WiFi-Based Channel Impulse Response Estimation

and Localization via Multi-Band Splicing

Mahdi Barzegar Khalilsarai∗ , Benedikt Gross† , Stelios Stefanatos‡ , Gerhard Wunder† , and Giuseppe Caire∗
∗ Communications and Information Theory Group, Technische Universität Berlin,
† Heisenberg Communications and Information Theory Group, Freie Universität Berlin,
‡ Qualcomm Technologies, Inc.

Emails: {m.barzegarkhalilsarai, caire}@tu-berlin.de, {benedikt.gross, g.wunder}@fu-berlin.de,

[email protected]

Abstract—Using commodity WiFi data for applications such AP 1

as indoor localization, object identification and tracking and (WiFi Device)
channel sounding has recently gained considerable attention. We
study the problem of channel impulse response (CIR) estimation
arXiv:2011.10402v1 [cs.IT] 20 Nov 2020

from commodity WiFi channel state information (CSI). The

accuracy of a CIR estimation method in this setup is limited
1

2
by both the available channel bandwidth as well as various CSI h
t

Path
distortions induced by the underlying hardware. We propose a Pa
multi-band splicing method that increases channel bandwidth
by combining CSI data across multiple frequency bands. In
order to compensate for the CSI distortions, we develop a per- Path 3
band processing algorithm that is able to estimate the distortion
parameters and remove them to yield the “clean” CSI. This AP 2
algorithm incorporates the atomic norm denoising sparse recovery (User Equipment)
method to exploit channel sparsity. Splicing clean CSI over M
frequency bands, we use orthogonal matching pursuit (OMP) as an c1
estimation method to recover the sparse CIR with high (M -fold) c2
CIR c3
resolution. Unlike previous works in the literature, our method
does not appeal to any limiting assumption on the CIR (other ⌧
0 ⌧1 ⌧2 ⌧3
than the widely accepted sparsity assumption) or any ad hoc
processing for distortion removal. We show, empirically, that the Fig. 1: An example schematic of two communicating APs.
proposed method outperforms the state of the art in terms of
localization accuracy.
Index Terms—Commodity WiFi, channel impulse response
estimation, WiFi-based localization, multi-band splicing, phase
scattering, and diffraction properties, and with different “travel
distortion removal, sparse recovery. times”. These characteristics are formally represented by the
channel impulse response (CIR), which is a signal defined over
I. I NTRODUCTION the (propagation) delay domain (see Fig. 1) and is related to the
CFR via the Fourier transform [8]. The CIR can be directly
Using commodity WiFi devices for purposes other than used in any of the applications stated above. Its first delay
classical wireless communication has sparked considerable component is equivalent to the channel Time-of-Flight (ToF),
interest in the recent years, especially with the advent of the the time taken by the signal to travel from the transmitter
Internet of Things (IoT) technology [1]. The reason is that to the receiver via a Line-of-Sight (LoS) path and can be
WiFi devices are ubiquitous, operate at a low cost and their used for localization [3]. The CIR also shows the power delay
acquired data provide highly useful information about the envi- profile, which can be used for channel sounding purposes [9].
ronment. Prominent examples of applications that make use of In addition, comparing the CIR at different times is applicable
commodity WiFi include indoor localization and ranging [2– in recognizing human gesture and activity [7].
4], channel sounding [5], object tracking [6], human activity The discussion above shows that meaningful application
recognition [7], etc. These applications rely on the channel of WiFi CSI data hinges upon an accurate estimation of
state information (CSI) of a communication link between a the CIR. A major challenge for achieving this goal is that,
pair of transmitter-receiver access points (APs). In general, limited channel bandwidth (BW) results in a low-resolution
the CSI refers to the received signal at the device in time, CIR estimate, due to the well-known time (delay)-frequency
frequency and space or a combination thereof. As an example, uncertainty principle. By a low resolution, we mean that the
in a single-antenna WiFi device, the CSI is equivalent to the smaller the channel BW is, the harder it is to distinguish
