You are on page 1of 5

Cramér-Rao Bound Analysis of Wi-Fi Indoor

Localization Using Fingerprint and Assistant Nodes


Qiyue LI∗ , Wei LI∗ , Wei SUN∗ , Jie LI† , Zhi LIU§
∗ School of Electrical Engineering and Automation, Hefei University of Technology, Hefei, China 230009
Email: liqiyue@mail.ustc.edu.cn, weili@mail.hfut.edu.cn, wsun@hfut.edu.cn
† School of Computer and Information, Hefei University of Technology, Hefei, China 23009
Email: lijie@hfut.edu.cn (Corresponding author)
§ Department of Mathematical and Systems Engineering, Shizuoka University, Japan 432-8561

Email: liu@ieee.org

Abstract—Location estimation in Wi-Fi environment has analysis framework and characterization of the properties of
gained considerable attention over the past years, and the this lower bound are presented to evaluate the impact of
Cramér-Rao Lower Bound (CRLB) can be used to evaluate different parameters on the accuracy of coordinates estimation.
the performance of the localization system. In this paper, we
analyze the CRLB of Wi-Fi indoor localization using fingerprint Furthermore, it can also provide important system design sug-
and assistant nodes. This localization method combines received gestions by revealing error trends with the indoor localization
signal strength (RSS) and Time of Arrival (TOA) into together, system deployment. As far as we know, this is the first time to
and constructs a fixed spatial model with several assistant nodes analyze the CRLB for localization system using hybrid RSS
to improve localization performance. There are two purposes of fingerprint and assistant nodes.
the CRLB analysis framework proposed in this paper. Firstly,
the expression of lower bound on location estimation error can The rest of this manuscript is organized as follows. Section
help in designing and refining efficient localization algorithm and II reviews the related work on the localization model. And in
parameters. Secondly, the error trends can provide suggestions Section III, analysis framework of the CRLB for the hybrid
for a positioning system design and deployment. Furthermore, RSS fingerprint and assistant nodes is depicted in detail. In
detailed analysis as well as experimental results are both pre-
Section IV, evaluation results are presented, and the papers
sented in this paper.
Index Terms—Localization System, Wireless Fingerprint, conclusions are presented in Section V.
Cramér-Rao Lower Bound, Time of Arrival, Assistant Node.
II. RELATED WORK
I. I NTRODUCTION
A. The Cramér-Rao Lower Bound
Currently, a wide range of Wi-Fi applications are applied in
It is well known that CRLB sets a lower limit for co-
indoor environment to monitor environmental and structural
variance matrix of any unbiased estimate of parameters. The
change, which requires knowledge of sensor locations [1].
CRLB is commonly used as a performance benchmark of
Meanwhile, there are a lot of localization estimation schemes
an estimator because it gives the lowest possible variance
used in long range communications, such as time of arrival
that any linear estimator can achieve [6]. And it provides a
(TOA) and received signal strength (RSS) [2]. With the greatly
benchmark against which we can compare the performance of
increasing demand for indoor localization services, so many
any unbiased estimator. Furthermore, it alerts us to the physical
localization algorithm are proposed and deeply researched.
impossibility of finding an unbiased estimator whose variance
Based on that, the hybrid RSS/TOA estimation schemes are
is less than the bound [7]. Therefore, the CRLB is widely
applied to improve indoor localization accuracy, which takes
used to evaluate the fundamental hardness of an estimation
full advantage of the respective character of RSS and TOA
problem. If 𝜃ˆ = (ˆ 𝑥, 𝑦ˆ)𝑇 is the estimate of the unknown
measurements [3]. In [4], we proposed a Wi-Fi indoor localiza-
device’s coordinates 𝜃 = (𝑥, 𝑦)𝑇 , then its covariance matrix
tion method based on collaboration of fingerprint and assistant
is,
nodes. Besides RSS fingerprint, we elaborately select several
[ 2 ]
assistant nodes around the unknown device in space by com-
ˆ ˆ ˆ 𝑇 𝜎𝑥ˆ 𝜎𝑥ˆ𝑦ˆ
𝐶𝑜𝑣𝜃 (𝜃) = 𝐸𝜃 {(𝜃 − 𝜃)(𝜃 − 𝜃) } = (1)
paring the RSS values sequence. And then with the distance 𝜎𝑦ˆ𝑥ˆ 𝜎𝑦ˆ2
information measured by TOA between all the assistants and
the unknown node, we try to establish a fixed spatial model where 𝐸𝜃 {⋅} is the expectation operator conditioned on 𝜃. The
to enhance localization accuracy. diagonal elements of (1) represent the mean squared errors
The Cramér-Rao Lower Bound (CRLB) is widely used to and the off-diagonal elements are the covariances between
evaluate the performance of indoor location algorithm [5]. different parameters. The CRLB is defined as the inverse of
In this paper, we analyze the CRLB of location estimation the Fisher information matrix (FIM) as [8]. If 𝑓𝜃 (𝑃 ) denotes
error using the RSS fingerprint and assistant nodes. A novel the probability density function of observations 𝑃 conditioned

