalman filter as shown
position estimates that do not conform to the dynamic model.
in Fig. 15.
As such, the overall accuracy can be significantly improved at
a low computational cost by exploiting the Kalman filter.
Specifically, in this experiment, the average root mean E. Comparison With State-of-the-Art Methods
square error (RMSE) of the location estimates averaged over In Table II, we compare the accuracy of the proposed
the 46 points along the reader trajectory is summarized in method with other state-of-the-art passive RFID localization
Table I. The RMSE of the location estimates of the conven- methods [29]–[33]. It is noted that, as different methods used
tional multilateration based method is approximately 0.45 m different tag grids with varying levels of tag densities, the
without wall, which degrades further to approximately 0.50 m localization accuracy has to be evaluated in relation to the
in the presence of a metallic wall. The ML estimation based respective density of tag deployment and minimum tag spac-
method reduces the error to as low as 0.1 m due to its abil- ing for fair comparison. We use the normalized error, defined
ity to compensate for propagation impairments. Application of as the product of the localization error and the respective tag
Kalman filtering on the output of ML estimates achieves an density, i.e.,
improved accuracy of less than 0.04 m by exploiting the prior  
q̂τ − qτ 2
knowledge of the reader’s instantaneous velocity. en,τ = 2
, (30)
It is worthwhile to mention that, while the Kalman filter A/K
is a generic framework that can be applied in tandem with as a basis of comparison in Table II, where en,τ is the normal-
the conventional multilateration based approach as well, it ized error, q̂τ and qτ represent the estimated and true positions
is important to note that the primary source of error in the of the reader at the τ th observation instant, A is the area of
SUBEDI et al.: ACCURATE LOCALIZATION AND TRACKING OF PASSIVE RFID READER BASED ON RSSI MEASUREMENTS 153

TABLE II
C OMPARISON OF L OCALIZATION ACCURACY OF THE S TATE - OF - THE -A RT U SING RFID S YSTEMS

surveillance, and K represents the number of tags deployed. A [9] T. Kanda, M. Shiomi, Z. Miyashita, H. Ishiguro, and N. Hagita, “A com-
comparison of the localization performance of these methods munication robot in a shopping mall,” IEEE Trans. Robot., vol. 26, no. 5,
pp. 897–913, Oct. 2010.
in terms of the normalized localization error shows that the [10] Y. Zhang, M. G. Amin, and S. Kaushik, “Localization and tracking
proposed method outperforms other methods for a given tag of passive RFID tags based on direction estimation,” Int. J. Antennas
density. Propag., vol. 2007, Dec. 2007, Art. no. 17426.
[11] G. Marrocco, E. Di Giampaolo, and R. Aliberti, “Estimation of UHF
RFID reading regions in real environments,” IEEE Antennas Propag.
Mag., vol. 51, no. 6, pp. 44–57, Dec. 2009.
VII. C ONCLUSION [12] F. Izquierdo, M. Ciurana, F. Barcelo, J. Paradells, and E. Zola,
“Performance evaluation of a TOA-based trilateration method to locate
In this paper, we have proposed an ML estimation technique terminals in WLAN,” in Proc. Int. Symp. Wireless Pervasive Comput.,
to directly estimate the position of an RFID reader in a 2-D Phuket, Thailand, Jan. 2006, pp. 1–6.
cartesian coordinates without the intermediate range estima- [13] T. Wang, “Novel sensor location scheme using time-of-arrival estimates,”
IET Signal Process., vol. 6, no. 1, pp. 8–13, Feb. 2012.
tion process. The proposed method allows us to compensate
[14] X. Li, Y. Zhang, and M. G. Amin, “Multifrequency-based range estima-
for the artifacts of the multipath propagation and the non- tion of RFID tags,” in Proc. IEEE Int. Conf. RFID, Orlando, FL, USA,
isotropic antenna pattern and obtain an optimal solution in the Apr. 2009, pp. 147–154.