channel frequency response (CFR) across the used radio bands. between two adjacent delay components in the CIR, even in
The CFR reveals important information about the propagation the absence of noise. Conversely, by increasing the BW, we
environment. In particular, in a multi-path channel, the signal achieve higher resolutions and minute details of the CIR can
travels through distinct paths each with different attenuation, be revealed. For example, a typical WiFi frequency band spans
‡ The work of Stelios Stefanatos was performed while he was with the Freie 20 MHz of BW. This is roughly equivalent to a resolution of
Universität Berlin. ∆τ = 1/BW = 50 ns in the delay domain, which means
that two paths can be discriminated if the difference of their transmitter and the receiver to obtain extra information about
delays is larger than 50 ns. Multiplied by the speed of light the channel. This information is then used to form a likelihood
ν ≈ 3 × 108 , this implies that the distance taken by the function, which we maximize with respect to the delay and
two paths must be larger than ∆d = ν∆τ = 15 m. In phase shift parameters and resolve the due ambiguities.
many practical scenarios (e.g. in small indoor environments),
such condition is violated. A solution that has been recently Notation: Scalars are denoted by simple alphabet letters,
studied is increasing the BW via multi-band splicing, which is while vectors and matrices are bold-faced. For an integer M ,
a technique that merges CFR samples across multiple bands we define [M ] := {1, 2, . . . , M }. We also define the set of
to obtain high-resolution CIR estimates. integers N as N = {− N 2−1 , . . . , N 2−1 }, where N is an odd
However, splicing CFR samples over multiple bands faces R of a signal h(τ ) at a point f is
integer. The Fourier transform
a major challenge due to hardware imperfections. The band denoted by F{h(τ )}|f = h(τ )e−j2πf τ dτ . For a matrix X,
carrier frequencies on the transmitter and receiver sides are XT denotes transpose, XH denotes Hermitian transpose, and
not precisely the same and the discrepancy between the two X† = (XH X)−1 XH denotes Moore-Penrose pseudo-inverse.
varies from band to band [10]. In addition, detecting a packet
at the receiver introduces a delay that is added to the natural
channel ToF [3]. Besides, frequency hopping across different II. S YSTEM S ETUP
bands causes a random phase offset in the CFR samples
since the phase-locked loop (PLL) in charge of generating the
band carrier frequency starts from a random initial phase. The We assume two access points (APs) communicating with each
combined effect of these hardware imperfections results in a other using OFDM signaling over M frequency bands. Each
phase distortion in the CFR samples, which needs to be treated band includes N subcarriers indexed via the integer set N ,
in implementing any reasonable multiband splicing method. As over which the pilot symbols are transmitted. The pilot signal
will be discussed in section II, this phase distortion follows a at the receiver side can (ideally) be written as [8]
piece-wise linear rule: it is linear with fixed unknown param-
y[m, n] = H[m, n]Sm,n + z[m, n], m ∈ [M ], n ∈ N , (1)
eters per band, while the parameters vary from band to band.
Several works in the literature have dealt with the problem where Sm,n = 1 is the symbol transmitted over subcarrier fm,n
of multi-band splicing. The Chronos system was proposed in (n-th subcarrier of band m) and assumed to be equal to one for
[3], which uses the squared CFR samples corresponding to simplicity, H[m, n] is the CFR sample at the same subcarrier
the carrier frequency in each band to estimate the ToF via and z[m, n] ∼ CN (0, 1/SNR) denotes additive white Gaussian
a compressed sensing sparse recovery method. This method noise (AWGN). The CFR is related to the CIR via a Fourier
is only designed to estimated the ToF and can not estimate transform. Suppose there exist K scatterers in the propagation
the CIR. It also does not work in scenarios where only a environment. Then the CIR is given as
few number of bands are used, since in this case the number K
X
of measurements is too small for any compressed sensing h(τ ) = ck δ(τ − τk ), (2)
method to be useful. Spectrum splicing for ToF estimation k=1

