Chenshu Wu*, Jingao - PowerPoint Presentation

Slide1

Chenshu Wu*, Jingao Xu*, Zheng Yang*, Nicholas D. Lane^, Zuwei Yin**School of Software, Tsinghua University^University College London and Bell Labs9.13 in Maui, Hawaii

Gain Without Pain: Accurate WiFi-based Localization using Fingerprint Spatial Gradient

Slide2

MotivationVarious location-based ubiquitous applications.And locating with Wi-Fi is superior in Ubiquitous: Almost everywhere installed infrastructure.Low-cost: Off-the-shelf Wi-Fi devices.Non-invasive: not required to wear/carry any special devices.2Indoor LocationNavigationMobile Tracking

Slide3

Major Problems3RSS bias over timeThe RSS bias is large, usually >10dbNo matter whether the AP is strong or notSpatial ambiguity and Temporal instability

Slide4

Major Problems4Fingerprint spatial ambiguityFingerprint temporal instabilityOne fingerprint might be similar to those from two or more quite distant locationsRe-collected fingerprints from same location can’t match with initially fingerprintMismatch &

Degrade the performance of original fingerprint database!

Slide5

Existing Arts5Argus, Ubicomp ’15Acoustic, Mobicom ’12Followme, Mobicom ’15While achieved better accuracy, these solutions degraded the delightful ubiquity and induce additional costs

Slide6

Existing Arts6Walkie-Markie, NSDI’13GIFT, TIE ’16Need a path and continuous fingerprints, can’t work when user is static!

Slide7

Problem StatementAccurate localization using RSS fingerprint.Overcomes spatial ambiguity & temporal instability using only Wi-Fi signals.Without any additional system costs7

Slide8

Problem Statement8Gain With PainGain Without PainHigh precisionMitigate spatial ambiguity Ease temporal instabilityInduce system costBring many restrictionsSpecial hardware

Additional information

A

new

form

Slide9

Key InsightThe spatial relationships of multiple RSS fingerprints from neighbouring locations would be more robust than individual RSS fingerprints from one single location.Leverage Fingerprint Spatial Gradient9

Slide10

FSG Specification10The FSG profile for defined as : : RSSI fingerprint of location : similarity of

defined as common p-norm distance

: physical distance between locations

 

Slide11

FSG Specification11The FSG profile for defined as : 

Slide12

FSG AdvantagesLeverage Fingerprint Spatial GradientTemporally stable & Spatially distinctiveCertain spatial neighbouring relationships would be more stable12

Slide13

System Overview13Inputs are no more than traditional fingerprintsComputed within constant time

Slide14

ChallengesHow to select the nearby points set from the single location to generate the most useful FSG?How to compare the similarity between two FSG?How to locate according to a query fingerprint?14

Slide15

ChallengesHow to select the nearby points set from the single location to generate the most useful FSG?How to compare the similarity between two FSG?How to locate according to a query fingerprint?15

Slide16

Profiling a Reference Location16Obtaining the most stable and distinctive FSGMinimizing Local Gradient (MLG)Maximizing Global Gradient (MGG)Optimizing Spatial Stability (OptSS)

Slide17

ChallengesHow to select the nearby points set from the single location to generate the most useful FSG?How to compare the similarity between two FSG?How to locate according to a query fingerprint?17

Slide18

Comparing FSG18Cosine SimilarityTurning Function Distance

Discrete

Frechet

Distance(Walking

Dog

distance..)

 

Slide19

ChallengesHow to select the nearby points set from the single location to generate the most useful FSG?How to compare the similarity between two FSG?How to locate according to a query fingerprint?19

Slide20

Profiling a Query FingerprintUnknown location Q, its fingerprint and candidate location C:

Where

is location

C

’s neighboring points set

 

20

Slide21

Profiling a Query FingerprintUnknown location Q, its fingerprint and candidate location C:

Where

is location

C

’s neighboring points set

 

21

Slide22

Profiling a Query FingerprintUnknown location Q, its fingerprint and candidate location C:

Where

is location

C

’s neighboring points set

 

22

Same

location

1-m

away

2

-m

away

Slide23

23Evaluations & ResultsHow much gains we can get without pains?

Slide24

Experiment4 DevicesGoogle Nexus 5Google Nexus 6p*2Google Nexus 7 padHUAWEI Mate 83 Experiment areasAcademic: 23.4k samplesOffice: 27.2k samplesClassroom: 87.0k samplesExperiment measurementsEach location 60 measurements244 BaselinesHorus (Winet’08)TW-KNN (Pervasive Computing’14)KLDivGIFT (TIE’16)

Slide25

Experiment25OfficeAcademicClassroom

Slide26

Explore best metric26OptSS+CS achieves the highest accuracy with mean accuracy of 3.6m and 95th percentile error of 6.7m.Three profiling methodsThree FSG similarity

Slide27

Performance in different areas27Average accuracy in three areas: 3.3m,4.3m,3.8mSmaller areas yields better accuracyDifferent areas

Slide28

Performance comparison28Average accuracy outperforms ：Horus by 18.2%GIFT by 21.7%TW-KNN by 25.0% KLDiv by 30.8%Our system can combine TW-KNN and HorusDifferent methods

Slide29

Performance comparison29System latency(Average time delays for 1000 queries):ViVi: 0.33s per queryHorus: 0.28s per queryTW-KNN: 0.77s per queryGIFT: 0.84s per queryParameter StudyDifferent AP numberDifferent nearby points’ range

Slide30

ConclusionA new form based on RSS fingerprintAn underlying spatial properties of nearby fingerprints More stable and distinctive than the original RSS fingerprints. Present a system ViVi to incorporate FSGGain without the pains of resorting to additional information or restrictions or vulnerable RSS model assumptions. Median accuracy outperforms 4 state-of-the-art approaches by more than 20%30

Slide31

31Thanks!Q&AJingao XuTsinghua Universityxujingao13@gmail.com