Authorized licensed use limited to: ASU Library. Downloaded on December 17,2023 at 03:22:38 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-5935-5/17/$31.00 ©2017 IEEE
on 𝜃, then the score function is defined as the gradient of its
Fingerprint Fingerprint Database
log-likelihood, i.e., RP Set Match
Training Phase

∂ Offline phase Online phase

𝑈 (𝜃) = ▽ln𝑓𝜃 (𝑃 ) = ln𝑓𝜃 (𝑃 ) (2) Initial localization


∂𝜃 RSS fingerprint
database
RSS sample
process

Then FIM, 𝐽 is defined as the variance of this score function, RSS Data Collect And Process
Assistant nodes selection

Location
∂ln𝑓𝜃 (𝑃 ) 2 TOA ranging and NLOS Estimation
𝐽(𝜃) = 𝐸{[ ] }
error mitigation
(3)
∂𝜃 Fixed spatial model
construction
The CRLB is defined as the inverse of FIM and obviously
ˆ ≥ {𝐽(𝜃)}−1 .
Cov𝜃 (𝜃)

B. CRLB for RSS based wireless localization


Fig. 1. Fingerprint and assistant nodes based localization model
RSS is defined as the voltage measured by a receivers
received signal strength indicator (RSSI) circuit. Often, RSS
is equivalently reported as measured power, i.e., the squared III. CRLB FOR FINGERPRINT AND ASSISTANT NODES
magnitude of the signal strength. Typically, the ensemble mean BASED LOCALIZATION
received power in a real-world, obstructed channel decays
A. Review of fingerprint and assistant nodes based localiza-
proportional to 𝑑−𝑛𝑝 , where 𝑛𝑝 is the path-loss exponent,
tion
typically between two and four, depending on complexity of
the indoor environment. The ensemble mean power at distance In [4], a Wi-Fi indoor localization method based on col-
𝑑 is typically modeled as laboration of fingerprint and assistant nodes is proposed. It
not only utilizes the RSS fingerprint database to get a coarse
𝑃 𝑑 = 𝑃0 − 10𝑛𝑝 log(𝑑/𝑑0 ) (4) result, but also selects assistant nodes which indicate the nodes
closed to the unknown node in space by comparing the RSS
where 𝑃0 is the mean received power (dBm) at the reference values sequence. And then we take the distances measured
distance 𝑑0 . The standard deviation of received power (when with TOA between all the assistants and the unknown node
received power is expressed in dBm), 𝜎𝑑𝐵 , is expressed into account. Aiming at the severe NLOS error caused by the
in units of dB and is relatively constant with distance [9]. complicated indoor environment, an adaptive Kalman filter
Typically, 𝜎𝑑𝐵 is as low as four and as high as 12. Thus, the with colored noise is used to eliminate the ranging error to
received power (dBm) at unknown device 𝑖 transmitted by 𝑗, enhance localization accuracy. Based on the additional ranging
𝑃𝑖,𝑗 is distributed as information, a fixed graph structure containing all the assistant
and unknown nodes is established to optimize localization
𝑓 (𝑃𝑖,𝑗 = 𝑝∣𝜃) = 𝑁 (𝑝; 𝑃¯ (𝑑𝑖,𝑗 ), 𝜎𝑑𝐵
2
) (5) error and reduce the effects of the environments complexity.
The fingerprint and assistant nodes based localization model
where 𝑁 (𝑥; 𝑦, 𝑧) is the notation for the value at 𝑥 of a Gaus- can be described in Figure 1.
sian probability density function with mean 𝑦 and variance Specifically, let us consider a scenario with 𝑚 APs and
𝑧. 𝑛 + 1 blindfolded devices which includes one target device
Then according to [10], the FIM can be expressed as, and 𝑛 assistants. Without loss of generality, we denote number
𝑚 0 as the target device, number 1 𝑚 as APs, and number
∑ (𝑥𝑖 − 𝑥0 )2
[𝑅]
𝐽𝑥𝑥 (𝜃) = 𝜌[𝑅] (6) 𝑚 + 1 𝑚 + 𝑛 as the selected 𝑛 assistants. In a 2-D system,
𝑖=1
𝑑4𝑖0 we denote 𝜃𝑖 = [𝑥𝑖 , 𝑦𝑖 ]𝑇 as the real coordinates of node
𝑖. The relative localization problem corresponds to estimate
𝑚
∑ (𝑥𝑖 − 𝑥0 )(𝑦𝑖 − 𝑦0 ) coordinates of the target device, 𝜃0 = [𝑥0 , 𝑦0 ]𝑇 , given the
[𝑅] [𝑅]
𝐽𝑥𝑦 (𝜃) = 𝐽𝑦𝑥 (𝜃) = 𝜌[𝑅] (7) known coordinates of APs, [𝜃1 , ⋅ ⋅ ⋅ , 𝜃𝑚 ], the estimated coarse
𝑑4𝑖0
𝑖=1 coordinates [𝜃ˆ𝑚+1 , ⋅ ⋅ ⋅ , 𝜃ˆ𝑚+𝑛 ], and all the TOA measurements
𝑚
𝑇𝑖𝑗 between each two blindfolded devices 𝑖, 𝑗, 𝑖, 𝑗 = 0 or 𝑚+
∑ (𝑦𝑖 − 𝑦0 )2 1 ≤ 𝑖, 𝑗 ≤ 𝑚 + 𝑛, 𝑎𝑛𝑑 𝑖 ∕= 𝑗.
[𝑅]
𝐽𝑦𝑦 (𝜃) = 𝜌[𝑅] (8)
𝑖=1
𝑑4𝑖0
B. CRLB for TOA measurement with eliminated NLOS error
10𝑛
where 𝜌[𝑅] = ( 𝜎𝑑𝐵 ln𝑝10 )2 , 𝑚 is the number of APs or anchor As detailed depicted in [8], we could denote 𝑇𝑖𝑗 as the
nodes, while 𝑑𝑖0 represents the true distance between AP 𝑖 one way ranging measurement taken for the radiated signal to
and the unknown node, which is numbered as 0. travel from the device 𝑖 to device 𝑗. The ranging measurement
Thus 𝐶𝑅𝐿𝐵𝑅 for RSS based localization is equal to 𝑇𝑖𝑗 can be modeled as 𝑇𝑖𝑗 = 1𝑐 × {𝑑𝑖𝑗 + 𝑛𝑖𝑗 + 𝑏𝑖𝑗 }, where 𝑐
{𝐽 [𝑅] (𝜃)}−1 , the matrix inverse of the FIM. is the speed of propagation, 𝑑𝑖𝑗 is the true distance between 𝑖