was also studied in [9] and [11], where heuristic methods where δ(·) denotes Dirac’s delta function, τk ∈ [0, 1/fs )
were proposed with overall performances inferior to that of and ck ∈ C are the delay and gain associated with path k,
[3]. ToneTrack [12] is another indoor localization system respectively, and where fs denotes subcarrier spacing. Note
which combines channel measurements across multiple bands that the path delay and gain parameters are independent of the
to increase resolution. However, it is not designed for use with frequency band. The CFR samples are then given as
K
commodity WiFi data and therefore it is not clear how it can X
handle the existent signal phase distortions in WiFi CSI. H[m, n] = F{h(τ )}|f =fm,n = ck e−j2πfm,n τk , (3)
k=1
A. Our Contribution for m ∈ [M ], n ∈ N . The pilot signal in a WiFi device
is subject to several phase distortions, caused by hardware
Our proposed method consists of three steps, devised to imperfections. These distortions result in the phase term
address the aforementioned challenges: first, a per-band pro-
ψ[m, n] = −2π(δm nfs + φm ), m ∈ [M ], n ∈ N . (4)
cessing algorithm that estimates the linear phase distortion
parameters and removes them from the CFR sample; a process The parameter δm ∈ [0, 1/fs ) represents receiver timing
that we shall call “cleaning”. This step involves atomic norm offset due to the packet detection delay (PDD) and receiver
denoising [13], which exploits channel sparsity to identify the sampling frequency offset (SFO). Besides, the phase offset
phase distortion components over the continuous domain and term φm ∈ [0, 1) represents the effect of the random phase
without assuming them to be included in a discrete set of offset introduced by the PLL when switching channel bands,
plausible values. Secondly, we use the clean CFR samples in addition to the accumulated phase offset across the slots due
gathered from all bands to obtain a high-resolution estimate of to the carrier frequency offset (CFO) between transmitter and
the CIR. This step is implemented via the greedy orthogonal receiver [3, 7, 11]. Notice that δm and φm differ from band
matching pursuit (OMP) sparse recovery method that exploits to band, such that ψ[n, m] is a piecewise linear function (on
signal sparsity. The CIR obtained at this step is a “relative” each band it is a linear function of the subcarrier index with
one, i.e. it contains ambiguities with respect to a shift in the different slope and constant terms).
delay domain as well as a global phase shift in the coefficients.
As a final step we employ a handshaking protocol between the The actual received pilot signal including the unavoidable
phase distortion is given by surements to always begin at τ = 0 and its first coefficient to
y[m, n] = ejψ[m,n] H[m, n] + z[m, n] have a zero phase. This is equivalent to shifting all CIR delay
K parameters by τ1 and subtracting the phase of the first coef-
ficient c1 from the phases of all coefficients. Mathematically,
X
= ejψ[m,n] ck e−j2πfm,n τk + z[m, n]
k=1 via this assumption we try to recover a “relative” CIR of the
K form
X K
= ck e−j2π(fm,0 τk +φm ) e−j2πnfs (δm +τk ) + z[m, n], h0 (τ ) =
X
ck δ(τ − τek ),
e (7)
k=1
(5) k=1
−j∠c1
where we have used the fact that the subcarriers of a single where eck = ck e , τek = τk − τ1 , for k = 1, . . . , K with
band are equispaced with a spacing equal to fs and therefore ∠c1 being the phase of the first CIR coefficient. Note that
we have fm,n = fm,0 + nfs , n ∈ N with fm,0 being the τe1 = 0 and e
c1 = |c1 |. Once h0 (τ ) is estimated, we can recover
carrier frequency of band m. We denote the noise term as τ1 and ∠c1 via the handshaking protocol.
before by z[m, n] since it is circularly symmetric and the
multiplication of the phase-distortion term does not change III. R EMOVING P HASE D ISTORTION FROM P ILOT S IGNALS
its distribution.
The main problem can be posed as follows: given the noisy, We propose a method to clean the pilot measurements by
phase-distorted pilot signals y[m, n], m ∈ [M ], n ∈ N , adopting the relative CIR model in (7) and removing the phase
estimate the CIR h(τ ). Our proposed solution to this problem distortion terms for each band separately. First, note that for
consists in the following steps: each band m, the pilot signals y[m, n], n ∈ N represent the
samples of a noisy mixture of K complex sinusoids (see (5)).
1) Estimate the distortion parameters {δm , φm } for all
Using the relation fm,n = fm,0 + nfs , n ∈ N , the vector of
bands m ∈ [M ] and remove the phase distortion from
pilot signals in band m can be written as
the pilot signals. K
2) Splice clean pilot data over all bands to obtain a high- X (m) (m)
y[m, n] = ck e−j2πnfs τk + z[m, n]
resolution CIR estimate. (8)
k=1
3) Use a hand-shaking procedure between the APs to
=: x[m, n] + z[m, n],
resolve ambiguities and estimate the ToF.
where
(m)
In the following we explain each of these steps in details. How- ck = ck e−j2π(fm,0 τk +φm ) ,
ever, first we need to clarify what we mean by “ambiguities” (m)
(9)
in step (3). τk = τk + δm ,
Observation 1 (Inherent Ambiguity): Suppose that a spe- and we have denoted the signal part of y[m, n] by x[m, n].
PK (1) (1)
cific CIR h(1) (τ ) = k=1 ck δ(τ −τk ) in combination with Our goal consists in estimating, for each band m, the set
(m) (m) K
(1) (1)
a specific set of phase distortion parameters δm , φm , m ∈ of parameters {τk }K k=1 and {ck }k=1 . Since the CIR is
[M ] and AWGN samples z (1) [m, n] generates pilot signals sparse (K  N ), sparse recovery via compressed sensing is
y (1) [m, n] according to (5). Now, consider a second CIR a natural choice for an estimation method. A powerful tool
PK (2) (2) in this regard is the atomic norm denoising superresolution
h(2) (τ ) = k=1 ck δ(τ − τk ) and associated phase dis-
(2) (2)
tortion parameters δm , φm , m ∈ [M ] and the same AWGN technique. Define the atoms a(τ, θ) ∈ CN , τ ∈ [0, 1/fs ], θ ∈
samples z (1) [m, n]. Let the parameters in the latter case be [0, 2π) with elements [a(τ, θ)]n = e−j(2πnfs τ −θ) , n ∈
such that: N . The set of atoms above constructs a continuous dic-
(2) (1) (2) (1) tionary A = {a(τ, θ) : τ ∈ [0, 1/fs ], θ ∈ [0.2π)}.
τk = τk − δ̄, ck = ck ej2πφ̄ , for all k, T
Now, let y(m) = y[m, − N 2−1 ], . . . , y[m, N 2−1 ]

