Crowdsourced Indoor Localization for Diverse Devices through ...

Introduction

Crowdsourced Indoor Localization for Diverse Devices through Radiomap Fusion

- Motivation

Crowdsourcing for Device Diversity

Christos Laoudias∗ , Demetrios Zeinalipour-Yazti† and Christos Panayiotou∗

- RSS Fingerprints - DIFF Fingerprints - SSD Fingerprints

Simulation Results - Simulation Setup - Varying number of APs - Varying noise - Varying number of devices

∗

KIOS Research Center for Intelligent Systems and Networks, University of Cyprus † Department of Computer Science, University of Cyprus

Experimental Validation - Measurement Setup

Conclusions - Concluding Remarks

Supported by the Cyprus Research Promotion Foundation under Grant TΠE/OPIZO/0609(BE)/06 International Conference on Indoor Positioning and Indoor Navigation, Montb´ eliard - Belfort, France 28 October 2013

Outline

Introduction - Motivation

Introduction

Crowdsourcing for Device Diversity - RSS Fingerprints - DIFF Fingerprints - SSD Fingerprints



Crowdsourcing for Device Diversity Simulation Results Experimental Validation Conclusions


International Conference on Indoor Positioning and Indoor Navigation, Montb´ eliard - Belfort, France

28 October 2013

Motivation of our work I Introduction - Motivation


Traditional RSS radiomap construction I I I I

Laborious: Collectors need to visit several locations Time consuming: A large volume of data is required Short-lived: Radiomap becomes obsolete with time Expensive: Cost can be prohibitive when the task is undertaken by trained professionals





28 October 2013

Motivation of our work I

I

Introduction

I

- Motivation

I


I



Experimental Validation

Traditional RSS radiomap construction

I

Laborious: Collectors need to visit several locations Time consuming: A large volume of data is required Short-lived: Radiomap becomes obsolete with time Expensive: Cost can be prohibitive when the task is undertaken by trained professionals

Crowdsourcing comes to the rescue I

I

Volunteers are collecting location dependent RSS samples, which they later contribute to the system Crowdsourced systems (e.g., Active Campus, Place Lab, Redpin, WiFiSLAM, Molé, Elekspot, FreeLoc)

- Measurement Setup



28 October 2013

Motivation of our work I

I

Introduction

I

- Motivation

I


I


Simulation Results

Traditional RSS radiomap construction

I

- Simulation Setup - Varying number of APs - Varying noise - Varying number of devices

Crowdsourcing comes to the rescue I

I



I

Laborious: Collectors need to visit several locations Time consuming: A large volume of data is required Short-lived: Radiomap becomes obsolete with time Expensive: Cost can be prohibitive when the task is undertaken by trained professionals Volunteers are collecting location dependent RSS samples, which they later contribute to the system Crowdsourced systems (e.g., Active Campus, Place Lab, Redpin, WiFiSLAM, Molé, Elekspot, FreeLoc)

Or maybe not? I I I I

Filtering incorrect contributions (aka polluted data) Handling non-uniform fingerprint distribution Managing the increasing radiomap size Copying with heterogeneous mobile devices


28 October 2013

Crowdsourcing with RSS Fingerprints Offline (training) phase Introduction - Motivation

I

Reference locations {L : ì = (xi , yi ), i = 1, . . . , l}, n APs

I

Device D (m) visits {L(m) : ì = (xi , yi ), i = 1, . . . , l (m) }, SM where m = 1, . . . , M, L(m) ⊆ L and L = m=1 L(m)

I

Reference fingerprint ri = [ri1 , . . . , rin ]T collected at ì is used to create the device-specific radiomap R(m) ∈ Z− l (m) ×n

I

Crowdsourced radiomap R ∈ Z− l×n



(m)

rij =


(m)

(m)

Mi 1 X (m) r , 1 ≤ Mi ≤ M Mi m=1 ij

(1)

- Measurement Setup


Online (localization) phase Use R and the new fingerprint s = [s1 , . . . , sn ]T measured at 0 the unknown location ` by the user carried device D (m ) 2 b = arg min` d 2 , d 2 = Pn I `(s) i i i j=1 rij − sj I


28 October 2013

Localization with DIFF Fingerprints Radio propagation model RSS[dBm] = A − 10γ log10 d + X , X ∼ N (0, σ 2 )






(2)

DIFF approach1 Takes the difference between all pairwise AP combinations I The new fingerprints contain n = n(n−1) RSS differences 2 2 I

