When Machine Learning Meets Wi-Fi

59

Transcript of When Machine Learning Meets Wi-Fi

Page 1: When Machine Learning Meets Wi-Fi
Page 2: When Machine Learning Meets Wi-Fi

Robust Device-Free Indoor Localization

When Machine Learning Meets Wi-Fi

Steve Liu, VP R&D, Chief-Scientist

Samsung AI Center โ€“ Montreal &

Professor & William Dawson Scholar

McGill University

Nov. 5th, 2019

Disclaimer: Any views or opinions presented in this talk are personal and do not represent those of people, institutions, organizations that the presenter may or may not be associated with in professional or personal capacity.

Page 3: When Machine Learning Meets Wi-Fi

Outline

Robust Device-Free Indoor Localization

โ€ข Indoor Localization & Background

โ€“ Fingerprinting-based Device-Free Localization Approach

โ€ข A Major Challenge: Environment Changes

โ€ข AutoFi

โ€ข Experiments

โ€ข More Recent Advances: Toward Robustness in Real World

โ€ข Conclusions & Future Work Directions

3

Page 4: When Machine Learning Meets Wi-Fi

Indoor Localization & Background

Page 5: When Machine Learning Meets Wi-Fi

Why this Topic?

The next big thing!

โ€ข Internet of Things

โ€ข AI and Machine Leaning (=> Intelligent)

โ€ข Advanced Communications like Wi-Fi 6 & 5G (=> Connected )

5

Page 6: When Machine Learning Meets Wi-Fi

Indoor Localization

GPS signal not available

6

Page 7: When Machine Learning Meets Wi-Fi

Applications

What can we do with location information?

Indoor navigation

Smart home automation

Automated industry

7

Page 8: When Machine Learning Meets Wi-Fi

State-of-the-Art

Device-Oriented

โ€ข Pro: High accuracy: 10cm-level (e.g., [1],[2],[3])

โ€ข Con: Complicated designโ€“ Customized hardware

โ€“ Antenna arrays (10+ antenna)

โ€“ Usually heavy computational overhead

โ€ข Locating devices, NOT users

[1] M. Kotaru, K. R. Joshi, D. Bharadia, and S. Katti, โ€œSpotfi: Decimeter level localization using WiFiโ€, Sigcomm โ€˜15. [2] X. Li, D. Zhang, Q. Lv, J. Xiong, S. Li, Y. Zhang, and H. Mei, โ€œIndotrack: Device-free indoor human tracking with commodity WiFi,โ€ IMWUTโ€™ 17[3] K. Qian, C. Wu, Y. Zhang, G. Zhang, Z. Yang, and Y. Liu, โ€œWidar2.0: Passive human tracking with a single WiFi link,โ€ MobiSys โ€˜18.

8

Page 9: When Machine Learning Meets Wi-Fi

Why Device-free Localization?

People are not always carrying devices

โ€ข Cannot, or Do not want to carry

9

Page 10: When Machine Learning Meets Wi-Fi

State of the Art

Device-Oriented

Device-Free

โ€ข Triangulation methodsโ€“ Utilize userโ€™s blocking/reflection of the signals

โ€“ Calculating Time-of-Fight (ToF), Angle-of-Arrival (AoA), Propagation features, etc.

โ€“ Dedicated devices required

10

Page 11: When Machine Learning Meets Wi-Fi

Disadvantages of Dedicated Devices

โ€ข Costโ€“ Equipment cost

โ€“ Deployment

โ€“ Triangulation: positions of devices must be known in advance

โ€ข Scalability

โ€ข Range limitation (motion sensors, RFID, โ€ฆ)

โ€ข Blocked by walls

โ€ข โ€ฆ

Motion Sensors CamerasRFID

Tag & Reader

11

Page 12: When Machine Learning Meets Wi-Fi

State of the Art

Device-Oriented

Device-Free

โ€ข Triangulation methods

โ€ข Fingerprinting-based methodโ€“ Correlating environment features with human locations

โ€“ Environment features:

โ€ข Wireless signals, e.g., Wi-Fi

โ€ข Light or sound

โ€ข Magnetic field

โ€ข โ€ฆ

12

Page 13: When Machine Learning Meets Wi-Fi

Benefits of Wi-Fi based DfP Localization

Work through walls!

Available Everywhere!

13

Page 14: When Machine Learning Meets Wi-Fi

Wi-Fi Based DfP Localization

Fingerprinting: Associate Wi-Fi features such as CSI (Channel State Information) or

RSSI (Received Signal Strength Indication) with usersโ€™ locations.

