1

Energy-efficient Localization Via Personal Mobility Profiling

Ionut Constandache

Co-authors: Shravan Gaonkar, Matt Sayler, Romit Roy Choudhury and Landon Cox

2

Context

Pervasive wireless connectivity +

Localization technology =

Location-based applications (LBAs)

3

Context

Pervasive wireless connectivity +

Localization technology =

(iPhone AppStore: 3000 LBAs, Android: 600 LBAs)

Location-based applications (LBAs)

4

Location-Based Applications (LBAs)

  Two kinds of LBAs:   One-time location information: Geo-tagging, location-based recommendations, etc.

5

Location-Based Applications (LBAs)

  Two kinds of LBAs:   One-time location information: Geo-tagging, location-based recommendations, etc.

  Localization over long periods of time: GeoLife: shopping list when near a grocery store TrafficSense: real-time traffic conditions

6

Localization Technology

  LBAs rely on localization technology to get user position

7

Localization Technology

  LBAs rely on localization technology to get user position

Accuracy Technology 10m GPS 20-40m WiFi 200-400m

GSM

8

Localization Technology

  LBAs rely on localization technology to get user position

  LBAs executed on mobile phones

Accuracy Technology 10m GPS 20-40m WiFi 200-400m

GSM

9

Localization Technology

  LBAs rely on localization technology to get user position

  LBAs executed on mobile phones

Accuracy Technology 10m GPS 20-40m WiFi 200-400m

GSM

Energy Efficiency is important (localization for long time)

10

Localization Technology

Ideally Accurate and Energy-Efficient Localization

11

Energy

… sample every 30s

Battery Lifetime: GPS ~ 10h WiFi ~ 40h GSM ~ 60h

12

Energy

… sample every 30s

Battery shared with   Talk time, web browsing, photos, SMS, etc.

Battery Lifetime: GPS ~ 10h WiFi ~ 40h GSM ~ 60h

13

Energy

… sample every 30s

Battery shared with   Talk time, web browsing, photos, SMS, etc.

Localization energy budget only percentage of battery   20% of battery = 2h GPS or 8h WiFi

Battery Lifetime: GPS ~ 10h WiFi ~ 40h GSM ~ 60h

14

Energy

… sample every 30s

Battery shared with   Talk time, web browsing, photos, SMS, etc.

Localization energy budget only percentage of battery   20% of battery = 2h GPS or 8h WiFi

Battery Lifetime: GPS ~ 10h WiFi ~ 40h GSM ~ 60h

For limited energy budget what accuracy to expect?

15

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Problem Formulation

16

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Localization Error

t0 t1 t2 t3 t4 t5 t6 t7 Time

Problem Formulation

17

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Localization Error

t0 t1 t2 t3 t4 t5 t6 t7 Time GPS

Problem Formulation

18

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Localization Error

t0 t1 t2 t3 t4 t5 t6 t7 Time GPS

Problem Formulation

Accuracy gain from GPS Eng.: 1 GPS read

19

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Localization Error

t0 t1 t2 t3 t4 t5 t6 t7 Time GPS

Accuracy gain from GPS Eng.: 1 GPS read

Problem Formulation

Accuracy gain from WiFi Eng.: 1 WiFi read

WiFi

20

L(t0) L(t1) L(t2) L(t3) L(t4)

L(t6) L(t7)

L(t5)

Localization Error

t0 t1 t2 t3 t4 t5 t6 t7 Time GPS

Accuracy gain from GPS Eng.: 1 GPS read

Problem Formulation

Accuracy gain from WiFi Eng.: 1 WiFi read

WiFi

21

Given energy budget B, known Trace T, location readings costs egps , ewifi , egsm :

Schedule location readings to minimize Average Localization Error (ALE)

Problem Formulation

22

Given energy budget B, known Trace T, location readings costs egps , ewifi , egsm :

Schedule location readings to minimize Average Localization Error (ALE)

Problem Formulation

ALE = Avg. dist. between reported and actual location of the user

23

Given energy budget B, known Trace T, location readings costs egps , ewifi , egsm :

Schedule location readings to minimize Average Localization Error (ALE)