(2) (1)
(6) ∈ CN
δm = δm + δ̄, φ(2) (1)
m = φm + φ̄, for all m, denote the vector of pilot samples in band m, x(m) =
T
where δ̄ and φ̄ are arbitrary values. Then it is easy to show x[m, − N 2−1 ], . . . , x[m, N 2−1 ]

its signal component and
that the generated pilot signals in the two cases coincide, i.e. z(m) the associated AWGN vector. It follows that
y (2) [m, n] = y (1) [m, n] for all m, n for all m, n. Therefore, K
(m) (m)
X
given the pilot signals, both sets of parameters in cases (1) x(m) = |ck |a(τk , ∠ck ), (10)
and (2) are equally plausible. This shows that there exist two k=1
inherent ambiguities in estimating the CIR parameters: (a) that is, x(m) is a linear combination of a few elements of A.
an ambiguity in terms of a (circular) shift of the CIR over We can define the atomic norm of a generic vector r ∈ CN
the delay domain and (b) an ambiguity in terms of a global over A is defined as [14]
phase shift of the CIR coefficients. Therefore, given only the krkA = inf
X X
{ wk : r = wk a(τk , θk )}.
phase-distorted pilot signals as in (5), we are undecided about wk ≥0, τk ∈[0,1/fs ]
k k
θk ∈[0,2π)
two scalar parameters, namely a delay shift and a phase shift. (11)
The delay shift is especially important here, as it effects the It is well known that the atomic norm promotes a sparse
estimation of the ToF. In step (3) of our proposed method, we representation of r in A and therefore is a suitable choice for
introduce a hand-shaking procedure to acquire some additional estimating the sparse vector x(m). If there were no noise, we
information about the CIR that helps resolve these ambiguities. would directly observe y(m) = x(m) and the sparse recovery
Assumption 1 (Relative CIR): The observation above con- problem could be posed as [14]:
veys an algorithmic implication. Without loss of generality
(w.l.o.g), we can assume the CIR estimated from pilot mea- minimize ke
xkA subject to [e
x]n = [y(m)]n , n ∈ N , (12)
e∈CN
x
which simply finds the vector that has the sparsest represen- IV. M ULTI -BAND S PLICING AND CIR E STIMATION
tation in A and whose elements are equal to the observed
CFR samples of band m. In the problem in hand, however, The idea of multi-band splicing is to combine the clean pilot
we have access only to the noisy pilot observations y(m) = measurements over all M bands to estimate the CIR h(τ ).
x(m) + z(m). In this case, the atomic norm denoising method The benefit of splicing several bands is that it increases the
was proposed in [13] which solves the following regularized achievable resolution in estimating h(τ ) by increasing the
convex problem: measurement bandwidth. While with a single band we have
1 a resolution of (∆τ )1 = 1/N fs over the delay domain,
minimize ke x − y(m)k2 + λm ke xkA , (13) using M bands we can potentially achieve a resolution of
e∈C
x N 2
where λm > 0 is a suitable regularization scalar. Fortunately, (∆τ )M = 1/M N fs which is M -fold smaller (finer).
the atomic norm denoising optimization problem has an equiv- A. Estimating the Relative CIR h0 (τ )
alent SDP form as [13]
1 λm For band m, define the vector of band subcarriers as
minimize x − y(m)k2 +
ke (t + [u]1 ) f (m) = [f N −1 , . . . , f
T
N −1 ] and the vector containing
t,u,e
x 2 2 m,− 2 m, 2
  (14) all subcarriers as f = [f (1)T , . . . , f (M )T ]T ∈ RM N . Simi-
T (u) x
subject to  0,
e
x
e H
t larly, we can define the vector of clean pilot measurements
corresponding to the subcarriers as follows. Let y e (m) =
where T (u) is the Toeplitz Hermitian matrix with u as its first T
ye[m, − N 2−1 ], . . . , ye[m, N 2−1 ]