14

Page 15: When Machine Learning Meets Wi-Fi

Wi-Fi Channel State Information

Channel State Information (CSI)

โ€ข Describe the Wi-Fi channel properties

โ€ข Sensitive to human locations

โ€ข Provides much more information than RSSI

โ€ข Readily available from commercial Wi-Fi cards

15

1 RSSI โ†’ 1 integer

56 subcarriers with CSI values(complex numbers) โ†’ 56 magnitudes + 56 phases

โ†’ 112 floating point numbers

Page 16: When Machine Learning Meets Wi-Fi

More Background on Wi-Fi CSI

Wi-Fi Multiple-Input Multiple-Output (MIMO)

e.g., Intel 5300 NICe.g., Linksys WRT160N16

Page 17: When Machine Learning Meets Wi-Fi

More Background on Wi-Fi CSI

MIMO in math

Received signals

Streams (Spatial Channels)CSI

Transmittedsignals

Noise

Source: http://www.sharetechnote.com/html/Communication_ChannelModel.html

๐’š = ๐‘ฏ ๐’™ + ๐’

17

Page 18: When Machine Learning Meets Wi-Fi

How does CSI look like?

CSI profiles at h11

โ€ข Captures the math representation of channel characteristics from transmitter antenna 1 to receiver antenna 1

โ€ข Time/frequency variant

Time (seconds)

Frequency(Subcarriers)

CSIMagnitude

18

Page 19: When Machine Learning Meets Wi-Fi

CSI Profiles are Location-Dependent

Subcarriers

CSI

Magnitude

19

Page 20: When Machine Learning Meets Wi-Fi

CSI Profiles to CSI Fingerprints

A CSI fingerprint of location L is a CSI profile associated with this location

โ€ข i.e., fingerprint = CSI profile at location L

20

Page 21: When Machine Learning Meets Wi-Fi

How to use ML for Localization?

Use CSI fingerprints to train ML classifiers

โ€ข Locations โ€“ labels (output), CSI profiles (input)

โ€ข SVM, Random Forest, KNN, Neural Networks, etc.

MLClassifiers

For training

21

Page 22: When Machine Learning Meets Wi-Fi

Classification for Localization

Trained models to classify the labels of new (online measured) CSI profiles

โ€ข Predicted labels = estimated locations

Trained ML

Classifiers

10 locations, (sub-)meter-level resolution - 99.7% accuracy.22

Page 23: When Machine Learning Meets Wi-Fi

A Major Challenge

- Environment Changes

Page 24: When Machine Learning Meets Wi-Fi

A Major Challenge

The changes in the indoor environment make the (old) fingerprints inconsistent with

the current situation (new).

24

Page 25: When Machine Learning Meets Wi-Fi

A Major Challenge

The CSI fingerprints will be โ€œcontaminatedโ€ by environment changes.

โ€ข The recorded fingerprints (old) no longer represent the changed environment (new)

25

Page 26: When Machine Learning Meets Wi-Fi

A Major Challenge

The CSI fingerprints will be โ€œcontaminatedโ€ by environment changes.

โ€ข The recorded fingerprints (old) no longer represent the changed environment (new)

Before After

26

Page 27: When Machine Learning Meets Wi-Fi

A Major Challenge

Localization accuracy drops significantly

โ€ข A case study:

Record the fingerprints again, and retrain the model?

27

Impractical: Inconvenient & time-consuming

Page 28: When Machine Learning Meets Wi-Fi

AutoFi

Page 29: When Machine Learning Meets Wi-Fi

Idea: Can We Reuse the Fingerprints?

Reconstruct vectors measured from the changed domain to the old fingerprint domain, so we can reuse the already trained model

๐‘€๐‘›๐‘’๐‘คโˆ’๐‘œ๐‘™๐‘‘

-------------------------------------------------------------------------------------

New vector domain(Contaminated CSI profiles)

Old vector domain(fingerprints)

๐‘€๐‘›๐‘’๐‘คโˆ’๐‘œ๐‘™๐‘‘

๐‘‰๐‘’๐‘š๐‘๐‘ก๐‘ฆ๐‘œ๐‘™๐‘‘ ๐‘‰๐‘1

๐‘œ๐‘™๐‘‘

๐‘‰๐‘’๐‘š๐‘๐‘ก๐‘ฆ๐‘›๐‘’๐‘ค ๐‘‰๐‘1

๐‘›๐‘’๐‘ค

29

Page 30: When Machine Learning Meets Wi-Fi

The Mapping Functions

Map CSI from the new domain to the old domain

Contaminated CSI profiles(locations unknown)

Unknown mapping How to determine?

