Deep Generative Model-based RSSI Synthesis for Indoor Localization

Presentation on ECTICON 2022 with title "Deep Generative Model-based
RSSI Synthesis for Indoor Localization"

1. Deep Generative Model-based
RSSI Synthesis for Indoor Localization
Dwi Joko Suroso, PanaratCherntanomwong, PitikhateSooraksa
School of Engineering
King Mongkut’s Institute of Technology Ladkrabang
25 May 2022

2. •Introduction to Indoor Localization and Fingerprint Technique
•The Pain Points and Database Synthesis
•Generative Deep Model and Variational Autoencoders (VAE)
•Measurement Campaign
•RSSI Synthesis via VAE
•Conclusions and Future Works
Outline
2

3. Location-based Service (LBS)
Falkowski, T, Günther, M, Jürgenhake C, Anacker H, Dumitrescu R. Towards a design approach for
industrial indoor location-based services (I2LBS). International Design Conference. 2018. 1043–1054. 3
LBS Applications
How about LBS indoor environment?
Yourstory.com
OneofLBS’featuresislocalization:
•Positioningtechnologyi.e.,GlobalPositioningSystem(GPS)
•Referencesystem(mapping)
LBS features

4. Global Positioning System (GPS)
GPS Illustration
GPSisnotreliabletoapplyforindoor:
•Thesignalblockageofwallsdegradesthe
signal.
•Theothertechnologiescanbeapplied
i.e.,radio-frequency(RF)-basedWi-Fi,
Bluetooth,ultrawideband(UWB),etc.
GPSissatellite-basednavigationsystem.
GPSinpositioning:
•Inoutdoorscenario,theerrorprediction
rangingfrom5-20m.
•Degradingtheperformanceinindoor:the
complexenvironment,non-lineofsight
(NLOS)condition,shadowingeffects.
GPS for outdoor vs indoor.
(Sumber : https://www.sewio.net/indoor-gps-tracking/)
O. Kerem, B. Ayhan dan I. Tekin, “Indoor positioning based on global positioning system signals,” Microw. Opt. Technol. Lett., pp. 1091-1097, 2013.4
Whatisindoorpositioningsystem(IPS)/indoorlocalization?

5. Indoor Localization
W. Sakpere, M. Adeyeye Oshin, and N. B. Mlitwa, “A State-of-the-Art Survey of Indoor Positioning and Navigation Systems and Technologies,” SACJ, vol. 29, no. 3, Dec. 2017
S. Sadowski and P. Spachos, “RSSI-Based Indoor Localization With the Internet of Things,” IEEE Access, vol. 6, pp. 30149–30161, 2018
5
Definition
Indoorlocalizationis“GPS”indoor.Itappliesthetechnologiesand
localizationtechniquestodevelopthepositioningsystem.
TechnologyAccuracy (m)Range (m)Power (W)
GPS1 –20Global500
RFID11 –500.02 –0.3
Wi-Fi1 –5< 1000.5 –1
UWB< 0.3< 3000.03
BLE1< 100.001
Zigbee1 -5< 300.02 –0.04
Technology
Light Wave
Mechanical
Acoustic
Wave
Radio
Frequency
(RF)
Ultra-wide Band (UWB)Radio Frequency Identification RFID)Bluetooth Low Energy (BLE)
Indoor Localization Technologies
•Low-cost,widelyavailable.
•Providinglocalizationparameter,e.g.,receivedsignalstrengthindicator(RSSI).Wi-Fi
Radio Frequency (RF)-based

6. Received Signal Strength Indicator (RSSI)
6
•RSSI values represent the relative quality of a received signal on a device.
•It follows log-loss distance path loss model.
•In non-line-of-sight (NLoS) condition, there are signal obstruction due to multipath effects.
!(#)=&−10·+·log!"(#!#)
P(d) = received power (dBm) in distance, d(m)
A = reference power (dBm) at d0, reference distance, typically 1 m.
n = path loss exponent
Multipath EffectsIdeal
Fingerprinting Localization method can be selected.
RSSI [dBm]
Distance [m]
K. Heurtefeux and F. Valois, "Is RSSI a Good Choice for Localization
in Wireless Sensor Network?,"2012 IEEE 26th International
Conference on AINA 2012, pp. 732-739
A. Goldsmith, Wireless Communications. Standford University, 2004.