column. The solution of (14) for x b(m) that has
b is a vector x denote the clean measure-
the following representation ments vector of band m and y e = [ey(1)T , . . . , y
e (M )T ]T ∈
MN
X (m) (m) (m)
X (m) (m) C as the multi-band spliced vector containing all CFR
b(m) =
x |b
ck |a(bτk , ∠b ck ) = ck a(b
b τk , 0). measurements. We can write the elements of ỹ as
k k
(m) (m) [e
y]i = F{h0 (τ )}|[f ]i + [e
z]i , i = 1, . . . , M N (18)
We are interested in the parameters τbk , θbk , as they contain
(m) where [z̃]i represents the AWGN plus the error produced by the
the phase distortion data. The harmonic parameters τbk can
phase-distortion removal procedure (i.e. the error in estimating
be obtained by solving the dual problem of (14) and finding
the parameters δm , φm , m ∈ [M ]). Since h0 (τ ) is sparse, a
those points in which the dual polynomial has a maximum
standard approach to estimating it from the frequency samples
modulus (details are omitted due to a lack of space; see
ỹ is using a compressed sensing method. Define a uniform grid
[13]). Denote the estimated support of x b(m) over the delay 1
(m) (m) b of size G over the delay domain as G = {0, G , . . . , G−1
G }/fs .
domain by τbk , k = 1, . . . , K.b The coefficients {b ck }K k=1 M N ×G
(m) (m) T Let D = [d(0), . . . , d(G − 1)] ∈ C be an overcomplete
are computed as b c (m)
= [bc1 , . . . , bcKb ] = A † y(m),
(m)
dictionary (G  M N ) associated with G, where each column
h i
(m) (m)
where A(m) := a(b τ1 , 0), . . . , a(b
τKb , 0) . d(i) is defined as
i i
d(i) = √ 1 [e−j2π[f ]1 ( G )/fs , . . . , e−j2π[f ]M N ( G )/fs ]T
∈ CM N
MN
Remark 1: The optimal value for the regularization param- (19)
eter λm depends on the dual norm of the noise term, and for for i = 0, 1, . . . , G − 1. Now, if the grid G is dense enough,
the AWGN model is derived as [13] we can approximate the vector form of (18) as
1 + 1/ log N p
λm = √ N log N + N log (4π log N ). (15) e ≈ Dh0 + e
y z, (20)
SNR G
where h0 ∈ C is a discrete approximation for h0 (τ ). In
Hereafter, we use this value for the regularization parameter
practice, setting G to a value of G = 2M N or G = 3M N
for solving (13).
yields a sufficiently dense grid.
Given ye , we estimate the h0 using the well-known or-
thogonal matching pursuit (OMP) sparse recovery method
A. Removing phase distortions [15]. Using the model (20), OMP can be seen as a greedy
iterative algorithm that selects, at each iteration, a column
of the dictionary D that has the highest correlation with the
Recall that via Assumption 1, we postulate (w.l.o.g) the CIR current residual and repeats this until a convergence condition
to be equal to the relative CIR h0 (τ ) in (7) with parameters is met. Then, the non-zero coefficients associated with each
ck , τek , k = 1, . . . , K where ∠e
e c1 = 0 and τe = 0. Assuming selected column to approximate the measurements vector y e are
this and since x b(m), m ∈ [M ] are estimated from the phase- computed by solving a simple least-squares problem. We refer
(m)
distorted measurements, from (9) we have that τb1 is an the reader to [15] for details. As for the stopping condition,
(m)
estimate of δm and ∠b c1 is an estimate of −2πφm . In other we assume that the sparsity order of h0 is given and halt the
words, the estimates of the distortion parameters for band m algorithm once the number of selected dictionary columns is
are given as: equal to the sparsity order.
(m) Remark 2: One may wonder why we do not use atomic
δbm = est(δm ) = τb ,
1
(m) (16) norm denoising once more in this setting. The reason is
∠bc1 twofold: first, the atomic norm has an SDP representation
φbm = est(φm ) =− .
2π when we have uniformly obtained samples of the complex
Therefore, we can remove the phase distortions by calculating exponential mixture. The frequency samples limited to a single
the clean pilot CFR samples as band are in fact taken over uniformly spaced subcarriers
ye[m, n] = ej2π(δm nfs +φm ) y[m, n]. (17) fm,0 + n fs for n = − N 2−1 , . . . , N 2−1 so that the SDP form is
b b
guaranteed. In contrast, in general the elements of the multi- exact, then we would have H b (τ ,∠c ) [m, 0] = H[m, 0], m ∈
1 1
band subcarrier vector f do not necessarily lie on a uniform [M ], i.e. the estimate coincides with the true CFR for parame-
grid. In this case a tractable formulation of the atomic norm ters τ̄ = τ1 and θ̄ = ∠c1 . This observation implies that a natu-
is unknown. Secondly, even if the subcarriers lie on a uniform ral way to estimate τ1 and ∠c1 is to compare H b (τ ,∠c ) [m, 0]2
1 1
grid and the SDP form exists, solving the corresponding SDP with yrx [m, 0] ytx [m, 0] = H[m, 0]2 + z 0 [m, 0] from (23) and
is not desirable due to the high computational complexity minimizing their difference simultaneously for all m ∈ [M ].
of SDPs in large dimensions. Therefore, using a grid-based Formally, this can be cast as the following problem:
compressed sensing method such as OMP is more reasonable.
PM  2
B. Handshaking, Resolving Ambiguities and ToF Estimation minimize C(τ̄ , θ̄) =