๐‘‰๐‘1๐‘œ๐‘™๐‘‘ = ๐‘€๐‘1 ร— ๐‘‰๐‘1

๐‘›๐‘’๐‘ค + ๐‘›๐‘1,

๐‘‰๐‘2๐‘œ๐‘™๐‘‘ = ๐‘€๐‘2 ร— ๐‘‰๐‘2

๐‘›๐‘’๐‘ค + ๐‘›๐‘2 ,

โ‹ฎ

๐‘‰๐‘๐‘๐‘œ๐‘™๐‘‘ = ๐‘€๐‘๐‘ ร— ๐‘‰๐‘๐‘

๐‘›๐‘’๐‘ค + ๐‘›๐‘๐‘ .

Fingerprints in Old Domain(associated with locations)

30

Page 31: When Machine Learning Meets Wi-Fi

The Mapping Functions

Approximation: the mapping functions are the same for different locations

โ€ข Domain-to-domain mapping

๐‘€๐‘1 = ๐‘€๐‘2 = โ‹ฏ = ๐‘€๐‘๐‘ = ๐‘€.

Contaminated domain (New) Fingerprint domain (Old)

M: New โ†’Old

31

Page 32: When Machine Learning Meets Wi-Fi

The Mapping Function

Mapping function: M

๐‘‰๐‘1๐‘œ๐‘™๐‘‘ = ๐‘€ ร— ๐‘‰๐‘1

๐‘›๐‘’๐‘ค + ๐‘›๐‘1,

๐‘‰๐‘2๐‘œ๐‘™๐‘‘ = ๐‘€ ร— ๐‘‰๐‘2

๐‘›๐‘’๐‘ค + ๐‘›๐‘2 ,

โ‹ฎ

๐‘‰๐‘๐‘๐‘œ๐‘™๐‘‘ = ๐‘€ ร— ๐‘‰๐‘๐‘

๐‘›๐‘’๐‘ค + ๐‘›๐‘๐‘ .

32

Fingerprints in Old Domain(associated with locations)

Contaminated CSI profiles(locations unknown)

Page 33: When Machine Learning Meets Wi-Fi

The Mapping Functions

How to automatically determine the mapping function M?

๐‘‰?๐‘œ๐‘™๐‘‘ = ๐‘€ ร— ๐‘‰?

๐‘›๐‘’๐‘ค + ๐‘›?

Both unknown!

โ€ข In order to get M, we need to automatically detect a location p as a

reference point such that both ๐‘‰๐‘๐‘›๐‘’๐‘ค๐‘Ž๐‘›๐‘‘ ๐‘‰๐‘

๐‘œ๐‘™๐‘‘are known

33

Page 34: When Machine Learning Meets Wi-Fi

Determine M

Observation: The status of an empty area can be detected using a rule-based algorithm

โ€ข Reference point (state) = Detecting Empty

โ€ข Both before and after the environmental changes

โ€ข How? Using variance of CSI magnitude (can be detected automatically)

Associate ๐‘‰?๐‘›๐‘’๐‘ค with ?=Empty => ๐‘‰๐‘’๐‘š๐‘๐‘ก๐‘ฆ

๐‘œ๐‘™๐‘‘ = ๐‘€ ร— ๐‘‰๐‘’๐‘š๐‘๐‘ก๐‘ฆ๐‘›๐‘’๐‘ค + ๐‘›

34

Page 35: When Machine Learning Meets Wi-Fi

Contaminant Removal

๐‘€ = ๐‘€E๐‘š๐‘๐‘ก๐‘ฆ

Contaminated profile (raw CSI measured in the newly changed environment)

๐‘€

? ?

De-contaminated (purified) profileโ€ข Transformed CSIโ€ข In the old domain, i.e. when

there is no environ changes

Trained ML model

Determine Location

35

Page 36: When Machine Learning Meets Wi-Fi

Contaminant Removal Example

An example of contaminant removal result @ P1

Still some residual errors!

36

Page 37: When Machine Learning Meets Wi-Fi

Autoencoder

A neural network learns efficient data representation (encoding) of the input data

โ€ข Encoder

โ€ข Decoder

โ€“ To (re)generate data

37

Page 38: When Machine Learning Meets Wi-Fi

Training the Autoencoder

The autoencoder tries to reconstruct the input

โ€ข Training: Inputs and outputs are the same โ€“ the fingerprint profiles recorded

โ€ข Learns an efficient representation in the recorded domain (fingerprint profiles)

โ€ข Use BP to train

Fingerprint profiles

The encoding

Fingerprint profiles

38

Page 39: When Machine Learning Meets Wi-Fi

Denoising with the Autoencoder (Inference)

โ€ข The autoencoder denoises measured new CSI profiles

โ€“ Using the (already) purified profiles as input

โ€“ Old fingerprint domain features are identified through the coding/decoding

โ€“ Features related to the environment changes are diluted/omitted

New (purified) profiles Reconstructed profiles

The encoding

39

Page 40: When Machine Learning Meets Wi-Fi

Putting All Components Together

The architecture of AutoFi (Training in red, localization in blue).