Authorized licensed use limited to: ASU Library. Downloaded on December 17,2023 at 03:22:38 UTC from IEEE Xplore. Restrictions apply.
and 𝑗, 𝑛𝑖𝑗 represents the ranging measurement noise subjected
to Gaussian distribution with zero mean and variance 𝜎𝑇2 , Node A
and 𝑏𝑖𝑗 is a positive random variable representing the error Node A
due to NLOS propagation, which can be modeled as different
statistical distributions in different channel environments.
Obviously, TOA measurements would change significantly Nodee D No B
Node
when there are many obstacles and moving people, especially Node
Node D Node B
in complex indoor environments, as indicated in [4]. Further- Searching
Searchin
ching Scopee
more, [11] has proven that in the indoor scenarios, such NLOS Unknown
Un known
n No
Node
N
error could modeled as colored noise instead of a Gaussian dis-
tribution noise. A NLOS mitigation algorithm is also proposed Node C
based on adaptive Kalman filter to eliminate such ranging Node C
error. Therefore, 𝑏𝑖𝑗 can be neglected and 𝑇𝑖𝑗 = 𝑑𝑖𝑗 + 𝑛𝑖𝑗 .
Therefore, 𝑇𝑖𝑗 follows Gaussian distribution with mean
value of 𝑑𝑖𝑗 /𝑐, and variance of 𝜎𝑇2 , i.e.:
Fig. 2. Fixed spatial structure with four assistant nodes.
𝑓 (𝑇𝑖,𝑗 = 𝑡∣𝜃) = 𝑁 (𝑡; 𝑑𝑖,𝑗 /𝑐, 𝜎𝑇2 ) (9)
i.e., 𝐸 has 𝑒 elements. Order the 𝑒 edges of 𝐺 in some way
where 𝑐 is the speed of propagation and 𝜎𝑇2 is the variance and define 𝑓𝐺 : ℝ2(𝑛+1) → ℝ𝑒 by:
of the time delay error. It is worth noting that the mean value 2
𝑓𝐺 (𝑝1 , ⋅ ⋅ ⋅ , 𝑝𝑛+1 ) = (⋅ ⋅ ⋅ , ∥𝑝𝑖 − 𝑝𝑗 ∥ , ⋅ ⋅ ⋅ ) (10)
of time delay error may be a non-zero value in practice, but
it can be estimated (as a nuisance parameter) by localization where 𝑖, 𝑗 ∈ 𝐸, 𝑝𝑘 ∈ ℝ2 for 1 ≤ 𝑘 ≤ 𝑛 + 1, and ∥⋅∥ denotes
algorithm so that it can be subtracted out and neglected here the Euclidean norm in ℝ2 .
[10]. Then we can use definition of rigid to measure the property
of graph 𝐺.
C. CRLB for estimation of fixed spatial model constructed by Definition 1: Rigid: Let 𝐺 be a graph with 𝑛 + 1 vertices,
assistant devices 𝐾 the complete graph with 𝑣 vertices, and 𝑝 ∈ ℝ2(𝑛+1) . The
graph 𝐺(𝑝) is rigid in ℝ2 if there exists a neighborhood 𝑈 of
After coarse coordinates of all the blindfolded devices 𝑝 in ℝ2(𝑛+1) such that 𝑓𝐾 −1 −1
(𝑓𝐾 (𝑝)) ∩ 𝑈 = 𝑓𝐺 (𝑓𝐺 (𝑝)) ∩ 𝑈 .
and the distance information between each two devices are According to [12] and [13], a fully connected graph is rigid.
obtained by fingerprints and TOA ranging methods, a fixed If one of the edges is randomly deleted from a rigid graph,
spatial model can be constructed. And this structure with which makes the graph no more rigid, the graph is called