ML sense. We further apply the Kalman filter for dynamically [15] A. H. Sayed, A. Tarighat, and N. Khajehnouri, “Network-based wireless
location: Challenges faced in developing techniques for accurate wire-
estimating the trajectory of the reader within the surveil- less location information,” IEEE Signal Process. Mag., vol. 22, no. 4,
lance region. Experimental results verified that the proposed pp. 24–40, Jul. 2005.
algorithm outperforms the conventional multilateration based [16] A. Milella, P. Vanadia, G. Cicirelli, and A. Distante, “RFID-based envi-
ronment mapping for autonomous mobile robot applications,” in Proc.
localization technique and the estimation accuracy is fur- IEEE/ASME Int. Conf. Adv. Intell. Mechatronics, Zürich, Switzerland,
ther improved at a low computational cost by exploiting the Sep. 2007, pp. 1–6.
Kalman filter. A comparison of the localization performance [17] J. Song, C. T. Haas, and C. H. Caldas, “A proximity-based method
for locating RFID tagged objects,” Adv. Eng. Informat., vol. 21, no. 4,
of these methods in terms of the normalized localization error pp. 367–376, Oct. 2007.
shows that the proposed method achieves significantly higher [18] Y. D. Zhang, X. Li, and M. G. Amin, “Array processing for RFID
localization accuracy compared to other methods for a given tag localization exploiting multi-frequency signals,” in Proc. SPIE,
tag density. Conversely, for a given accuracy, the proposed vol. 7349. Orlando, FL, USA, Apr. 2009, Art. no. 734907.
[19] M. Bouet and A. L. Dos Santos, “RFID tags: Positioning principles
method requires a sparser tag deployment. and localization techniques,” in Proc. IFIP Wireless Days Conf., Dubai,
UAE, Nov. 2008, pp. 1–5.
[20] S. Subedi, Y. D. Zhang, and M. G. Amin, “Precise RFID localization
R EFERENCES in impaired environment through sparse signal recovery,” in Proc. SPIE,
vol. 8753. Baltimore, MD, USA, May 2013, Art. no. 87530H.
[1] K. Finkenzeller, RFID Handbook: Radio-Frequency Identification [21] K. Yamano et al., “Self-localization of mobile robots with RFID system
Fundamentals and Applications. New York, NY, USA: Wiley, 1999. by using support vector machine,” in Proc. IEEE/RSJ Int. Conf. Intell.
[2] Y. Zhang, X. Li, and M. G. Amin, “Principles and techniques of Robots Syst., vol. 4. Sendai, Japan, Sep. 2004, pp. 3756–3761.
RFID positioning,” in RFID Systems Research Trends and Challenges, [22] S. A. Parkison, E. T. Psota, and L. C. Perez, “Automated indoor
M. Bolić, D. Simplot-Ryl, and I. Stojmenović, Eds. New York, NY, RFID inventorying using a self-guided micro-aerial vehicle,” in Proc.
USA: Wiley, 2010, pp. 389–412. IEEE Int. Conf. Electro Inf. Technol., Milwaukee, WI, USA, Jun. 2014,
[3] N. C. Karmakar, Handbook of Smart Antennas for RFID Systems. pp. 335–340.
Hoboken, NJ, USA: Wiley, 2011. [23] I. Amir and A. Naim, “Frequency hopping range estimation with low
[4] C. Hekimian-Williams, B. Grant, X. Liu, Z. Zhang, and P. Kumar, power consumption,” U.S. Patent, 7 061 428 B1, Jun. 2006.
