localization11 aug 09

Upload: krishna-ram-budhathoki

Post on 04-Apr-2018

220 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/29/2019 Localization11 Aug 09

    1/14

    1

    1

    Shinsuke Hara, Ph. D.Gradate School o f Engineering, Osaka City University

    11/August/2009

    [email protected]

    Localization-1

    2

    Contents of Shins seminar

    Whats localization?Why localization now?Context-aware servicesLocation-based servicesImportance of localizationCategoryPrinciple

    Performance bounds

    3

    Whats localization?

    To estimate the 2-dimensional or 3-dimensionallocation of a target, such as a mobile user and node

    Location estimationPositioning

    Localization (technical term in robotics)

    Ranging is to estimate (measure) the distancebetween a pair of transceivers

    Usually, localization is based on multiple rangingwith different known locations, such as basestations and anchor nodes

    4

    Assumptions for localization

    Some nodes are available whose locations areknown in advance, which are called referencenodesor anchor nodes. In this seminar, we callthem anchor nodes.

    From a target node, anchor nodes can get someinformation related to the distance to the targetnode

    Examples are: TOA (time of arrival), TDOA (timedifference of arrival), AOA (angle of arrival), RSSI(received signal strength indication, receivedpower) and so on

  • 7/29/2019 Localization11 Aug 09

    2/14

    2

    5

    Anchornode #3

    Anchornode #1

    Anchornode #4Target node

    Anchornode #2

    1=g(d1)

    TOA (time of arrival)

    2=g(d2)

    3=g(d3)

    4=g(d4)

    6

    How can we measure TOA? (1)

    One-Way Ranging (synchronization, the same

    clock among nodes)

    time

    time

    0

    0

    Anchor node

    Target node

    d=cc: speed of light (3.0x108 m/s)

    Establishment of the synchronization is very difficult!

    7

    How can we measure TOA? (2)

    Round trip time measurement

    time

    time

    0Anchor node

    Target node

    Td(known in advance)

    TR

    =(TR-Td)/2

    d=c

    8tt

    t

    0

    1

    0

    1Anchornode

    (A)

    Targetnode(B)

    TTA

    TTA

    TRB

    TTB

    TTB

    TRB

    TRA

    How can we measure TOA? (3)

    Two-Way Ranging

  • 7/29/2019 Localization11 Aug 09

    3/14

    3

    9

    TTA+=TR

    B+tTR

    A-t=TTB+ 2)()( BR

    BT

    AT

    AR

    p

    TTTTt

    TWR does not need synchronization !

    How can we measure TOA? (4)

    10

    Anchornode #3

    Anchornode #1

    Anchornode #4Target node

    Anchornode #2

    TDOA (time difference of arrival)

    21=2-1 31=3-1

    41=4-1

    11

    How can we measure TDOA?

    Anchor nodes are synchronized and can talkwith each other

    time

    time

    0

    0

    Anchor node 1

    Anchor node 2

    time

    1

    2

    21=2-1

    Target node

    12

    Anchornode #3

    Anchornode #1 Anchor

    node #4

    Target node

    Anchornode #2

    AOA (angle of arrival )

    2

    1

    3

    4

  • 7/29/2019 Localization11 Aug 09

    4/14

    4

    13

    How can we measure AOA?

    A target node or anchor nodes need to havearray antenna

    14

    Pros and cons of TOA, TDOA and AOA

    Their performance has been long believed to bemuch better (than RSSI)They do not work in non-line-of-sight (NLOS)environments (tall buildings and people walkingaround)TWR or synchronization is required for TOAArray antenna is required for AOA

    15

    Anchornode #3

    Anchornode #1

    Anchornode #4Target node

    Anchornode #2

    P1=f(d1)

    RSSI (received signal strength indication)

    P2=f(d2)

    P3=f(d3)

    P4=f(d4)

    16

    RSSI is easily measurable

    Cellularphones

    Wireless LANterminals

    Many current wireless communication standardssupport RSSI measurement

    IEEE 802.15.4

    LQI (link quality indicator)8bit resolution,-173dBm to 82dBm(1dB step)

  • 7/29/2019 Localization11 Aug 09

    5/14

    5

    17

    Pros and cons of RSSI

    Its performance has been long believed to bemuch worse (than TOA), because of fadingand shadowingThey can work even in NLOS environmentsIt is simply implementableOWR is easily possible

    18

    Why localization now?