I

˜ contains ˜ri = [˜ri12 , . . . , ˜ri(n−1)n ]T Crowdsourced radiomap R where ˜rijk = rij − rik , 1 ≤ j < k ≤ n

˜s = [˜s12 , . . . , ˜s(n−1)n ]T where ˜sjk = sj − sk , 1 ≤ j < k ≤ n b s) = arg min` d˜ 2 , d˜ 2 = Pn Pk−1 ˜rijk − ˜sjk 2 I `(˜ i i i k=2 j=1

I

I

Higher dimensionality leads to increased computations

1 F. Dong, et al., A calibration-free localization solution for handling signal strength variance, in MELT, 2009. International Conference on Indoor Positioning and Indoor Navigation, Montb´ eliard - Belfort, France

28 October 2013

Localization with SSD Fingerprints SSD approach2 Introduction - Motivation

I

Subtracts the RSS value of an anchor AP from the other RSS values in the original fingerprint

I

The new fingerprints contain n − 1 independent RSS differences

I

ˇ contains ˇri = [ˇri1 , . . . , ˇri(n−1) ]T Crowdsourced radiomap R where ˇrij = rij − rik , j = 1, . . . , n, j 6= k




ˇs = [ˇs1 , . . . , ˇsn−1 ]T where ˇsj = sj − sk , j = 1, . . . , n, j 6= k b s) = arg min` dˇ 2 , dˇ 2 = Pnj=1 ˇrij − ˇsj 2 I `(ˇ i i i I

j6=k


I

Lower dimensionality leads to higher localization errors

2 A. Mahtab Hossain, et al., SSD: a robust RF location fingerprint addressing mobile devices’ heterogeneity, in IEEE Transactions on Mobile Computing, 2013. International Conference on Indoor Positioning and Indoor Navigation, Montb´ eliard - Belfort, France

28 October 2013

Simulation Setup AP4

AP15

AP7

AP11

AP2

AP12

AP14

AP8

AP6

AP16

AP10





AP1

AP5

AP13

AP3

(1)

I

Radiomap R(1) contains RSS values rij generated by the propagation model of (2) with A = −22.7 dBm, γ = 3.3

I

Radiomap R(m) contains RSS values such that (m) (1) rij = α1m rij + β1m , m = 2, . . . , M,

I

All M devices contribute their radiomaps R(m) to get the ˜ and R ˇ crowdsourced RSS radiomap R according to (1), R

I

User carries D (1) and may reside at any location

I

Probability of correct location estimation Pc =

- Measurement Setup


AP9


Nc Ns 28 October 2013

Varying number of APs 1

0.9

Probability of correct location estimation [Pc]

0.8



Simulation Results

0.6

0.5

0.4

0.3

0.2



0.7

RSS DIFF SSD DS

0.1

0

4

6

Pc for localizing device D

8 10 Number of access points [n]

(1)

12

14

16

with M = 2 devices and σ = 3 dBm

- Measurement Setup


I

DIFF is better than SSD and performs equally well with DS

I

RSS usually performs poorly, e.g., Pc = 0.45 for n = 6 APs

I

For a large number of APs, e.g., n > 11, RSS looks fine

I

For RSS, there are peaks in the Pc curve at n ∈ {4, 8, 12, 16} because the APs are evenly distributed around the area


28 October 2013

Varying noise standard deviation 1 RSS DIFF SSD DS

0.9


0.8



Simulation Results

0.6

0.5

0.4

0.3

0.2



0.7

0.1

0

2

4


6 8 10 Standard deviation of noise [σ]

(1)

12

14

with M = 2 devices and n = 6 APs

- Measurement Setup

Conclusions

I

Under low noise conditions (σ = 1, 2 dBm), the performance of SSD is similar with DIFF

I

When σ ≥ 3 dBm, Pc is decreased by 5%–10% for SSD

I

DIFF attains the same level of performance with DS

I

For RSS, Pc < 0.5 even under low noise conditions

- Concluding Remarks


28 October 2013

Varying number of crowdsourcing devices 1

0.9


0.8



Simulation Results

0.7

0.6

0.5

0.4

0.3

0.2


RSS DIFF SSD DS

0.1

0

1

2

3

4



5 6 Number of devices [M]

(1)

7

8

9

10

with σ = 3 dBm and n = 6 APs

- Measurement Setup


I

Pc decays linearly for the RSS approach

I

DIFF and SSD approaches are extremely robust and their performance is not affected as more devices contribute data

