Seamless indoor/outdoor location cognition with confidence in wireless systems
This paper proposes a method of real-time seamless location cognition in WLAN systems based on monitoring, learning and recognizing the statistics of received data traffic. It uses the property that statistics such as average and variance of throughput are often correlated with the location. The method also uses a confidence parameter to quantify the confidence in the location identification, based on comparing matches with multiple candidates. The proposed method can support seamless location cognition by WLAN terminals, indoor and outdoor, without depending on any infrastructure service.
Seamless Indoor/Outdoor Location Cognition with
Confidence in Wireless Systems
S. Aust, T. Ito
NEC Communication Systems, Ltd.
Abstract—This paper proposes a method of real-time seamless
location cognition in WLAN systems based on monitoring,
learning and recognizing the statistics of received data traffic. It
uses the property that statistics such as average and variance of
throughput are often correlated with the location. The method
also uses a confidence parameter to quantify the confidence in the
location identification, based on comparing matches with
multiple candidates. The proposed method can support seamless
location cognition by WLAN terminals, indoor and outdoor,
without depending on any infrastructure service.
Keywords-Location; WLAN; distance; cognition; confidence
Location cognition is a key function for cognitive radio
(CR) systems. The use of radio spectrum resources highly
depends on the location. The Federal Communications
Commission (FCC) has regulated that TV white space CRs
need to be aware of their geo-location and the coverage area of
TV broadcast services . GPS has been the standard service
for positioning of CRs with some extensions to loosen
requirements . Location information is also important for
navigation or detecting movements of radio transmitters .
GPS is widely used for position measurement outdoors, but its
use indoors is limited . Other methods are used to
complement GPS for providing position information – for
example, beacons, triangulation or fingerprints from
surrounding access points (APs).
Comparing with outdoor, indoor location is usually more
difficult due to the complexity of the effect of the surrounding
physical environment on the propagation of radio waves. A
recent survey in  describes indoor positioning systems (IPS)
for wireless networks, addressing security, privacy, cost,
performance, robustness, complexity and limitations. The
authors classified techniques into four types, namely
triangulation, fingerprinting, proximity and vision analysis.
Fingerprinting is widely applied for indoor positioning ,
. Fingerprinting typically uses location-related data
(“fingerprints”) obtained in an off-line training phase. An
enhanced fingerprint method based on triangulation of radio
signal strength (RSS) obtained in real-time was proposed in .
The method estimates the propagation model that best-fit the
propagation environment. However, a vast amount of
measurements is needed to obtain accurate values for the
proposed log-normal path-loss model.
We propose a novel approach for location cognition which
is independent of infrastructure such as GPS, APs or any
centralized database (DB) containing fingerprints. We propose
a location cognition method based on the monitoring, learning
and recognizing the statistical characteristics of data
transmission between radio terminals. The proposed method
can support seamless location cognition by WLAN terminals,
indoor and outdoor. Our proposal combines learning and
classification of statistical distributions, with a measure of the
confidence of classification results – to realize an advanced
cognition function. Knowing the confidence contributes
significantly to the intelligence of a cognitive radio, allowing it
to make more reliable decisions, for example in regard to
switching channels according to location.
OCATION COGNITION ENGINE (LOC)
The idea underlying our proposal is that locations have
transmission characteristics which can be learnt and recognized.
We propose to monitor the data transmission at the receiver
terminal to obtain statistics which are used to classify the
location of the data transmission. We consider a two stage
procedure, whereby data is collected and templates extracted in
a learning stage, and templates are used to classify the location
in the location recognition stage. The difference from other
fingerprinting methods is that the fingerprints are identified in
an autonomous way and based on high-level transmission
characteristics, rather than physical layer information such as
RSS or delay. We also introduce novel system parameters
indicating the confidence in the best-fit.
In  we argued that monitored traffic of constant data
transmission show characteristic distributions and can be used
to calculate statistics such as mean and standard deviation to
obtain specific distributions of received data at each different
location. In particular, the received number of data packets has
been used as monitoring parameter. Further, we proposed in
 the use of the Jeffrey Divergence (JD) with Gaussian
approximation to calculate statistical dissimilarities.
