1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin...

20
1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban Sensing is collaborative work of many faculty, staff, and students in partnership with NSF NeTS-FIND, Cisco, Nokia, Schematic, Sun, UCLA REMAP, UCLA ITS, Walt Disney Imagineering R&D Jeffrey Hightower Intel Labs Seattle Ramesh Govindan USC Embedded Networks Laboratory Mobile phones as instruments to understand physical processes in the world

Transcript of 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin...

Page 1: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

1

Discovering Semantically Meaningful Places from Pervasive RF-Beacons

Donnie Kim, Deborah EstrinUCLA Center for Embedded Networked Sensing

CENS Urban Sensing is collaborative work of many faculty, staff, and studentsin partnership with NSF NeTS-FIND, Cisco, Nokia, Schematic, Sun, UCLA REMAP, UCLA ITS, Walt Disney Imagineering R&D

Jeffrey HightowerIntel Labs Seattle

Ramesh GovindanUSC Embedded Networks Laboratory

Mobile phones as instruments to understand physical processes in the world

Page 2: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

2

Places We Go

Indoor PlacesMost of the places we go are indoors. A single building can have multiple places.(e.g., multiplex building, shopping malls, etc.)

Visit FrequencySome places are visited more often than others.

Outdoor PlacesSome are outdoors. (e.g., bus stops, tennis courts, plazas, etc.)

Visit DurationSome places are visited longer than others.

Page 3: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

3

Why Finding Places Matters?

Location Aware RemindersTo-do lists

Social Networking Applications Twitter, Facebook, etc.

Health management + InterventionContext triggered behavior interventions, self-monitoring

Human Spatial and Temporal Behavior Research DataResearch for urban planning, architecture, epidemics

Page 4: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

4

Visit HistoryVisit HistoryPlace ServicePlace ServicePlace Signatures

Place Signatures

GPS WiFi GSM

App X App Y App Z

System Overview

Page 5: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

5

Discovering Places from RF-Beacons

Page 6: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

6

Discovering Places from RF-Beacons

Page 7: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

7

PlaceSense Designed to discover places by continuously monitoring the radio beacons

Involves Two Steps:1.Entering: Detecting when the radio environment is stabilizing2.Exiting: Detecting when the radio environment is changing

Stable Radio Environment?Familiar beacons: if the previous scan windows contained it New beacons: if none of the previous scan windows contained it

(e.g., WiFi, Cell tower)

Entering ExitingIntermittent beacons

EnteringExiting

Scan Window:A window size w defines the smallest time unit in which the algorithm will determine entrance/departure to a place(non-overlapping)

Page 8: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

8

Step 1: Sensing EntranceContinuously seen stable scan windows imply a potential entrance to a place.

Stable Scan Window?If a scan window does not contain any new beacons*, its stable.

* if none of the previous scan windows contained it

Previous Scan Windows?Current scan window is saved and compared against the following scan.Scan windows are accumulated until entrance is determined or a new beacon is found. How many continuous stable scans?Stable depth, smax, specifies how many stable scan windows must be seen.

{ }PreviousScanWindows

Conservative approach**Empties previous scan windows when a new beacon is found.(to filter out “hallway beacons”)

Hallway beacons

** [05 Hightower] proposed to tolerate some scan windows with new beacons instead of rapidly emptying.

Page 9: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

9

Step 2: Sensing DepartureDetecting a changing radio environment that indicates a departure from a place.

Changing radio environment?Detecting new beacons or missing familiar beacons* implies the device is leaving.

* If the previous scan windows contained it

Problem: Infrequent BeaconsMissing: detected at the beginning but disappears does not imply a departure Late coming: not detected at the beginning but appears does not imply a departure

Filtering Infrequent Beacons outRepresentative Beacons: Focusing on beacons with high response rate

Rk,x: response rate of beacon x at place knk : total scan count since the place was entered

Hybrid approach**Missing representative beacons & detecting new beacons

Missing infrequent beacons

Latecomers

Representative beacons

** [05 Hightower] only relies on detecting new beacons.

Page 10: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

10

More Perks for Sensing Entrance/Departure

Avoiding a single scan window determining a departureTolerance depth, tmax, specifies at least how many scan windows must be unstable.Prevents infrequent beacons dividing a single visit into multiple visits.

Visiting closely located places head to headTolerance depth introduces delays on determining a departure.The delay may effect detecting entrance to the subsequent place. (If the travel time between two places is less than the delay)

Traveling between closely located places

Buffering StrategyAllows rapidly detecting place entry after quick transitions.Buffers overlapping data and starts entry determination in parallel, as soon as the t value is below tmax.