“Accurate localization of RFID tags using phase difference,” in Proc. [24] L. M. Ni, Y. Liu, Y. C. Lau, and A. P. Patil, “LANDMARC: Indoor
IEEE Int. Conf. RFID, Orlando, FL, USA, Apr. 2010, pp. 89–96. location sensing using active RFID,” in Proc. IEEE Int. Conf. Pervasive
[5] J. S. Choi, H. Lee, R. Elmasri, and D. W. Engels, “Localization systems Comput. Commun., Fort Worth, TX, USA, Mar. 2003, pp. 407–415.
using passive UHF RFID,” in Proc. Int. Joint Conf. INC IMS IDC, Seoul, [25] P. Setlur, G. E. Smith, F. Ahmad, and M. G. Amin, “Target localization
South Korea, Aug. 2009, pp. 1727–1732. with a single sensor via multipath exploitation,” IEEE Trans. Aerosp.
[6] D. Fortin-Simard, K. Bouchard, S. Gaboury, B. Bouchard, and Electron. Syst., vol. 48, no. 3, pp. 1996–2014, Jul. 2012.
A. Bouzouane, “Accurate passive RFID localization system for smart [26] Q. Wu, Y. D. Zhang, M. G. Amin, and F. Ahmad, “Through-the-wall
homes,” in Proc. IEEE Int. Conf. Netw. Embedded Syst. Every Appl., radar imaging based on modified Bayesian compressive sensing,” in
Liverpool, U.K., Dec. 2012, pp. 1–8. Proc. IEEE ChinaSIP, Xi’an, China, Jul. 2014, pp. 232–236.
[7] J. Maneesilp, C. Wang, H. Wu, and N.-F. Tzeng, “RFID support [27] J. Wang, M. G. Amin, and Y. D. Zhang, “Signal and array processing
for accurate 3D localization,” IEEE Trans. Comput., vol. 62, no. 7, techniques for RFID readers,” in Proc. SPIE, vol. 6248. Orlando, FL,
pp. 1447–1459, Jul. 2013. USA, Apr. 2006, Art. no. 624807.
[8] S. Chia, A. Zalzala, L. Zalzala, and A. Karim, “Intelligent technologies [28] E. Pauls and Y. D. Zhang, “Experimental studies of high-accuracy
for self-sustaining, RFID-based, rural e-health systems,” IEEE Technol. RFID localization with channel impairments,” in Proc. SPIE, vol. 9497.
Soc. Mag., vol. 32, no. 1, pp. 36–43, Mar. 2013. Baltimore, MD, USA, Apr. 2015, Art. no. 94970I.
154 IEEE JOURNAL OF RADIO FREQUENCY IDENTIFICATION, VOL. 1, NO. 2, JUNE 2017

[29] S. Han, H. Lim, and J. Lee, “An efficient localization scheme for a Eric Pauls received the B.S. and M.S. degrees
differential-driving mobile robot based on RFID system,” IEEE Trans. in electrical engineering from Villanova University,
Ind. Electron., vol. 54, no. 6, pp. 3362–3369, Dec. 2007. Villanova, PA, USA, in 2014 and 2016, respectively.
[30] S. Park and S. Hashimoto, “Autonomous mobile robot navigation using He is currently a member of the Advanced Systems
passive RFID in indoor environment,” IEEE Trans. Ind. Electron., Technologies Group, Harris Corporation. His work is
vol. 56, no. 7, pp. 2366–2373, Jul. 2009. focused on machine learning, geolocation, and radar
[31] S. Park and H. Lee, “Self-recognition of vehicle position using UHF pas- systems.
sive RFID tags,” IEEE Trans. Ind. Electron., vol. 60, no. 1, pp. 226–234,
Jan. 2013.
[32] B.-S. Choi, J.-W. Lee, and J.-J. Lee, “Localization and map-building of
mobile robot based on RFID sensor fusion system,” in Proc. IEEE Int.
Ind. Inform., Daejeon, South Korea, Jul. 2008, pp. 412–417.
[33] H. J. Lee and H. J. Lee, “Localization of mobile robot based on
radio frequency identification devices,” in Proc. SICE-ICASE, Busan,
South Korea , Oct. 2006, pp. 5934–5939.