Fig. 1 shows the LOC comparing an input test distribution
with a set of reference distributions (RD). The test distribution
is obtained in real-time from the transmission performance
4th IEEE Workshop On User MObility and VEhicular Networks On-MOVE 2010, Denver, Colorado
978-1-4244-8389-1/10/$26.00 ©2010 IEEE 653
samples using a sliding window and an outlier-filter, and
removal of transient states. The LOC calculates the JDs for
each of the RDs. The RD with the minimal JD, which indicates
least dissimilarity between the RD and test data, is selected as
best fit. A measure of the “goodness of fit”, which we call the
“dynamic confidence̍, is then calculated as explained next in
Figure 1. Proposed Location Cognition (LOC)
III. ESTIMATION OF LOCATION CONFIDENCE
In addition to identifying a location by choosing the best-fit,
we introduce new confidence parameters describing the
reliability of the location classification. The motivation for this
is that the “best-fit” result might not be reliable in all
situations, either because the training was not sufficient or the
monitored data was poor, or the location does not correspond
to a known location.
A. Dynamic Confidence Algorithm
A measure of the reliability of the “best-fit” identification
of a location, which we call the Dynamic Confidence (dc) is
obtained by comparing the JD of the all the reference
distributions with that of the best-fit.
Calculating dc for sample k is done as follows:
is the divergence with the candidate distribution i, and
is the minimum divergence corresponding to the best-fit.
The value of is set to = 0.5 for optimal effect. The result is
a value with a range of [0-100]. A high confidence value close
to 100 indicates large distances (large dissimilarities) between
best-fit and other templates. A low confidence value close to 0
indicates a small distances (small dissimilarities).
In the example of Fig. 1 neighbor (B) has the smallest
distance a to (A) and its distance contributes to the confidence.
If the distance a is close to the divergence of (A) the
confidence will be low. If a is larger than the divergence of (A)
the confidence increases. Neighbor (C) has the distance b to
(A) which is larger compared to (B).
B. Static Confidence Algorithm
The LOC has an intrinsic classification ability that depends
on the set of RDs selected during the off-line phase. We
introduce a measure of this ability called the Static Confidence
(sc) of the LOC.
For the calculation of sc the following algorithm was
l is the number of true classifications at a location l
l is the number of false location classifications Fl, and
bestFit is the RD which maximizes sc with the maximum
number of true classifications T
l and minimal number of false
location classifications F
The static confidence sc has value in the range of [0-1]. sc
higher than 0.8 indicates to the user a high confidence of the
location cognition. The value sc is static and does not change
until a new set of RD
i is acquired. A low sc value may trigger a
new RD selection process including new reference
MPLEMENTATION AND TESTING OF A LOC PROTOTYPE
In this section we discuss the implementation and testing of
a prototype of the LOC for WLAN terminals.
A. Implementation and Test Site
The LOC engine was implemented as wireless middleware
on a Linux platform using WLAN IEEE 802.11g. The LOC
uses as monitoring parameter, the number of received data
packets, TxCount. We have also implemented an outlier filter
for data pre-processing. For the test, we considered a scenario
where location cognition would be useful for implementing
local WLAN channel usage policies in and around a building.
Transmission data was obtained by transmitting data between a
pair of WLAN terminals at 5 different locations under various
conditions. Figure 2 shows a map of the test site, including the
indoor floor plan.
The following five key locations were used for testing:
x Indoor locations
1. lab: office-style research lab (size: 10m x 20m x 3m).
2. corridor: inside the building, consisting of doors,
walls and windows (size: 3m x 30m x 3m).
3. entrance: an entrance hall in the building mainly
consisting of glass windows (size: 25m x 25m x 10m).
x Outdoor locations
4. building: a location 5m beside the building.
5. road: a location at the road 50m far from the building.
Figure 2. Map of the test site
Transmission between sender and receiver was line-of-sight
(LOS) in each case. At each location, transmission data was
obtained for multiple distances at 5 m, 10m and 15m. Data was
obtained for both short packets (200 bytes) and long packets
(1500 bytes) capturing the effect of different packet sizes. A
single UDP flow was sent continuously for the duration of 300s.
Transmission was monitored at the receiver in intervals of 1s at
maximum throughput obtaining reference distributions.
The procedure for acquisition of data during the off-line
phase, selection of the reference distributions and then on-line
location cognition are described below.
B. Selection of Reference Distributions
The mean and deviation of the reference distributions are
shown in Fig. 3 and Fig. 4, respectively. The top graphs show
the results for short packet obtaining max throughput at 7 Mbps
before saturation occurred. The bottom graphs show the results
for long packets at 28 Mbps before saturation occurred.