Page 11: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

11

Place Learning – Two Classes

Geometry-based Fingerprint-based

InputLocation coordinates

(e.g., GPS, WiFi/Cell tower triangulation)

Radio environment(e.g., currently visible cell towers,

WiFi access points)

ProsTightly coupled with the

geographical location of the place

Does not depend on the underlying positioning system’s accuracy

(especially indoors)

Cons

Depends on the underlying positioning system’s accuracy and

availability

Radio environment may change over time (affecting recognition not

necessarily detection)

Page 12: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

12

Experiments – Data CollectionMobile Device’s Hardware/SoftwareNokia N95 mobile phone: integrated GPS and built-in WiFiCampaignr: Software configured to collect GPS/WiFi/GSM traces every 10 secondsData uploaded to a server every night

Data CollectionThree data collectors

Scripted Tour: for accurate ground-truth (on UCLA campus)Each data collector individually selected 10 places they go often (30 visits for 8,10,15 min)

Real-life Data: for further validationCollected 4 week-long trace logs from each collectors as they went about their normal life Ground-truthEach data collector kept a diary of place visits (≥ 5 min) [enter time, leave time, name]Webpage illustrating the GPS coordinates:Provided for reviews/corrections(however, GPS data was not available in most of the indoors)

Time accuracy of the dairy deteriorated within the first few days. (~ 5 min)

Page 13: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

13

Experiments – Evaluation Metrics

* [07 Zhou] did not considered merged and divided

Four types of erroneous place discovery

Remembered Places: recorded by peopleDiscovered Places: found by algorithms

Interesting Places: forgotten place visits

(+) Correct, Interesting(−) False, Missed, Merged, Divided

Precision = # Correct + # Interesting

# Discovered

Recall = # Correct

# Remembered

Page 14: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

14

Experiments – Results

PlaceSense reduces the number of missed places while also increasing the number of interesting and false places.

Many indoor places were merged as a single visit

Page 15: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

15

Experiments – Results by Users

Aaron Bryan Chris All

PS BP KA PS BP KA PS BP KA PS BP KA

Cor.233 156 81 251 182 63 242 175 79 726 513 223

Int.10 1 2 6 1 0 13 0 2 29 2 4

Mer.6 14 138 23 15 185 16 15 138 45 44 461

Div.2 21 1 2 30 3 6 12 2 10 63 6

Mis.0 50 21 3 52 28 3 65 48 6 167 97

Fal.6 2 10 14 2 20 14 4 21 34 8 51

Recall0.97 0.65 0.34 0.90 0.65 0.23 0.91 0.66 0.30 0.92 0.65 0.28

Precision0.95 0.81 0.36 0.87 0.80 0.23 0.88 0.85 0.33 0.89 0.82 0.30

Both Precision and Recall is improved by significantly increasing the number of correct places

Precision = # Correct + # Interesting

# DiscoveredRecall =

# Correct

# Remembered

PS: PlaceSense, BP: BeaconPrint, KA: Kang et al.* Names are pseudonyms

Page 16: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

16

Experiments – Does it help recognition?

Yes!

Significantly improves discovering and recognizing short visits

Frequently visited places are often briefly visited

Page 17: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

17

Summary

PlaceSense provides a significant improvement in discovering and recognizing places.• PlaceSense (precision: 89%, recall: 92%)• BeaconPrint (precision: 82%, recall: 65%)• PlaceSense accuracy gains are particularly noticeable in challenging radio environments where beacons are inconsistent and coarse• PlaceSense detects entrance/departure time with over twice the precision of previous approaches (thanks to judicious use of buffering and timing)• PlaceSense is accurate at discovering places visited for short durations* (less than 30 minutes) or places where the device remains mobile

* Valuable to emerging applications like life-logging and social location sharing

Page 18: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

18

Thanks for your time.Questions?

http://www.cs.ucla.edu/[email protected]

Page 19: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

19

Appendix – Discovered Places

HomeCENSEng-IV 14-129BChevron VermontBH4404Chipotle WestwoodFowler A103BKinsey Pav 1220MS5200Borders WestwoodBruin PlazaTarget Highland SDYin-Yin ChineseiMax RegalWooden CenterFamima!

Ranch 99 DublinBH5436Powell LibraryTarget La BreaMarc MelroseHaines 220Ralphs WestwoodAckerman Post OfficeParking Lot4Barnes & Novel WestwoodBH3276Verra’s OfficeKerkhoff PatioAckerman Bus stopAmi RestaurantLAX terminal 2

BH3803Bus stop HomeBH4760Mr. NoodlesIn-and-out WestwoodYamato JapaneseCoffee bean WestwoodBH3771Ackerman TsunamiTrader Joes NationalSeas CaféRalphs OverlandCoffee Bean GaleyTennis court PalmsWhole Foods WestwoodStarbucks Venice…

Page 20: 1 Discovering Semantically Meaningful Places from Pervasive RF-Beacons Donnie Kim, Deborah Estrin UCLA Center for Embedded Networked Sensing CENS Urban.

20

Appendix – Time Accuracy