L OCATING IN F INGERPRINT S PACE : W IRELESS I NDOOR LOCALIZATION WITH L ITTLE H UMAN I NTERVENTION Zheng Yang, Chenshu Wu, and Yunhao Liu MobiCom 2012.

LOCATING IN FINGERPRINT SPACE:WIRELESS INDOOR LOCALIZATION WITH LITTLE HUMAN INTERVENTION

Zheng Yang, Chenshu Wu, and Yunhao Liu

MobiCom 2012

- Sowhat 2012.08.20

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

MOTIVATION

RSSI fingerprinting-based localization

Site survey Time-consuming Labor-intensive Vulnerable to environmental dynamics Inevitable

OBJECTIVE

Wireless Indoor Localization Approach

RSSI Floor Plan User Movement

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

LIFS, SYSTEM ARCHITECTUREGeographical

dist.≠

Walking dist.

RSSI + Distance

MULTIDIMENSIONAL SCALING (MDS)

Information visualization for exploring similarities/dissimilarities in data

STRESS-FREE FLOOR PLAN

MDS

Geographical distance ≠ Walking distance,Ground-truth floor plan –conflict with measured distance

Sample grids in a floor plan (grid length l = 2m)

Distance matrix D = [dij],dij = walking distance between point i and j

Stress-free floor plan – 2D & 3D

FINGERPRINT SPACE – FINGERPRINT & DISTANCE MEASUREMENT

Fingerprints and distance collection Record while walking Footsteps every consecutive steps by accelerometer Set of fingerprints, F = {fi, i = 1~n}

Distance(footsteps) matrix, D’=[d’ij]

Pre-processing Merge similar fingerprints (δij<ε)

Accelerometer readingTwice integration Distance: NoiceLocal variance threshold method Step count

Stride lengths vary? MDS tolerate measurement errors

FINGERPRINT SPACE – FINGERPRINT SPACE CONSTRUCTION

Adequate fingerprints & distance1. 10x sample locations in stress-free floor plan2. First several days for training

d’ij unavailable d’ij = d’ik + d’kj

Shortest path update D’ all-pairs of fingerprints Floyd-Warshall algorithm

MDS Fingerprint space 2D & 3D

MAPPING –CORRIDOR & ROOM RECOGNITION

Corridor recognition (Fc) Higher prob. on a randomly chosen shortest path Minimum spanning tree Betweenness Watershed

1. Size(corridor) / Size(all)2. Large gap of betweenness values

Room recognition (FRi) k-means algorithm

(k = number of rooms)

Classify fingerprints into the corridor or rooms

Fingerprints collected near “doors”

PD = {p1, p2, …, pk}, stress-free floor planFD , fingerprint space

distance matrix D and D’ l = (lp1, lp2, …, lp k-1)l’ = (lf1, l’f2, …, l’f k-1)

cosine similarity

MAPPING –REFERENCE POINT

Near-door fingerprints, FD,labeled with real locations

1. Map near-door fingerprintsto real locations (FD → PD)

2. Map rooms to rooms

Floor-level transformation Stress-free floor plan ≠ Fingerprint space

∵ translation, rotation, reflection Transform matrix,

xi = coordinate of fi ∈ FD

yi = coordinate of pi ∈ PD

For fingerprint with coordinate xreal location = sample location closest to Ax + B

Room-level transformation Room by room Doors and room corners as reference point Transformation matrix

MAPPING –SPACE TRANSFORMATION

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

HARDWARE AND ENVIRONMENT

2 Google Nexus S phones Typical office building covering 1600m2

16 rooms,5 large – 142m2, 7 small, 4 inaccessible

26 Aps, 15 are with known location 2m x 2m grids, 292 sample locations

EXPERIMENT DESIGN

5 hours with 4 volunteers Fingerprints recording – every 4~5 steps

(2~3m) Accelerometer –

work in different frequency based on detecting movement

600 user traces, with 16498 fingerprints Corridor, >500 paths

Small rooms, >5 pathsLarge rooms, >10 paths

Half of data used for training,half …………………... in operating phase

THRESHOLD VALUE OF FINGERPRINT DISSIMILARITY

STEP COUNT 5 ~ 200 footsteps

Error rate = 2% in number of detected steps

Accumulative error of long path Unobvious performance drop ∵ only use inter-fingerprint step counts

FINGERPRINT SPACE 795 fingerprints when ε = 30

CORRIDOR RECOGNITION

Refining Perform MST iteratively Sift low betweenness Until MST forms a single line

ROOM RECOGNITION

REFERENCE POINT MAPPING

POINT MAPPING

• 96 percentile < 4m• Average mapping error = 1.33m

LOCALIZATION ERROR Emulate 8249 queries using real data on LiFS Location error

Average,LiFS = 5.88mRADAR = 3.42m

Percentile of LiFS80 < 9m / 60 < 6m

Caused bysymmetric structure

Fairly reasonable!

Room error = 10.91%

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

DISCUSSION

Global reference point Last reported GPS location

Locations of APsSimilar surrounding sound signature…

Could be added in LiFS for more robust mapping Key for symmetric floor plans / multi-floor fuildings

Large open environment

OUTLINE Introduction

System Design

Evaluation

Discussion

Conclusion

CONCLUSION

LiFS Spatial relation of RSSI fingerprints + Floor plan Low human cost

Comments Clear architecture Not specific descriptions in evaluation

THANKS FOR LISTENING ~