additional ranging information can be used to estimate the minimum rigid graph, and we define the edges set 𝐸 =
optimal position based on coordinates obtained in the initial 𝑚𝑟𝑔(𝐺). Thus, the number of independent measurements in
localization process. Under these circumstances, we can search the spatial model mentioned above is the number of edges of
for the ultimate localization result by setting a searching minimum rigid graph constructed by the model.
scope (circle of radius R) around the initial position. Dur- Theorm 1: For a graph 𝐺 = (𝑉, 𝐸) in ℝ2 , ∣𝑉 ∣ = 𝑛 + 1, the
ing searching process, the spatial structure consisted of the necessary and sufficient condition of graph 𝐺 is a minimum
unknown and assistant nodes keeps intact. A simple example rigid graph is: (1) ∣𝐸∣ = 2𝑛 − 1, (2) for any subgraph 𝐺′ =
of the constructed spatial model with four assistant nodes is (𝑉 ′ , 𝐸 ′ ), ∣𝑉 ′ ∣ = 𝑘, ∣𝐸 ′ ∣ ≤ 2𝑘 − 3.
illustrated in Figure 2. Due to space limitations, we omit the proof details. The
In the above fixed spatial model, there are 𝑛 + 1 vertices first condition ensures that there are enough edges in 𝐺 to
(including the target node and 𝑛 assistant nodes) and 𝐶𝑛2 edges keep rigid, while the second condition ensures that the number
when all the distance information is measured between each of edges will not be too much to prevent from some edges
two nodes, where each edge can be expressed by TOA ranging repeatedly connecting the same two vertices.
measurement. Then our purpose is to estimate accuracy of this Thus, the number of independent TOA measurements in the
spatial model. Obviously, these 𝐶𝑛2 ranging measurements are spatial model is 2𝑛 − 1. And the joint probability density for
redundant for constructing the model. estimation the fixed spatial model can be described as:
We regard a graph 𝐺 as a set 𝑉 = (1, ⋅ ⋅ ⋅ , 𝑛 + 1) whose 2𝑛−1
∏ 1 1 𝑑𝑖𝑗 2
elements are called vertices of 𝐺 together with a nonempty set 𝑓 (𝑡∣𝜃) = √ exp(− 2 (𝑡 − ) ) (11)
𝐸 of two-element subsets of 𝑉 which are referred to as edges 2𝜋𝜎𝑇 2𝜎𝑇 𝑐
𝑘=1
of 𝐺. A graph 𝐺(𝑝) in the 2-D space ℝ2 is a graph 𝐺 = (𝑉, 𝐸) Then the FIM for estimation of fixed spatial model con-
together with a point 𝑝 = (𝑝1 , ⋅ ⋅ ⋅ , 𝑝𝑛+1 ) ∈ ℝ2 × ⋅ ⋅ ⋅ × ℝ2 = structed by assistant devices can be expressed as:
ℝ2(𝑛+1) . We refer to the points 𝑝𝑖 for 𝑖 ∈ 𝑉 as the vertices of
∑ (𝑥𝑘 − 𝑥𝑖 )2
𝐺(𝑝) and the line segments [𝑝𝑖 , 𝑝𝑗 ] in ℝ2 for 𝑖, 𝑗 ∈ 𝐸 as the [𝑇 ]
𝐽𝑥𝑥 (𝜃) = 𝜌[𝑇 ] (12)
edges of 𝐺(𝑝). Consider a graph 𝐺 = (𝑉, 𝐸) with 𝑒 edges, 𝑑2𝑘𝑖
(𝑘,𝑖)∈𝑚𝑟𝑔(𝐺)