    What are the paradigms of next generation mobilecommunication systems?

    1G system: analogue (1980s)2G system: digital (1990s)3G system: multimedia (2000s)4G system: heterogeneous network & system

    and context-aware services

    5G system: green, learning,

    The word of contextmeans the overall situation in which an eventoccurs (The American Heritage Dictionary)

    19

    NGN

    ADSL WLAN

    WiMAX

    CPU capabilit y

    Traffic load condition

    Channel bandwidth

    Indoor

    Outdoor

    In motion

    Stationary

    3G

    Context in communicationsmeans the overall situationsurrounding any parts of the entities in a communication system

    Context-aware services

    20

    What is an essential difference of wireless from wired?

    End device (terminal) can move aroundThe location is worth estimatingServices with the location information are uniqueonly for wireless

    Location-based services

  • 7/29/2019 Localization11 Aug 09

    6/14

    6

    21

    Location-based services (1/3)

    Switch your cellphones to sleepmode

    My cell phoneisautomaticallyswitched to sleepmode

    My cell phonei sautomatically switchedto vibration modewhen getting on asilencecart

    (a) School (b) Transportation

    22

    You areapproaching toagood French restaurant My cell phoneis

    automaticallyswitched off whenentering atheater

    (c) Navigation andadvertisement

    (d) Theater

    Location-based services (2/3)

    23

    (e) Resource management

    My current basestation is busy withhandling a lot of traffic, but my nextbasestation which I can reach in afewseconds is not busy. So the systemhas just decided not to assign a channelwith lower datarateto mein thecurrentcell and to assign channel with muchhigher dataratein thenext cell

    Location-based services (3/3)

    (f) Advice

    24

    Importance of localization (1/4)

    GPS (Global Positioning System)Cellular (3G, 4G)WMAN (WiMAX)WLAN (WiFi)

    WPAN (Bluetooth)WSN (Zigbee, UWB)

    RF-ID (Passive, Active)

    Providing a variety of location-based services depends onthe accuracy of the estimated location of a target (node, cellphone and so on)

    To localize a target more accurately, we can use any kindsof wireless systems

    WiMAX

    GPS

    RF-ID

    WiFi

    Zigbee

    Bluetooth

  • 7/29/2019 Localization11 Aug 09

    7/14

    7

    25

    [1] G. Y. Delisle, Location Awareness and Positioning Methods for LocationBased Services in Wireless Communication Networks,IEEE VTC 2006-Fall.

    From always best connected to always best

    located [1]

    Importance of localization (2/4)

    References

    26

    Importance of localization (3/4)

    Wireless Enhanced 911 (FCC Third Order)

    Base station-based localization: error67%)error95%)

    GPS-based localization: error67%)error95%)

    All terminals are not equipped with GPS receiversCellular base stations and GPS satellites are not availablein indoor environments

    27

    Importance of localization (4/4)

    Increasing number of sessions in international conferencesIncreasing number of special issues in international journalsSpecialized conferences

    Specialized projects

    MELT 2008 (The 1st ACM International Workshop onMobile Entity Localization and Tracking in GPS-less

    Environments), San Francisco, CA, USA, 19 Sep. 2008WPNC 2009 (The 6th Workshop on Positioning,

    Navigation and Communication, Leibniz, Germany, 19Mar. 2009

    WHERE (Wireless Hybrid Enhanced Mobile RadioEstimators), FP7-ICT-2007-1, 01.01.2008-30.06.2010,5.551 Million Euros

    28

    Category

    Range-freeAPS [1]

    Range-based

    Deterministic (fingerprinting)

    Probabilistic

    RADAR [2]Parametric

    Non-parametric

    ML, LS, MAP

    BP-iterative

    BP [3]

    [1] D.Niculescu and B.Nath, Ad hoc positioning system (APS), IEEE Globecom2001, pp.2926 2931, Nov. 2001.