rx [m, 0] y tx [m, 0] − b (τ̄ ,θ̄) [m, 0]2  .
H
m=1 y
(τ̄ ,θ̄)∈Ξ
By Observation 1 we know that with the given phase- (25)
distorted pilot signals, there always exists an ambiguity with The variable domain Ξ can be chosen in various ways. We
respect to a circular shift of the CIR over the delay domain and have empirically observed that choosing a discrete set as
a global phase shift in the coefficients. From this observation Ξ = {(τi , θi ) ∈ [0, 1/fs ) × [0, 2π), i = 1, . . . , I}
we concluded that, w.l.o.g, we can assume the relative CIR is sufficient. Then minimizing the cost C(τ̄ , θ̄) is a simple
h0 (τ ) that has its first delay component at τ = 0 with a zero- and fast minimization over a discrete set. Defining the optimal
phased coefficient (see (7)). The relation between the relative parameters as
and the true CIRs is as follows:
(τ ? , θ? ) = arg min C(τ̄ , θ̄), (26)
h(τ ) = ej∠c1 h0 (τ − τ1 ). (21) (τ̄ ,θ̄)∈Ξ

Therefore it remains to estimate τ1 and c1 using extra infor- the estimate of the true CIR is given by
mation about the channel. This extra information is obtained G
jθ ? b 0 ]i δ(τ − i − 1 − τ ? ).
X
via a handshaking procedure that was suggested in [3] and is h(τ ) = e
b [h (27)
i=1
G
explained in the following.
From the affine phase distortion model developed before, In addition, the ToF is estimated as
we know that the zero subcarrier in each band (band carrier d = τb1 = τ ? .
ToF (28)
frequency) is not polluted by the PDD and SFO phase errors,
but only by the constant phase error term φm induced by V. S IMULATION R ESULTS
the CFO and PLL phase offsets. This constant error term has
the same absolute value but different signs on the transmitter- In this section, we compare our method to Chronos [3] in
receiver APs (see [3] Eqs. (11) and (12)). Therefore, for each terms of localization accuracy. To this end, let us first define
band m, we can write the pilot CSI at the two APs and on the the average ranging estimation error as
center carrier frequency as
ed = |τ1 − τb1 |ν, (29)
ytx [m, 0] = ejφm H[m, 0] + ztx [m, 0] 8
(22) where ν ≈ 3 × 10 is the speed of light. To generate random
yrx [m, 0] = e−jφm H[m, 0] + zrx [m, 0], CIRs, we consider channels with K = 3 delay taps. The
where ztx [m, 0] and zrx [m, 0] denote noise terms at the trans- maximum path distance is set to dmax = 100 m, which
mitter and receiver sides, respectively. Notice the difference means that the longest path takes a total distance of dmax
in the sign of the phase distortion terms on both sides. to reach the receiver, which is a reasonable choice for an
Also note that the CFR H[m, 0] is the same on both ends indoor environment. This is equivalent to setting the maximum
due to channel reciprocity [8]. During a frequency-hopping delay spread equal to τmax = dmax /ν ≈ 333 ns. The
procedure the transmitter sends packets to the receiver through delay components are chosen uniformly at random in the
which the receiver obtains the zero carrier measurements interval [0, τmax ]. The path gains {ck }Kk=1 are generated as
ytx [m, 0], m = 1, . . . , M . Therefore, at the receiver side we complex Gaussian random variables, with decreasing variance
have both yrx [m, 0] as well as ytx [m, 0] for all m. Multiplying as σk2 = 4−k , k = 1, . . . , K. This is a natural choice, since
the two values for each m we get typically the path with a larger delay is attenuated more than
yrx [m, 0] ytx [m, 0] = H[m, 0]2 + z 0 [m, 0], m ∈ [M ], (23) a path with a smaller delay. We consider M = 16 frequency
where z 0 [m, 0] := ejφm H[m, 0]zrx [m, 0] + bands each with N = 65 subcarriers. With a subcarrier spacing
e −jφm
H[m, 0]ztx [m, 0] + zrx [m, 0]ztx [m, 0] denotes the of fs = 312.5 kHz, each band occupies (N − 1)fs = 20 MHz
signal and noise cross-terms, which especially in high-SNR of bandwidth. As a proof of concept, we have considered half
regimes can be safely assumed to be of small value. Equation of the bands (8 bands) in the (2, 2.19) GHz range and the
(23) gives the noisy squared CFR samples on the zero carrier other half in the (5, 5.19) GHz range. Note that the 5 GHz
per band. On the other hand, using the discrete estimate of range is already in use by the IEEE 802.11a WiFi standard
the relative CIR h b 0 obtained in the previous step and the [16]. For each band m, the phase distortion parameters are
relation (21), we can estimate H[m, 0] as a function of the randomly generated as follows: we assume the PDD and SFO
unknown delay shift τ̄ and phase shift θ̄ parameters as to cause a delay distortion of up to 960 ns, so that δm is chosen
G
uniformly at random (and independently across bands) in the
b 0 ]i e−j2πfm,0 ( i−1 range [0, 960] ns. Also the PLL and CFO induced constant
X
b (τ̄ ,θ̄) [m, 0] = ej θ̄ e−j2πfm,0 τ̄
H [h G )/fs