[34] P. Asadzadeh, L. Kulik, and E. Tanin, “Gesture recognition using RFID
technology,” Pers. Ubiquitous Comput., vol. 16, no. 3, pp. 225–234,
Mar. 2012. Yimin D. Zhang (SM’01) received the Ph.D. degree
[35] D. M. Dobkin, The RF in RFID: Passive UHF RFID in Practice. from the University of Tsukuba, Tsukuba, Japan, in
Burlington, MA, USA: Newnes, 2008. 1988.
[36] H. Naseri and V. Koivunen, “Indoor mapping based on time delay esti- He joined the Department of Radio Engineering,
mation in wireless networks,” in Proc. IEEE ICASSP, Brisbane, QLD, Southeast University, Nanjing, China, in 1988. He
Australia, Apr. 2015, pp. 2424–2428. served as the Director and the Technical Manager
[37] T. S. Rappaport, Wireless Communications: Principles and Practice. with Oriental Science Laboratory, Yokohama, Japan,
Upper Saddle River, NJ, USA: Prentice-Hall, 2002. from 1989 to 1995, a Senior Technical Manager with
[38] Y. Zhao, N. Patwari, P. Agrawal, and M. G. Rabbat, “Directed by direc- the Communication Laboratory Japan, Kawasaki,
tionality: Benefiting from the gain pattern of active RFID badges,” IEEE Japan, from 1995 to 1997, and a Visiting Researcher
Trans. Mobile Comput., vol. 11, no. 5, pp. 865–877, May 2012. with ATR Adaptive Communications Research
[39] N. Gordon, B. Ristic, and S. Arulampalam, Beyond the Kalman Filter: Laboratories, Kyoto, Japan, from 1997 to 1998. He was a Research Professor
Particle Filters for Tracking Applications. Boston, MA, USA: Artech with the Center for Advanced Communications, Villanova University,
House, 2004. Villanova, PA, USA, from 1998 to 2015. Since 2015, he has been an Associate
[40] ALN-9640 Squiggle Inlay Datasheet, Alien Technol., San Jose, CA, Professor with the Department of Electrical and Computer Engineering,
USA, Feb. 2014. [Online]. Available: http://www.alientechnology.com/ College of Engineering, Temple University, Philadelphia, PA, USA. His
wp-content/uploads/Alien-Technology-Higgs-3-ALN-9640-Squiggle.pdf research interests include statistical signal and array processing, compressive
sensing, convex optimization, time-frequency analysis, MIMO system, radar
imaging, target localization and tracking, adaptive beamforming, and their
applications to radar, communications, satellite navigation, RFID, and radio
astronomy. He has published over 300 journal articles and conference papers
Saurav Subedi received the Ph.D. degree in and 13 book chapters in the above areas. He was a recipient of the 2016
electrical engineering from Villanova University, Premium Award from the Institution of Engineering and Technology for Best
Villanova, PA, USA, in 2016. He is currently Paper published in IET Radar, Sonar and Navigation.
a Data Scientist/Researcher-RF localization with Dr. Zhang is an Associate Editor for the IEEE T RANSACTIONS ON S IGNAL
Bastille Networks, Atlanta, GA, USA. He was a P ROCESSING, and serves on the Editorial Board of the Signal Processing
Research Assistant with the Center for Advanced journal. He was an Associate Editor for the IEEE S IGNAL P ROCESSING
Communications, Villanova University from 2012 to L ETTERS from 2006 to 2010, and an Associate Editor for the Journal of
2016. His research interests include statistical sig- the Franklin Institute from 2007 to 2013. He is a member of the Sensor
nal processing, radar signal processing, sparse signal Array and Multichannel Technical Committee of the IEEE Signal Processing
reconstruction and compressive sensing, and data Society, and the Technical Committee Co-Chair of the 2018 IEEE Sensor
science. Array and Multichannel Signal Processing Workshop.