Authorized licensed use limited to: ASU Library. Downloaded on December 17,2023 at 03:22:38 UTC from IEEE Xplore. Restrictions apply.
Figure 4 shows the simulation results of localization accu-
∑ (𝑥𝑘 − 𝑥𝑖 )(𝑦𝑘 − 𝑦𝑖 )
[𝑇 ] [𝑇 ] [𝑇 ] racy with the number of assistant nodes. It shows an almost
𝐽𝑥𝑦 (𝜃) = 𝐽𝑦𝑥 (𝜃) =𝜌
𝑑2𝑘𝑖 monotonic increase in average localization accuracy as the
(𝑘,𝑖)∈𝑚𝑟𝑔(𝐺)
(13) number of assistants increases. However there is almost no
any improvement in localization accuracy when the number of
∑ (𝑦𝑘 − 𝑦𝑖 )2 assistants is greater than 8. Besides, with the threshold used
[𝑇 ]
𝐽𝑦𝑦 (𝜃) = 𝜌[𝑇 ] (14)
𝑑2𝑘𝑖 when selecting the assistant nodes, it is considered that when
(𝑘,𝑖)∈𝑚𝑟𝑔(𝐺)
𝑛 = 8, the actual effect is the best.
[𝑇 ] 1
where 𝜌 = 𝑑𝑘𝑖 is the true distance between 𝑘 and
(𝑐𝜎𝑇 )2 ,
In the above experimental results, if the number of assistants
𝑖, and (𝑘, 𝑖) ∈ 𝑚𝑟𝑔(𝐺) means the edge between 𝑘 and 𝑖 is is 8, the ∣𝑚𝑟𝑔(𝐺)∣ = 15, which indicates that in the fixed
included in the minimum rigid graph of 𝐺. spatial model there are 15 edges in the minimum rigid graph
Subsequently, the CRLB for the estimation of fixed spatial of 𝐺 and there are 15 independent TOA measurements (where
model constructed by assistant devices can be expressed as, the number of vertices in spatial model is 9, composed of 8
[ ] assistant nodes and one target node).
[𝑇 ]
[𝑇 ] −1 1 𝐽𝑦𝑦 (𝜃) −𝐽 [𝑇 ] 𝑥𝑦(𝜃) From (17), it can be seen that the lower bound of local-
{𝐽 (𝜃)} =  [𝑇 ]  [𝑇 ] [𝑇 ] (15)
𝐽 (𝜃) −𝐽𝑥𝑦 (𝜃) 𝐽𝑥𝑥 (𝜃) ization error is related to 𝜎𝑇 for TOA part and 𝜎𝑑𝐵 /𝑛𝑝 for
  RSS part. 𝜎𝑇 is the variance of time delay error, and can
where 𝐽 [𝑇 ] (𝜃) = 𝐽𝑥𝑥 (𝜃) ⋅ 𝐽𝑦𝑦 (𝜃) − (𝐽𝑥𝑦 (𝜃))2 .
[𝑇 ] [𝑇 ] [𝑇 ]
be vary with the different TOA measurements device. For
From the CRLB property, we know that Cov(𝜃, ˆ 𝜃) ≥ wideband DSSS measurements, the empirical value of 𝜎𝑇 is
{𝐽(𝜃)}−1 , i.e., the matrix Cov(𝜃, ˆ 𝜃) − {𝐽(𝜃)}−1 is positive 6.1 ns, while in UWB devices is 1.9 ns. 𝜎𝑑𝐵 stands for the
semidefinite. Sine the diagonal elements of positive semidefi- standard deviation of received power and is relatively constant
nite matrices are larger or equal to zero, we have, with distance. Typically 𝜎𝑑𝐵 is as low as 4 and as high as
[𝑇 ] [𝑇 ]
12 [10]. The path-loss exponent 𝑛𝑝 represents complexity
ˆ ≥ tr{𝐽 [𝑇 ] (𝜃)}−1 = 𝐽𝑥𝑥 (𝜃) + 𝐽𝑦𝑦 (𝜃) of propagation environment, and typically between 2 and 5,
var(𝜃)   (16)
𝐽 [𝑇 ] (𝜃) which 2 means free space environment with no obstacles.
We also perform some simulations to show the localization
D. CRLB for fingerprint and assistant nodes based localiza-
accuracy with the propagation parameters, 𝜎𝑇 and 𝜎𝑑𝐵 /𝑛𝑝 .
tion
The results are shown in Figure 5 and Figure 6, respectively.
Since the respective errors based on the RSS fingerprint