i=1
phase distortion term φm is chosen uniformly at random (and
(24) independently across bands) in [0, 2π). We assume the signal-
for m ∈ [M ]. Note that if the estimate of the relative CIR were to-noise-ratio (SNR) to be equal to SNR = 20 dBs.
 1 VI. C ONCLUSION
Proposed Method
0.9 Chronos
We proposed a method for CIR estimation and indoor localiza-
0.8 tion using commodity WiFi CSI data. This method performs
0.7 multi-band CFR splicing and exploits channel sparsity to
achieve high resolution CIR estimates. In order to compensate
0.6
for inherent hardware-induce distortion in the CSI. we pro-
CDF

0.5 posed a per-band processing based on atomic norm denoising

which estimates the distortion parameters and remove them
0.4
from the CSI samples. Via empirical simulations we showed
0.3 that our method outperforms the state-of-the-art in terms of
0.2
ranging error.

0.1 R EFERENCES
0 [1] L. Li, H. Xiaoguang, C. Ke, and H. Ketai, “The applications
10−5 10−4 10−3 10−2 10−1 100 101 102 of WiFi-based wireless sensor network in internet of things
ed (m) and smart grid,” in 2011 6th IEEE Conference on Industrial
Electronics and Applications. IEEE, 2011, pp. 789–793.
Fig. 2: CDF of the ranging error for Chronos and our proposed [2] C. Yang and H.-R. Shao, “WiFi-based indoor positioning,” IEEE
method. Here we have M = 16, N = 65 and SNR = 20 dBs. Communications Magazine, vol. 53, no. 3, pp. 150–157, 2015.
[3] D. Vasisht, S. Kumar, and D. Katabi, “Decimeter-level localiza-
tion with a single WiFi access point.” in NSDI, vol. 16, 2016,
pp. 165–178.
[4] M. B. Khalilsarai, S. Stefanatos, G. Wunder, and G. Caire,
The Chronos method uses only the M handshaking carrier
“WiFi-based indoor localization via multi-band splicing and
frequency CFR samples in (23), which are noisy versions phase retrieval,” in 2019 IEEE 20th International Workshop
of {H[m, 0]2 }. Then, using the CIR sparsity assumption, it on Signal Processing Advances in Wireless Communications
applies the basis pursuit denoising (BPDN) algorithm to these (SPAWC). IEEE, 2019, pp. 1–5.
measurements. BPDN can be formulated as [5] D. Maas, M. H. Firooz, J. Zhang, N. Patwari, and S. K. Kasera,
“Channel sounding for the masses: Low complexity gnu 802.11
minimize
0
kxk1 subject to kq − Fxk ≤ , b channel impulse response estimation,” IEEE transactions on
x∈CG
wireless communications, vol. 11, no. 1, pp. 1–8, 2011.
T
where q = [yrx [1, 0] ytx [1, 0], . . . , yrx [M, 0] ytx [M, 0]] , F ∈ [6] K. Qian, C. Wu, Y. Zhang, G. Zhang, Z. Yang, and Y. Liu,
0
CM ×G is the over-sampled Fourier matrix where [F]m,i = “Widar2. 0: Passive human tracking with a single Wi-Fi link,”
i−1
√1 e−j2πfm,0 G0 /fs , m ∈ M, i ∈ [G0 ] with G0 is typ-
in Proceedings of the 16th Annual International Conference on
M Mobile Systems, Applications, and Services, 2018, pp. 350–361.
ically chosen as a multiple of M . Also  is an estimate [7] X. Shen, L. Guo, Z. Lu, X. Wen, and Z. He, “WiRIM: Resolu-
on the `2 norm of the noise plus cross-terms vector z0 = tion improving mechanism for human sensing with commodity
T
[z 0 [1, 0], . . . , z 0 [M, 0]] (z 0 [m, 0] is defined in (23)). Although Wi-Fi,” IEEE Access, vol. 7, pp. 168 357–168 370, 2019.