40

Page 41: When Machine Learning Meets Wi-Fi

Experiments

Page 42: When Machine Learning Meets Wi-Fi

Experiment Setup

A Linksys WRT160N router, a laptop with Intel 5300 NIC.

โ€ข Wi-Fi traffic: 10 โ€“ 20 pings per second.

โ€ข Meter-level resolution (minimum distance 80cm)

42

Page 43: When Machine Learning Meets Wi-Fi

Experiment Setup

Introducing the โ€œcontaminantsโ€

43

Page 44: When Machine Learning Meets Wi-Fi

Experiment Results

Baseline, no mapping was applied, Random Forest (RF).

Opening windows Opening doors

Accuracy Mean Min Mean Min

Baseline 18.8% 0% 41.7% 0%

44

Page 45: When Machine Learning Meets Wi-Fi

Experiment Results

Using only contaminant removal technique

Opening windows Opening doors

Accuracy Mean Min Mean Min

Baseline 18.8% 0% 41.7% 0%

Contaminant Rmv 69.0% 0% 70.0% 0%

45

Page 46: When Machine Learning Meets Wi-Fi

Experiment Results

AutoFi (using both contaminant removal and autoencoder)

Opening windows Opening doors

Accuracy Mean Min Mean Min

Baseline 18.8% 0% 41.7% 0%

Contaminant Rmv 69.0% 0% 70.0% 0%

AutoFi 84.9% 47.6% 90.2% 71.3% 46

Page 47: When Machine Learning Meets Wi-Fi

Brief Summary

โ€ข Problem for robust localization

โ€“ Small variations in the environment may significantly contaminate the

fingerprints

โ€ข Solution - AutoFi

โ€“ Reuse the fingerprints and the trained ML model with a contaminant

removal technique

automatically maps the contaminated profiles back to the fingerprint

domain

โ€“ Utilize an autoencoder to further denoise the purified profiles

47

Page 48: When Machine Learning Meets Wi-Fi

More Recent Advances

- Toward Robustness in Real World

Page 49: When Machine Learning Meets Wi-Fi

More Robust Motion Detection

AutoFi used CSI variance as an indicator of human motion

โ€ข CSI variance is not a robust indicator of human motion

โ€ข In some empty rooms, CSI variance caused by noise can be even higher than that caused by human motion

โ€ข Simply applying white noise filtering does not work

โ€“ Because this also removes the variance caused by human motion

Our solution MoFi [to appear in RTSS@Work 2019]

49

Page 50: When Machine Learning Meets Wi-Fi

More Robust Motion Detection

โ€ข Robustness Study on Motion Detection

โ€“ Different users

โ€“ Different rooms

โ€“ Different AP/device placements

โ€“ Environment changes, e.g., moved furniture, open/close doors/windows

โ€“ Interference from other 2.4GHz devices, e.g., microwave ovens and Bluetooth beacons

50

Page 51: When Machine Learning Meets Wi-Fi

Addressing User Diversity

Different body shapes yield different CSI profiles

โ€ข A system trained on one user may not work for a new user

โ€ข Labelling CSI profiles for a new user is mostly impractical

โ€“ Requiring user involvement

โ€“ Time consuming

โ€ข Work under submission and review

51

Page 52: When Machine Learning Meets Wi-Fi

Wi-Fi Sensing

A new IEEE TIG (Topic Interest Group) on Wi-Fi sensing:

52

Page 53: When Machine Learning Meets Wi-Fi

Wi-Fi based Motion Detection

Photo credit: Linksys: https://www.linksys.com/us/linksys-aware/ 53

Page 54: When Machine Learning Meets Wi-Fi

Conclusions & Future Directions

Page 55: When Machine Learning Meets Wi-Fi

Conclusions and Future Directions

โ€ข Wi-Fi based device-free localization is very promising

โ€ข Robust ML solutions have been developed for real-life scenarios

โ€“ Robust to environment changes: AutoFi

โ€“ Robust to random noises: MoFi

โ€“ Robust to user diversity: Under submission

โ€ข Future: Many more to explore!

โ€“ Multi-user localization

โ€“ Fast bootstrap/adaptation

โ€“ โ€ฆ

55

Page 56: When Machine Learning Meets Wi-Fi
Page 57: When Machine Learning Meets Wi-Fi

BACKUP SLIDES

57

Page 58: When Machine Learning Meets Wi-Fi

Physics Behind This

Signals affected by human body

58

Page 59: When Machine Learning Meets Wi-Fi

Physics Behind This

Different body locations introduce different effects

59