I

DIFF performs better than SSD and is very close to DS


28 October 2013

Measurement Setup I

Experimental data collected at the KIOS Research Center3 I



I


I


I

Performance evaluation I


Conclusions

I


I

3

RSS samples collected with 5 devices (HP iPAQ PDA, Asus eeePC laptop, HTC Flyer Android tablet, HTC Desire and Samsung Nexus S Android smartphones) 2100 location-tagged fingerprints for each device collected at 105 reference locations 960 location-tagged fingerprints for each device collected at 96 test locations Used the reference data to build device-specific radiomaps and crowdsourced radiomaps with different device combinations Used the test data to evaluate various crowdsourcing approaches in terms of the localization error RSS, DIFF and SSD approaches for crowdsourcing compared with DS (device-specific) RSS radiomap

The KIOS dataset is available to download at http://goo.gl/u7IoG


28 October 2013

Two-device crowdsourced radiomaps 8

9

7

8 7

- Motivation


Simulation Results

Localization Error [m]

Introduction

Localization Error [m]

6 5 4 3 2

4 3

1

0

RSS

DIFF

SSD

DS

0

RSS

DIFF

SSD

DS

Localization of the iPAQ (left) and Desire (right) devices

I

Two contributing devices (iPAQ, Nexus) that fully cover the localization area for crowdsourcing the radiomap

I

Differential approaches reduce error that is comparable to DS

I

For iPAQ, the median error is 3.4 m for RSS against 2 m for DIFF and SSD (75th percentile drops from 5.2 m to 3 m)

I

DIFF approach filters out high errors more effectively

- Measurement Setup

Conclusions

5

2

1



6



28 October 2013

Crowdsourcing with multiple devices 4.5


RSS DIFF SSD DS

3 Median Localization Error [m]

- Motivation

Median Localization Error [m]

3.5

Introduction

3.5

RSS DIFF SSD DS

4

3 2.5 2 1.5 1

2.5

2

1.5

1

0.5

0.5




0

0 1

2 3 4 Number of Crowdsourcing Devices

5

1

2 3 4 Number of Crowdsourcing Devices

5

Localization of the iPAQ with fully overlapping radiomaps (left) and the eeePC with non-overlapping radiomaps (right) I RSS performs poorly, e.g., for 5 devices the median error is 4.3 m

compared to 1.8 m for DIFF and 2.3 m for SSD I For DIFF and SSD the localization error does not vary

significantly as suggested by the simulations I DIFF outperforms SSD for any number of devices International Conference on Indoor Positioning and Indoor Navigation, Montb´ eliard - Belfort, France

28 October 2013

Concluding Remarks I

Notes I


I


I

Simulation Results

Our Contributions I


I


I

- Measurement Setup


I

Crowdsourcing stands as the only viable solution for building the radiomap considering effort, time and cost Our community has not appreciated its potential (0 papers in IPIN’10-11, 1 in IPIN’12 and 2 in IPIN’13) Evaluated DIFF and SSD methods for creating the RSS differences from the original RSS fingerprints Simulation and experimental findings indicate that differential fingerprinting is a promising solution DIFF performs better than SSD at the expense of higher computational complexity

Future Work I

Investigate other issues related to crowdsourcing, e.g. polluted data, non-uniform fingerprint distribution and the fast growing radiomap size


28 October 2013



Thank you for your attention




Contact Christos Laoudias KIOS Research Center for Intelligent Systems and Networks Department of Electrical & Computer Engineering University of Cyprus Email: [email protected]



28 October 2013




Extra Slides




28 October 2013

- Motivation




−20

−20

−30

−30

−40 −50 α=0.9155 β=13.3612

−60 −70 −80 −90 −100 −100

RSS pairs Linear Fit −80 −60 −40 Mean RSS from HTC Flyer [dBm]

−20

−40 −50 α=0.9439 β=1.321

−60 −70 −80 −90 −100 −100

RSS pairs Linear Fit −80 −60 −40 Mean RSS from HTC Desire [dBm]

−20

I

Several studies report a linear relation between the RSS values measured by heterogeneous devices

I

rij 2 = αm1 m2 rij 1 + βm1 m2 , m1 , m2 ∈ {1, . . . , M}, where (αm1 m2 , βm1 m2 ) are the coefficients between D (m1 ) and D (m2 )

I

Direct fusion of the different RSS radiomaps using (1) may degrade the quality of the crowdsourced radiomap

- Measurement Setup

Conclusions

Mean RSS from Asus eeePC [dBm]

Introduction

Mean RSS from HP iPAQ [dBm]

Linear relation between RSS values


(m )

(m )


28 October 2013