this can not be done in practice, to be concrete, we set  to [8] D. Tse and P. Viswanath, Fundamentals of wireless communi-
cation. Cambridge university press, 2005.
be exactly equal to kz0 k in our simulations. The result is an [9] Y. Xie, Z. Li, and M. Li, “Precise power delay profiling with
sparse estimate of the CIR convoluted with itself h(τ ) ∗ h(τ ) commodity Wi-Fi,” IEEE Trans. on Mobile Computing, 2018.
(since H[m, 0]2 = F{h(τ ) ∗ h(τ )}|f =fm,0 )and therefore, its [10] T.-D. Chiueh, P.-Y. Tsai, L. I-Wei, and T.-D. Chiueh, Baseband
first significant component is expected to be located at 2τ1 . receiver design for wireless MIMO-OFDM communications.
Dividing this location by 2 gives an estimate of τ1 , i.e. the Wiley Online Library, 2012.
[11] Y. Zhuo, H. Zhu, H. Xue, and S. Chang, “Perceiving accurate
ToF. We do not further explain this method due to a lack of CSI phases with commodity WiFi devices,” in INFOCOM 2017-
space and refer the reader to [3] for details. Fig. 2 illustrates IEEE Conference on Computer Communications, IEEE. IEEE,
the ranging error CDF for our proposed method as well as 2017, pp. 1–9.
for Chronos. The CDF is obtained after 200 Monte-Carlo [12] J. Xiong, K. Sundaresan, and K. Jamieson, “Tonetrack: Lever-
simulations. As we can see, our proposed method achieves aging frequency-agile radios for time-based indoor wireless
localization,” in Proceedings of the 21st Annual International
a ranging error that is 2 to 3 orders of magnitude smaller Conference on Mobile Computing and Networking, 2015, pp.
than Chronos: in 90 % of the instances our method achieves 537–549.
a ranging error of ≈ 5 cm or less, whereas the ranging [13] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic norm denoising
error for Chronos is ≈ 34 m or less. This figure also shows with applications to line spectral estimation,” IEEE Transactions
that in about 5 % of the instances the ranging error of our on Signal Processing, vol. 61, no. 23, pp. 5987–5999, 2013.
[14] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed
method is 1 meters or larger. The reason for such outliers is sensing off the grid,” IEEE transactions on information theory,
that sometimes the phase distortion parameters are estimated vol. 59, no. 11, pp. 7465–7490, 2013.
poorly, either due to noise or two delay components being [15] T. T. Cai and L. Wang, “Orthogonal matching pursuit for sparse
very close such that the atomic-norm denoising step fails signal recovery with noise,” IEEE Transactions on Information
to distinguish between them. In such cases, compensating theory, vol. 57, no. 7, pp. 4680–4688, 2011.
[16] IEEE802.11a, “Supplement to IEEE 802.11 standard —part
for phase distortions induces an error in the estimation of 11: Wireless LAN, medium access control (MAC),and physical
the relative CIR and eventually the handshaking step fails to layer (PHY) specifications: High-speed physical layer in the 5
correctly identify the ToF. However, such errors happen rarely GHz band,” Sept. 1999.
and do not effect the average performance of our method.