Note that for all RDs the mean shows a higher number of
successfully received data packets for outdoor locations,
indicating a better quality link with less interference and multi-
path effects. Regarding indoor locations, the RDs for the
“entrance” show the highest number of received data packets,
followed by the RDs for “lab” and “corridor” which show the
lowest number of received data packets. An explanation is that
for these locations indoor wireless activities lead to a reduced
number of successfully received data packets. However, due to
the exposed multi-path environment in the “entrance” the
number of received packets is less than in “outdoor” location.
Multi-path fading and increased wireless activities result in
lower throughput in particular for the “lab” and the “corridor”
as shown in Fig.3. The characteristics for receiving data of long
data packets are similar for short packets, except that the
maximum value for outdoor is reduced for long data packets.
Fig. 4 shows the standard deviation of all locations. It
shows the highest value for the location “lab” for short packets
and decreases at the locations “corridor” and “entrance”. For
the outdoor locations the standard deviation decreases
significantly, were the location “building” shows a higher
standard deviation than “road”. Multi-path fading at the long
side of the building is a possible cause for the large deviation
there. For the location “corridor” lower standard deviation can
be observed for short distances.
For long packet length (1500 byte) the results show a
different characteristic. The standard deviation is significantly
reduced for all locations, expect for the location “lab”. Our
explanation for this is that it is due to the significantly larger
background wireless activity inside the location “lab”.
C. Location Cognition Tests
Two types of location cognition were tested, generic
location cognition and an extended location cognition which
identifies the distance between the transmitter and receiver.
The generic location cognition uses a generic reference
distribution (RD) with the maximum of static confidence,
whereas the extended location cognition uses an extended RD
set corresponding to different distances, specifically, distances
5m, 10m and 15m. The reason for the test with the extended
location cognition is to get additional information regarding the
resolution of the location identification. Table I lists the sc
results for short packets (200 byte) and long packets (1500
During an evaluation phase of the experiment, the
correctness of the location cognition was evaluated. In order to
obtain quantitative results, tests were run based on 3 trials,
considering short packet length, long packet length, 3 different
distances (5m, 10m, 15m) and 5 different locations (lab,
corridor, entrance, building, road) and a window size (ws) at 30.
We evaluated the correctness of the location cognition by
counting the number of correct classifications (true positive)
and incorrect classification (false positive).
Figure 3. Mean for short packets (top graph) and long packets (bottom
Figure 4. Standard deviation for short packet length (top graph) and long
packets (bottom graph)
SC RESULTS FOR 5 LOCATIONS (L=5) AND 5 LOCATIONS AT 3
DISTANCES (L=15) AUTONOMOUSLY SELECTED BY THE LOC
Generic RD set
Full RD set
200 byte 1.00 0.80
1500 byte 0.80 0.73
D. Results for Identification of Location
The result for short packets showed a true positive ratio
(TPR) TPR=0.99 and a false positive ratio (FPR) FPR=0.01
which indicates a perfect classification of location for short
packets (graphs excluded for reasons of brevity).
We show the receiver operating characteristic (ROC) graph
for location cognition results of long packets in Fig. 5. The
graph shows that the location “lab” and “corridor” both have a
TPR=1.0 and FPR=0.0 which indicates a perfect classification.
The location “entrance” a shows a reduced performance with
TPR=1.0 and FPR=0.17 which indicates a correct location
cognition at this location, but with an increased false rate at
other locations. The location “building” shows a similar
characteristic with TPR=1.0 and FPR=0.14.
In Fig. 5 the location “road” shows a reduced TPR=0.78
and FPR=0.2. All false results at “road” were “building”,
which means that the LOC succeeded in recognizing that the
location was outdoor. We conclude that the LOC shows a high
cognition performance compared to a random guess.
E. Results for Identification of Distance
Next, we discuss the correctness of the distance cognition
between sender and receiver for long packets at all locations
(short packet graphs excluded for reasons of brevity). Figure 6
shows the distance cognition at 5m, 10m, and 15m. It can be
observed that distance cognition for 5m was at TPR=0.86 and
FPR=0.66. Distance cognition was TPR=1.0 and FPR=0.4 at
10m, and TPR=0.8 and FPR=0.6 at 15m. Fig. 6 shows the
improvement for the distance cognition comparing ws=30 and
ws=100. Distance cognition at 5m can be improved to
TPR=0.93, FPR=0.33 and to TPR=0.8, FPR=0.4 at 15m. We
confirm that distance cognition can be optimized by using
larger window sizes.