7. Offline
Area of Interest
Fingerprint
Database
Localization
Algorithm
Location
Estimation
Signal
Measurement
(Fingerprint
Location,
Parameter)
Online
-36-48-51
Fingerprint LocationTarget
Area of Interest
Fingerprint Technique
General Fingerprint Technique Illustration
S. He and S. -. G. Chan, "Wi-Fi Fingerprint-Based Indoor Positioning: Recent Advances and
Comparisons," inIEEE Communications Surveys & Tutorials, vol. 18, no. 1, pp. 466-490,
Firstquarter 2016, doi: 10.1109/COMST.2015.2464084.7

8. Fingerprint Technique: Advantages
Multipath and non-line-of-sight propagation effects considered in the preconstructed database.
Simple power-based parameter i.e., Received signal strength indicator (RSSI)
Relatively high accuracy (depends on database features and quality)
V. Moghtadaiee, S. A. Ghorashi and M. Ghavami, "New Reconstructed Database for Cost Reduction in Indoor Fingerprinting Localization," inIEEE Access,
vol. 7, pp. 104462-104477, 2019, doi: 10.1109/ACCESS.2019.2932024.8
Need high density database for high accuracy

9. X. Wang, Z. Chen, S. Zhang, and J. Zhu, “Super-Resolution Based Fingerprint Augment for Indoor Wi-Fi Localization,” 2020 IEEE Glob. Commun. Conf.
GLOBECOM 2020 -Proc., vol. 2020-January 2020. 9
High-labor costTime InefficientNot flexible
Reducing the density of site surveyCutting down large survey
Pain Points
Splitting to the small portion Increase density by database synthesis
e.g., crowdsourcing-data
Classical InterpolationPropagation ModelImage MethodGenerative Deep Model
Solutions
Fingerprint Technique: Pain Points and Database Synthesis

10. 10
Generative Deep Model
Thebasicunderstandingofthegenerativemodelistolearntheinputandgenerate
similarvaluesastheoutputbythelatentdistributionlearnedfromtheinput
distribution.
•David Foster, “Generative deep learning: teaching machines to paint, write, compose, and play,” O’Reilly, 2019.
•Goodfellow, I. (2016). Nips 2016 tutorial: Generative adversarial networks.arXivpreprint arXiv:1701.00160.
Explicit Density estimates
the true pdf or cdf over the
sample space.
Implicit Density does not produce
explicit densities but generates a
function that can draw samples from
the true distribution.

11. 11
Existing Literature
•Somepublicationsconsideredtoutilizechannelstateinformation(CSI)forVAE
input.However,acquiringCSIisquitecomplexandneedshighcostand
sophisticatedequipmentforitsextraction.
•Otherpublicationsproposedthesemi-supervisedVAEinindoorlocalizationusing
RSSI(Byaddedonemoreneuralnetworkasdiscriminator).
•Theotherresearchersapplythegenerativeadversarialnetworks(GAN)toaugment
RSSIvaluesandlocalizethetarget.GANismorereliablewiththevastdataset;
however,itcoststhetrainingtimesinceitisbasedonadversarialnetworkswhich
employstwoduelingneuralnetworks.
•OurresearchoffersmorestraightforwardsolutioninRSSIsynthesisby
implementingVAEtoenhancethefingerprintdatabase.
•K.M.ChenandR.Y.Chang,“Semi-SupervisedLearningwithGANsforDevice-FreeFingerprintingIndoorLocalization,”GLOBECOM2020-Proceedings,2020
•X.Chen,et.al,“Fidora:RobustWiFi-basedIndoorLocalizationviaUnsupervisedDomainAdaptation,”IEEEInternetofThingsJournal,pp.1–1,2022,
•W.Qian,et.al,“Supervisedandsemi-superviseddeepprobabilisticmodelsforindoorpositioningproblems,”Neurocomputing,vol.435,pp.228–238,May2021,
•W.Njima,et.al,“IndoorLocalizationUsingDataAugmentationviaSelectiveGenerativeAdversarialNetworks,”IEEEAccess,vol.9,no.Ml,pp.98337–98347,2021
•M.Nabati,et.al,“UsingSyntheticDatatoEnhancetheAccuracyofFingerprint-BasedLocalization:ADeepLearningApproach,”IEEESensorsLetters,vol.4,no.4,2020.

12. 12
/01|/ 0/|13/
4
5$
1
Input OutputEncoder DecoderMean
Variance
Latent
Variable
Variational Autoencoders (VAE)
D. P. Kingmaand M. Welling, “Auto-encoding variational bayes,” ICLR 2014 -Conference Track Proceedings, pp. 1–14, 2014.
Thebasicunderstandingofthegenerativemodelistolearntheinputandgenerate
similarvaluesastheoutputbythelatentdistributionlearnedfromtheinput
distribution.

13. Measurement Campaign
Actual Measurement Setup
13
•Thetotalareaofmeasurement
campaignis10×15.14'".
•Weconsideredthe25fingerprint
locationsina5×5'"areaofinterestand
recordedtheRSSIvaluesfrom8
referencenodes(RNs).
Measurement Layout
NameDevice/toolsSpecificationUse
Reference/
sink/server
node
ESP32 (Wi-Fi
standard)
IEEE 802.11
Memory 520 kB
Wi-Fi transceiver, RSSI
values
SoftwareArduino IDE,
Jupyter
Notebook
1.8.5 version.
Python 3.7,
Keras
To program the ESP32.
To embedded the
generative model
algorithm.
Measurement Devices/Tools

14. Measurement Campaign
8 reference nodes (RNs), 1 sink node illustration setup
14
ESP32 for reference/sink/server node

Server node
•8RNs,oneasthetarget/sinknode,andoneservernodeconnectedtoacomputerfor
RSSIdataacquisition.
•Thetotaldatasetfortheselocationsis1,489rows,and8(eight)columnsrepresentthe
numberofRNs.(3-minutesspan,60setsRSSIdataforeachfingerprint).