Problem Formulation

ALE = Avg. dist. between reported and actual location of the user

Find the Offline Optimal Accuracy

24

Results

B = 25% Battery Opt. GPS/WiFi/GSM

Trace 1 78.5m

Trace 2 58.6m

Trace 3 62.1m

25

B = 25% Battery Opt. GPS/WiFi/GSM

Trace 1 78.5m

Trace 2 58.6m

Trace 3 62.1m

Offline Optimal ALE > 60m

Results

26

Offline Optimal ALE > 60m

Results

Online Schemes Naturally Worse

B = 25% Battery Opt. GPS/WiFi/GSM

Trace 1 78.5m

Trace 2 58.6m

Trace 3 62.1m

27

Our Approach: EnLoc

  Reporting last sampled location increases inaccuracy

28

Our Approach: EnLoc

  Reporting last sampled location increases inaccuracy

  Prediction opportunities exist   Exploit habitual paths   Leverage population statistics when the user has deviated

29

Our Approach: EnLoc

  Reporting last sampled location increases inaccuracy

  Prediction opportunities exist   Exploit habitual paths   Leverage population statistics when the user has deviated

  EnLoc Solution:   Predict user location when not sampling   Sample when prediction is unreliable

30

EnLoc: Overview

Deviations

EnLoc

Habitual Paths

E.g. Regular path to office E.g. Going to a vacation

31

EnLoc: Overview

Deviations

EnLoc

Habitual Paths

E.g. Regular path to office

Per-user Mobility Profile

E.g. Going to a vacation

32

EnLoc: Overview

Deviations

EnLoc

Habitual Paths

E.g. Regular path to office E.g. Going to a vacation

Per-user Mobility Profile Population Statistics

33

Profiling Habitual Mobility

  Intuition: Humans have habitual activities   Going to/from office   Favorite grocery shop, cafeteria

34

Profiling Habitual Mobility

  Intuition: Humans have habitual activities   Going to/from office   Favorite grocery shop, cafeteria

  Habitual activities translate into habitual paths   E.g. path from home to office

35

Profiling Habitual Mobility

  Intuition: Humans have habitual activities   Going to/from office   Favorite grocery shop, cafeteria

  Habitual activities translate into habitual paths   E.g. path from home to office

  Habitual paths may branch   E.g., left for office, right for grocery   Q: How to solve uncertainty?   A: Schedule a location reading after the branching point.

36

Per-User Mobility Graph

User Habitual Paths

  Graph of habitual visited GPS coordinates

37

Per-User Mobility Graph

User Habitual Paths Logical Representation

  Graph of habitual visited GPS coordinates

38

Per-User Mobility Graph

  Graph of habitual visited GPS coordinates

  Sample location after branching points   Predict between branching points   # of BPs < # of location samples

(BP = branching point)

User Habitual Paths Logical Representation

39

Evaluation: Habitual Paths

  30 days of traces, loc. battery budget 25% per day   Assume phone speed known

40

Evaluation: Habitual Paths

  30 days of traces, loc. battery budget 25% per day   Assume phone speed known

41

Evaluation: Habitual Paths

  30 days of traces, loc. battery budget 25% per day   Assume phone speed known

Average ALE 12m

42

  Predict based on population statistics   If user on a certain street, at the next intersection

predict the most probable turn.

Deviations from habitual paths

43

  Predict based on population statistics   If user on a certain street, at the next intersection

predict the most probable turn.   Probability Maps computed from Google Map simulation

Deviations from habitual paths

44

  Predict based on population statistics   If user on a certain street, at the next intersection

predict the most probable turn.   Probability Maps computed from Google Map simulation

Deviations from habitual paths

Goodwin & Green

U-Turn Straight Right Left

E on Green 0 0.881 0.039 0.078

W on Green 0 0 0.596 0.403

N on Goodwin

0 0.640 0.359 0

S on Goodwin

0 0.513 0 0.486

45

Evaluation: Population Statistics

OptX: report last sampled location using sensor X (offline)

EnLoc-Deviate: Equally spaced GPS + population statistics (online). ALE ~ 32m

B = 25% Battery

46

Future Work/Limitations

  Assumed phone speed known   Infer speed using accelerometer   Energy consumption of accelerometer relatively small

  Deviations from habitual paths   Quickly detect/switch to deviation mode

  Probability Map hard to build on wider scale   Statistics from transportation departments

47

Conclusions

  Location is not for free   Phone battery cannot be invested entirely into localization

  Offline optimal accuracy computed   For specified energy budget   Known mobility trace

  However, online localization technique necessary

  EnLoc exploit prediction to reduce energy   Personal Mobility Profiling   Population Statistics

48

Questions?

Thank You!

Visit the SyNRG research group @ http://synrg.ee.duke.edu/