F. Dynamic Confidence Results
Fig. 7 shows dc results obtained for location cognition and
location cognition including distance cognition during a
seamless location transition from indoor to outdoor. Sender and
receiver jointly moved during the transition while the LOC
estimated the location and dynamic confidence.
Figure 5. ROC graph of location cognition results
Figure 6. ROC graph of distance cognition
The dc for the location cognition in the entrance location is
100%. The switch in location classification occurs at 48s. The
intermediate points during location changes are indicated by
lower confidence. The dc for extended cognition of distance in
addition to location is lower and affected earlier by the location
Figure 7. dc during a seamless location transition
We proposed a method for seamless location cognition
indoor and outdoor, which does not rely on any AP
infrastructure. The method is based on learning and
recognizing the location dependent differences in wireless
transmission characteristics. We implemented a location
cognition engine (LOC) and demonstrated that it is able to
recognize indoor and outdoor locations and distance between
sender and receiver. Finally, we proposed and demonstrated the
use of confidence with LOC.
Our results show that location cognition is possible by
observing the received number of data packets which is
different from other location techniques that use RSS. We also
demonstrated that by using novel parameters such as static
confidence and dynamic confidence, the system behaviour of
our proposed LOC can be easily verified. In particular, the
dynamic confidence has the potential for classifying the
location cognition results by a cognitive engine.
The proposed LOC is highly versatile and can be applied
for indoor/outdoor environments. All performance tests have
been conducted in real WLAN environments including typical
dynamics such as changes of wireless activities or surrounding
environments. It can be concluded that the proposed method
for seamless indoor/outdoor location cognition has been tested
successfully and has the potential to be an integral part of
intelligent WLAN systems in the future.
 A. Arslan, “Cognitive Radio, Software Defined Radio, and Adaptive
Wireless Systems”, Springer 2007.
 S. J. Shellhammer, A. K. Sadek, W. Zhang, ”Technical Challenges for
Cognitive Radio in the TV White Space Spectrum”, Information Theory
and Applications Workshop, February 2009.
 C. Fischer and H. Gellersen, “Location and Navigation Support for
Emergency Responders: A Survey”, Pervasive Computing, March 2010.
 Indoor Galileo/GPS Indoor Navigation & Positionierung (in German),
funded by the Deutsches Zentrum fuer Luft & Raumfahrt (DLR) and the
German Ministry, www.indoor-navigation.de, on-line, 2010.
 Y. Gu, A. Lo, I. Niemegeers, “A Survey of Indoor Positioning Systems
for Wireless Personal Networks”, IEEE Communications Surveys &
Tutorials, Vol. 11, No. 1, First Quarter 2009.
 A. Papliateseyeu, N. Kotilainen, O. Mayora, V. Osmani, “FINDR: Low-
cost Indoor Positioning Using FM Radio”, Lecture Notes of the Institute
for Computer Sciences, Social Informatics and Telecommunications
Engineering, Vol. 7, Springer Berlin Heidelberg, 2009.
 M.B. Kjaergaard, G. Treu, P. Ruppel, A. Kuepper, “Efficient Indoor
Proximity and Separation Detection for Location Fingerprinting”, in the
Proc. of Mobileware 2008, Innsbruck, Austria, February 13-15, 2008.
 V. Honkavirta, T. Perala, S. Ali-Loevtty, R. Piche, “A Comperative
Survey of WLAN Location Fingerprinting Methods”, in the Proc. of the
6th Workshop on Positioning, Navigation and Communication, 2009.
 S. Mazuelas, A. Bahillo, R. M. Lorenzo, P. Fernandez, F. A. Lago, E.
Garcia, J. Blas, E. J. Abril, “Robust Indoor Positioning Provided by
Real-Time RSSI Values in Unmodified WLANs”, IEEE Journal of
Selected Topics in Signal Processing, Vol. 3, No. 5, October 2009.
 S. Aust, A. Matsumoto, T. Ito, P. Davis, “Supervised Classification
Using Jeffrey Divergence for Location Cognition in Cognitive Radio”,
in the Proc. of the 12th International Symposium on Wireless Personal
Multimedia Communications, Sendai, Japan, September 7-10, 2009.