algorithm and the TOF ranging between all the assistant nodes V. C ONCLUSION
based calculation are independent to each other, CRLB for the In this paper, we studied the CRLB of location based on the
hybrid algorithms can be calculated separately, and then sum collaboration of RSS and assistant nodes. We obtain the total
together [14]. That is, 𝐶𝑅𝐿𝐵 = 𝐶𝑅𝐿𝐵𝑅 + 𝐶𝑅𝐿𝐵𝑇 CRLB for the hybrid RSS/ TOA measurements with mitigating
According to the above equations, the final expression of the NLOS ranging error based on the collaboration of RSS and
estimation variance for the proposed scheme can be described assistant nodes. Subsequently, we utilize the theory of rigid
in (17). graph to analyze the estimation error of the fixed spatial model
10𝑛 𝑑 𝑑𝑖𝑗
where 𝑏 = ( 𝜎𝑑𝐵 ln𝑝10 )2 , 𝐴𝑖𝑗 = 0,(𝑖𝑗)
𝑑𝑖0 𝑑𝑗0 a non-dimensional to obtain the complete CRLB expression. Besides, we also
parameter, representing the geometric conditioning of devices explore characterization of the properties of this lower bound
𝑖 and 𝑗 with respect to target node, and 𝑑0,(𝑖𝑗) is the length of to evaluate the impact of different parameters on the accuracy
shortest distance between target node and the edge connecting of coordinates estimation, which can provide suggestions for
devices 𝑖 and 𝑗. Also 𝐴𝑖,𝑗 can be considered as the distance a positioning system design and deployment.
between each two devices of target node, 𝑖 and 𝑗, normalized
by 𝑑𝑖0 𝑑𝑗0 . ACKNOWLEDGMENT
This research is supported in part by grants from the
IV. P ERFORMANCE EVALUATION OF FINGERPRINT AND
National Natural Science Foundation of China (61301114,
ASSISTANT NODES BASED LOCALIZATION
51304058).
From (17), it is evident that the CRLB depends on the
number of APs, the geometry of the AP, the number of R EFERENCES
assistants nodes, the shape of the constructed spatial model, [1] S. He and S. H. G. Chan, “Wi-fi fingerprint-based indoor positioning:
propagation model parameters. Next, we elaborately discuss Recent advances and comparisons,” IEEE Communications Surveys
the properties of these parameters on localization error bound. Tutorials, vol. 18, no. 1, pp. 466–490, Firstquarter 2016.
[2] A. M. Hossain and W.-S. Soh, “A survey of calibration-free indoor
Using these findings, we provide insights into positioning positioning systems,” Computer Communications, vol. 66, no. C, pp.
system deployment issues and the criteria of assistant nodes 1–13, Jul. 2015.
selection. [3] H. Liu, J. Yang, S. Sidhom, Y. Wang, Y. Chen, and F. Ye, “Accurate
wifi based localization for smartphones using peer assistance,” IEEE
During the simulations, we use the same indoor environment Transactions on Mobile Computing, vol. 13, no. 10, pp. 2199–2214,
as depicted in [4], as illustrated in Figure 3. Oct 2014.