15. VAE Implementation: RSSI-to-image Conversion
•RSSI_1indicatestheRSSIvaluesFrom
RN1,andeachdifferentvaluesfrom
RNsyielddifferentcolor.
•Thenormalizationisdonebyhighest
value(100dBm)andlowestvalue(-
100dBm)asreferencefor0(min
value)and255(maxvalue)for
grayscaleconversion.
•Thenullvalueinthelastpixelinour
representedimageissetto100dBm
(yellow).
%='−100×255
100
•%and'isthenormalizedandreal
RSSIvalue.
S.Tiku,et.al,“QuickLoc:Adaptivedeep-learningforfastindoorlocalizationwithmobiledevices,”ACMTransactionsonCyber-PhysicalSystems,vol.5,no.4,pp.1–28,2021.15

16. (5,5)
(4,4)
(3,3)
(2,2)
(1,1)
0~60&ℎ
360~420&ℎ
760~820'ℎ
1120~1180&ℎ
1400~1460&ℎ
Sampled Locations and Represented RSSI Values
•TheRSSIvaluesinlocation(1,1)
arerepresentedbytherow
1~60!ℎinthedataset.
•WecomparetheVAEsynthesis
resultinthis5sampledlocations
andcomparetheRSSIvaluesfor
actualandsynthesis.
16

17. Latent Dimension 1
Latent Dimension 2
Results: Latent Distribution vs. Epochs
Latent Dimension 1
Latent Dimension 2
Latent Dimension 1
Latent Dimension 2
10 epochs100 epochs1000 epochs
•Thelatentdistributionsonepoch10werepilingup;inepoch100and1000theclustercorrelationis
startingtoreduce.
•Thispillingupandhighcorrelatedbetweenvalue1to5affecttheRSSIvalues,sinceintheactual
RSSItheyarearedifferent,especiallywhenweobservedRSSIvaluescorrespondingtotheRN
sequentialandposition.
17

18. Epoch=10, Coordinate (1,1)Synthetic RSSI
Latent Dimension 1
Latent Dimension 2
-1-8-19
-5-9-13
-8-80
RSSI
Discrepancy
RSSI Synthesis Results: 10 Epochs
−62−53−49
−51−55−49
−52−50
−61−61−68
−56−64−62
−60−58
Actual RSSI
18

19. Epoch=100, Coordinate (1,1)Synthetic RSSI
Latent Dimension 1
Latent Dimension 2
-3-3-15
-3-6-15
-8-110
RSSI
Discrepancy
−62−53−49
−51−55−49
−52−50
−61−61−68
−56−64−62
−60−58
Actual RSSI
−64−58−53
−53−58−47
−52−47
19
RSSI Synthesis Results: 100 Epochs

20. Epoch=1000, Coordinate (1,1)Synthetic RSSI
Latent Dimension 1
Latent Dimension 2
-10-19-4
-8-6-19
-2-70
RSSI
Discrepancy
−62−53−49
−51−55−49
−52−50
−61−61−68
−56−64−62
−60−58
Actual RSSI
−64−58−53
−53−58−47
−52−47
−71−80−64
−48−58−41
−58−51
20
RSSI Synthesis Results: 1000 Epochs

21. AverageRSSI Synthesis Error for All Sampled Locations
•Theaccuracyisacceptable,especiallyfor
1000epochresultssincetheerrorfrom5
positionsarerelativelylowi.e.,-4.09dBm
averageerrorin(1,1).
•Wecanimprovealmostallaverage
discrepancyerrorsbetweenactualand
RSSIsynthesisRSSIfromthesampled
locationsbyincreasingepochsinplotting
latentdistribution.
21

22. Conclusions and Future Works
•WeproposedVAEimplementationtosynthesizeRSSIvaluesfromactualRSSI
fromarelativelysmallmeasurementcampaign.
•TheresultsshowthatthelatentdistributioninourVAEimplementationcanbe
usedforRSSIsynthesis.
•Observingthelatentdistributionfromallepochsshowsthateachclusterof
coordinatesisnotsodistinctfromotherclusters,makingsomeRSSIsyntheses
inaccurate.
•ThenumberofRSSIdatasetswehave,andthesmallimagesizeassumption(3-
by-3)couldaffectthishighcorrelationbetweenclusters.
•Ournear-futureworkistoexploredenserdatasetsviabothpubliclyavailable
datasetsandconductingmeasurementcampaigns.
22

23. 23
Thank you very much..