Académique Documents
Professionnel Documents
Culture Documents
Abstract—Nowadays, the popular adoption of smart phones RSSI measurements RSSI measurements
has enabled the development of several location-based applica- Database
at known positions
tions such as indoor localization for shopping centers, airports,
hospitals, among others. Typically, these services are based on
Wireless Local Area Networks (WLANs), where the selection of
RSSI measurements
access points (APs) plays an important role for the localization RSSI measurements
Matching
at unknown locations
accuracy. In this paper, we evaluate different criteria of AP
selection as well as the number of required APs, the impact of
AP power transmission and the usage of different frequency
bands. The results are obtained by means of experimental
User localization
evaluations that provide more realistic comparisons. Finally,
we provide novel guidelines for the design of improved indoor
positioning systems. Fig. 1. Fingerprint matching scheme where RSS measurements are stored
at the database, and then compared to the online measurements to estimate
I. I NTRODUCTION the user position.
Building A 6.3 m
1m
it is compared against a database of fingerprint, so that
the user position is estimated. Let us representing this
collection as a vector composed of RSS measurements 8.9 m
r(z) = {r1 (z), r2 (z), r3 (z), ..., rM (z)} at the position z =
(x, y), where rk (z) refers to the measured RSS due to the Fig. 3. The established grid for collecting the measurements at each room.
kth access point APk , and each of these measurements rk Typically a grid resolution of 1.2m x 1m is considered.
follows a shadowing model [15] given by,
d
rk = tk + 10 log10 (K) − 10γ log10 ( ) − ψdB (1) measurements at each position are quite different. In fact,
do
measurements from APs that provide similar features within
where tk denotes the transmitted power in dBm, K is a
an indoor environments should be avoided [7]. In this work,
unitless constant that depends on the antenna characteristics
we evaluate different AP selection criteria which are applied
and the average channel attenuation, γ is the path loss
to both the fingerprint database, as well as to the online data.
exponent, d and do are the distance to the target location and
The following criteria are considered:
reference distance for the antenna far-field, respectively, and
ψdB is a random variable following a Gaussian distribution, • Energy: The APs are selected in a descending order
N (0, σk2 ). according to the average energy of the measurements
rk .
The estimation of the user location is obtained by • Maximum RSS: this criterion selects APs in a descend-
measuring the similarity between the current measurement ing order according to the maximum RSS value.
and the stored one. An estimator for the position ẑ is • Entropy: APs are sorted in descending order ac-
obtained by minimizing the error E = ẑ − z for a given cording to the calculated entropy, i.e. H(rk ) =
cost function C(E), i.e. minJ (E) = min E {C(E)}. From a
P
s∈ψ P (s)log(P (s)), where ψ is the alphabet of pos-
Bayesian point of view, the optimal estimator
R is given by sible values for the random variable rk , and P (s) the
the mean of the posterior P(z/r), i.e. ẑ = zP(z/r)dz. probability mass function.
For a more practical approach, the maximum likelihood is • RSS variance: with this criterion, APs with larger
commonly employed, which is given by, variance in their corresponding RSS measurements are
"M # selected. It follows a Gram–Schmidt orthogonalization
1 X −(rk (z) − µk (z))2
P(r/z) = QM exp . and is given by,
2
k=1 (2πσk (z))
1/2 2σk2 (z) 1: procedure RSS variance (G)
k=1
(2) 2: S ← set of APs
where µk (z) refers to the received signal attenuated by 3: Vc ← constant
the free-space path loss (i.e. µk (z) = tk + 10 log10 (K) − 4: for each index in the set S do
Vr
10γ log10 ( ddo )), and σk2 (z) the variance of the received 5: Vr⊥k ← Vrk − P royVc k
signal rk (z). Nevertheless, the variance estimates requires 6: end for
collecting a large amount of data, and the uncertainty of its 7: I ← argmax
Vr⊥k
2
k
estimation is increased as few measurements are available. 8: V c ← V rk
For this reason, it is replaced by a constant. Thus, the max- 9: S ←S\k
imum likelihood reduces to the minimization of euclidean 10: if S is empty then
distance between the current measurement rk (z) and the 11: return I
stored one at the database, µk (z)1 . 12: else
13: go to 4
III. AP SELECTION CRITERIA
14: end if
Ideally, it is desirable to select APs that provide useful 15: end procedure
information for the user localization, i.e. where the collected where G = {Vr1 , Vr2 , ..., VrM }, Vrk is a vector com-
1 It is usually obtained by averaging several times the received signal posed by measurements of the kth AP which are
rk (z), so that the shadowing effect can be mitigated. collected within one room, and Vc a constant vector.
Criteria Energy Max. RSS Entropy Variance Optimal Criteria Energy Max. RSS Entropy Variance
Mean 1.97 1.97 1.97 1.97 1.48 Mean 2.6 2.76 2.56 2.27
error error
(m) (m)
TABLE I
T HE OBTAINED ACCURACY FOR EACH SELECTION METHOD ALONG
WITH THE REQUIRED NUMBER OF AP S Criteria Energy Max. RSS Entropy
• Optimal: in this case, the APs with the best localization 5 APs 1.93 2.01 1.98
results are obtained by brute force.
4 APs 2.17 2.16 2.06
IV. E XPERIMENTAL EVALUATION
The experiments were conducted at the buildings of TABLE III
the Pontifical Catholic University of Peru, where RSS T HE OBTAINED RESULTS WHILE OPERATING IN THE 5 GH Z BAND OF
measurements were collected in geographical points that FREQUENCY, AND FOR DIFFERENT NUMBERS OF AP S