Authorized licensed use limited to: ASU Library. Downloaded on December 17,2023 at 03:22:38 UTC from IEEE Xplore. Restrictions apply.
(2𝑛−3)𝑚 ∑𝑚
(𝑐𝜎𝑇 )2 + 𝑏 𝑖=1 𝑑−2
𝑖0
ˆ ≥
var(𝜃)
1
∑𝑚 ∑𝑚+𝑛 ∑ 𝑏
∑𝑚 ∑𝑚+𝑛 ∑ 𝐴2𝑖𝑗 ∑𝑚 ∑𝑚+𝑛 𝐴2𝑖𝑗
(𝑐𝜎𝑇 )4 𝑖=1 𝑗=𝑚+1 (𝑖,𝑗)∈𝑚𝑟𝑔(𝐺) 𝐴2𝑖𝑗 + (𝑐𝜎𝑇 )2 𝑖=1 𝑗=𝑚+1 (𝑖,𝑗)∈𝑚𝑟𝑔(𝐺) 𝑑2𝑗0 + 𝑏2 𝑖=1 𝑗=𝑚+1 𝑑𝑖0 𝑑𝑗0
(17)

Window 1 A1 Window Window

Cumulative Probability
AP3
AP3 0.8
A22
A

Experiment Table
σdB/n p=1
0.6
AP1
AP1 Desk σdB/n p=2
0.4 σdB/n p=3
σdB/n p=4
AP4
AP
AP44 0.2 σdB/n p=5
AP2
A3
0
A4 0 1 2 3 4 5 6 7 8 9 10
Localization Error (m)

