Synchronicity: Pushing the Envelope of Fine-Grained Localization with Distributed MIMO

Jie Xiong*, Kyle Jamieson*, Karthikeyan Sundaresan†

*University College London †NEC Labs America

Sep 11th, 2014

ACM HotWireless ‘14

•  Outdoors: GPS •  Accurate enough for navigation •  Signals fade in indoor environments

•  Precise indoor location enables: •  Augmented reality on smart devices •  Localization of items in huge space

Precise location services are important

2

MIMO technology is popular

3

MIMO 802.11n

MU-MIMO 802.11ac

Distributed MIMO

802.11xx ?

•  Increasing number of antennas on an access point •  Higher bandwidth

AoA localization (ArrayTrack [NSDI’13])

Channelization 5.17 GHz 5.33 GHz

20 MHz

40 MHz

New with 802.11ac

ToA (SAIL [Mobisys’14]) TDoA localization

80 MHz

160 MHz

Time difference of arrival (TDoA)

4

A1

A2

A3

d2-d1=d21(hyperbola 2)

d3-d1=d31 (hyperbola 1)

Client

d3 d1

d2

Previous disadvantageous: •  APs were not time synchronized •  Limited bandwidth (20MHz)

New opportunities for TDoA: •  Larger bandwidth: upto 160MHz for 802.11ac •  Distributed MIMO: APs achieve precise time

synchronization (SourceSync [SIGCOMM’10], Clock Sync [MobiCom’14 poster])

•  Channel State Information available for each sub-carrier

TDoA: New opportunities

5

Challenge: TDoA resolution is limited

6

Physical Layer Bandwidth Resolution 802.11g 20 MHz 15 m 802.11n 40 MHz 7.6 m 802.11ac <160 MHz >1.9 m

Super resolution schemes can improve the naïve resolution above to some extent

0

0.5

Nor

mal

ized

am

plitu

de

D=8.1m

MUSIC’s resolution limit

7

Naïve: 15m resolution (20MHz sampling rate)

MUSIC: around 8m resolution

The two peaks merged into one eventually

When two paths get closer, MUSIC fails to resolve them

correctly

0

0.5 D=10.7mGround truthGround truth

0

0.5D=5.4m

0 20 40 60 80 100 1200

0.5

Time (ns)

D=2.7m

Synchronicity: key observation When the sampling rate is not high enough to resolve all the signals correctly, the stronger peak is more accurate!

8

0 50 100 150 200 2500

2

4

6

8

Norm

alize

d amp

litude

Ground truth

MUSIC peak

ns

Key observation verification

9

!  Stronger peak has less error

!  The error becomes smaller with larger signal strength difference

0

0.5

D=5.4m

3dB 6dB 9dB0

0.5

1.0

1.5

2.0

2.5

3.0

Peak

Pos

ition

Erro

r (m

eter

s)

Signal strength ratio (LOS/NLOS)

Error for LOS peakError for NLOS peak

3dB 6dB 9dB0

0.5

1.0

1.5

2.0

2.5

3.0

Peak

Pos

ition

Erro

r (m

eter

s)

Signal strength ratio (LOS/NLOS)

Error for LOS peak

ns

Merged peak can still be accurate

10

As long as the LOS path is stronger

!  The larger signal strength difference, the smaller the peak error

0 20 40 60 80 100 1200

0.5

Time (ns)

D=2.7 m

22 18 15 11 9 5 2 0 −4 −70

0.5

1

1.5

2

2.5

Peak

pos

ition

erro

r in

met

ers

Signal strength ratio in dB (direct−path/reflection−path)

" Spectrum identification (SI) on a single AP

" Triangle inequality (TI) across multiple APs

" Clustering and outlier rejection across multiple APs

Synchronicity step outline

11

•  Compare the peak distance with the threshold: verify whether the spectrum falls in the inaccurate zone

•  Compare the amplitudes of the peaks: discard the spectrum with weaker first peak

12

Spectrum identification (SI)

✔

0 25 50 75 10002468

Time in ns

0 25 50 75 10002468

Nor

mal

ized

am

plitu

de

Peakdistance

Peakdistance

0 25 50 75 10002468

Time in ns

0 25 50 75 10002468

Nor

mal

ized

am

plitu

de

Useful (keep)

Useless (discard)✗

0 50 100 150 200 2500

2

4

Time in ns

0 50 100 150 200 2500

2

4

6

Norm

alize

d am

plitu

de

Useless (discard)

Useful(keep)

The peak is mergedtowards LOS peak

The peak is mergedtowards NLOS peak

Spectrum identification (SI) of merged peak

13

•  Identify whether the first peak is a single signal peak or merged peak

•  If it’s a merged peak, check the

skewness of the peak: discard the spectrum with forward skewness

✔

✗

" Spectrum identification (SI) on a single AP

" Triangle inequality (TI) across multiple APs

" Clustering and outlier rejection across multiple APs

Synchronicity step outline

14

Multi-AP processing for LOS path blockage: triangle inequality (TI)

15

A1 A2 dc

Client

d2

d1

dr

The LOS path difference is always smaller than the AP distance dc

If dr – d1 > dc is detected dr is NLOS path and should be discarded

" Spectrum identification (SI) on a single AP

" Triangle inequality (TI) across multiple APs

" Clustering and outlier rejection across multiple APs

Synchronicity step outline

16

Multi-AP processing for LOS path blockage: clustering and outlier rejection

17

A5 is blocked for NLOS

E125 is far away from other location estimates

The LOS-path blocked AP localizes client to random positions while LOS AP estimates are clustered

A3

A2

S

A5

VS E123

E124

A1

A4

E234

E125

Combines MUSIC and Matrix Pencil

18

!  MP : Matrix Pencil on raw data !  MUSIC: MUSIC on noise space data !  Synchronicity: (1) MUSIC to divide data (2) Apply Matrix Pencil on signal-space data

MP location estimate

MUSIC location estimate

Synchronicity location estimate

Raw data Noise space

Signal space

Matrix Pencil

MUSIC

MUSIC

Matrix Pencil

Experimental setup •  Each of the 6 antennas on WARP emulates one AP

•  Soekris boxes equipped with 802.11 radios as clients

19

Synchronicity improves the median accuracy from 3.6m (MUSIC) to 1.7m

Overall performance

20

0 200 400 600 800 10000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

MUSIC

0 200 400 600 800 10000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

SynchronicityMUSIC

0 200 400 600 800 1000 Location error in cm

*We borrow time synchronization error data from SourceSync

Performance/w time synchronization error

21

0 100 200 300 400 500 6000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

Synchronicity wo synchronization error

0 100 200 300 400 500 6000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

Synchronicity wo synchronization error Synchronicity w 5 ns (95th percentile) error

0 100 200 300 400 500 6000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

Synchronicity wo synchronization error Synchronicity w 5 ns (95th percentile) errorSynchronicity w 10 ns (95th percentile) error

0 100 200 300 400 500 6000

0.2

0.4

0.6

0.8

1

Location Error in cm

CD

F

Synchronicity wo synchronization error Synchronicity w 5 ns (95th percentile) errorSynchronicity w 10 ns (95th percentile) errorSynchronicity w 20 ns (95th percentile) error

0 100 200 300 400 500 600 Location error in cm

Conclusion and future work •  Synchronicity: 1.7 m accuracy from just TDoA

–  Single antenna APs; 20MHz channel bandwidth

•  Can augment other schemes including AoA

•  Significant accuracy improvement on higher bandwidth

•  Merge adjacent channel data for more accurate localization

22 Thank you!