Experiment Table
Fig. 6. Localization accuracy with 𝜎𝑑𝐵 /𝑛𝑝 .
Door Door

Desk Chair RP AP Unknown Node Assistant Nodes


[4] Q. Li, W. Li, W. Sun, J. Li, and Z. Liu, “Fingerprint and assistant nodes
based wi-fi localization in complex indoor environment,” IEEE Access,
vol. 4, pp. 2993–3004, 2016.
[5] H. L. Van Trees, Parameter Estimation I: Maximum Likelihood. John
Wiley Sons, Inc., 2002.
Fig. 3. Deployment of simulation environment. [6] A. K. M. M. Hossain and W. S. Soh, “Cramer-rao bound analysis of
localization using signal strength difference as location fingerprint,” in
2010 Proceedings IEEE INFOCOM, March 2010, pp. 1–9.
1 [7] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
0.9 Theory. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1993.
0.8 [8] J. Y. Huang and Q. Wan, “The crlb for wsns location estimation in nlos
Cumulative Probability

0.7 n=10 environments,” in 2010 International Conference on Communications,


n=9 Circuits and Systems (ICCCAS), July 2010, pp. 83–86.
0.6 n=8
0.5 n=7
[9] A. J. Coulson, A. G. Williamson, and R. G. Vaughan, “A statistical
n=6 basis for lognormal shadowing effects in multipath fading channels,”
0.4
n=5 IEEE Transactions on Communications, vol. 46, no. 4, pp. 494–502,
0.3 n=4
n=3
Apr 1998.
0.2
[10] N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero, R. L. Moses, and
0.1 N. S. Correal, “Locating the nodes: cooperative localization in wireless
0
0 1 2 3 4 5 6 7 8 9 10
sensor networks,” IEEE Signal Processing Magazine, vol. 22, no. 4, pp.
Localization Error (m) 54–69, July 2005.
[11] L. Qiyue, W. Zhong, L. Jie, S. Wei, and W. Jianping, “A novel
adaptive kalman filter based nlos error mitigation algorithm,” IFAC-
PapersOnLine, vol. 48, no. 28, pp. 1118 – 1123, 2015.
Fig. 4. Localization accuracy with the number of assistant nodes. [12] G. Laman, “On graphs and rigidity of plane skeletal structures,” Journal
of Engineering Mathematics, vol. 4, no. 4, pp. 331–340, 1970.
[13] B. Servatius and H. Servatius, “Rigidity, global rigidity, and graph
1 decomposition,” European Journal of Combinatorics, vol. 31, no. 4, pp.
0.9 1121 – 1135, 2010, rigidity and Related Topics in Geometry.
Culumative Probability

0.8 σT=1 ns [14] R. Kumarasiri, K. Alshamaileh, N. H. Tran, and V. Devabhaktuni,


0.7 σT=2 ns “An improved hybrid rss/tdoa wireless sensors localization technique
0.6 σT=3 ns utilizing wi-fi networks,” Mobile Networks and Applications, vol. 21,
0.5 no. 2, pp. 286–295, 2016.
σT=4 ns
0.4
0.3 σT=5 ns
0.2 σT=6 ns
0.1 σT=7 ns
0
0 1 2 3 4 5 6 7 8 9 10
Localization Error (m)

Fig. 5. Localization accuracy with 𝜎𝑇 .

Authorized licensed use limited to: ASU Library. Downloaded on December 17,2023 at 03:22:38 UTC from IEEE Xplore. Restrictions apply.

You might also like