    [2] P.Bahl and V.Padmanabhan, RADAR: An in-building RF-based user locationand tracking system,IEEE Infocom2000, pp.775 - 784 Mar. 2000.

    [3] A.T.Ihler, et al., Nonparametric belief propagation for self-localization of sensornetworks,IEEE J SAC, vol.23, no. 4, pp.809-819, Apr. 2005.

    References

  • 7/29/2019 Localization11 Aug 09

    8/14

    8

    29

    Range-free localization (APS)

    Anchor #1Anchor #2

    Anchor #3Anchor #4

    Anchor #5

    3 hops

    3 hops

    3 hops

    4 hops

    5 hops

    Target

    The location of a target is estimated with the knownlocations of anchor nodes and the numbers of hops tothem 30

    Problems of APS

    Hop count does not correspond to physicaldistance

    The localization performance is very badThey believe ranging is difficult (ranging function isrealized in current wireless communicationstandards)

    31

    Deterministic localization (RADAR)

    a b c d

    Anchor #1

    Anchor #2

    Anchor #3

    RSSI (received signal strength indication)=received power

    Database[a, -20, -34, -50][b, -21, -32, -65][c, -34, -11, -45]

    RSSI at Anchor #1

    [?, -33, -15, -22]

    Find thebestmatching

    32

    Problems of RADAR

    Exhausted pre-measurement for making databaseis requiredOnce the room layout is changed, the currentdatabase becomes useless, so a new measurement

    is requiredTo include the effect of moving objects intodatabase is difficult

  • 7/29/2019 Localization11 Aug 09

    9/14

    9

    33

    Principle of localization

    Targetnode

    Targetnode

    Mobile

    terminal

    Unknownlocation

    (x, y, z)

    Anchornode

    Accesspoint

    Basestation

    Knownlocation

    (xn, yn, zn)

    WSNWLANCellular

    N>kfor the k-dimensional localization problemn=1, , N

    Mn: measurement vector at the n-th anchor node for a target

    Mn=[Mn1, , MnL], L: the number of measurements for the target

    34

    What is the measurement?

    The measurement is related to the location of the target node,(x, y, z), such as

    TOA (Time Of Arrival)TDOA (Time Difference Of Arrival) [1]AOA (Angle Of Arrival) [2]RSSI (Received Signal Strength Indication, received power)

    Localization: to estimate the location of the target node,=(x, y, z)

    Ranging: to estimate the distance to the target node

    Note:

    35

    An example of TOA-based method (1/2)

    2-dimensional localization: z=0, zn=0(parameter) vector to be estimated: =(x, y)One-shot measurement: L=1Measurement TOAn: Mn=tn (the propagation time

    between the mobile terminal andthe n-thbase station

    Distance: nn ctd

    c=3.0x108 m/sec

    [1] N. Priyantha, et al., CRICKET location-support system,Proc. ACM MOBICOM,Aug. 2003.

    [2] H. Kawano and M. Kawano, Development of a pedestrian navigation systemworking on the 2.4 GHz band,11th World Congress on Intelligent TransportSystems,Oct. 2004 (http://www.radio-com.jp/rcs_technology.html).

    References

    36

    Basestation 4

    Basestation 1

    Basestation 2

    Mobile terminal =(x, y)

    Basestation 3

    d1

    d3 d2

    d4

    NLOS

    An example of TOA-based method (2/2)

    Localization is based on several ranging from differentplaces (base stations)

    NLOS: non-line-of-sight

  • 7/29/2019 Localization11 Aug 09

    10/14

    10

    37

    How can we localize the target?

    p42

    p14

    p12

    The target must belocated at theweight of thetrianglep42-p12-p14

    38

    ctn

    LS (Least Squares) estimation (1/2)

    find which minimizes

    2

    1 |)(

    ~

    |)( N

    nnn dde

    ),( yxLS estimation problem [1]

    22 )()( nn yxxx

    n-th true distance: 22 )()()),(( nnn yyxxyxd

    n-th square (ranging) error:2|)(

    ~|)( nnn dde

    Sum of the square errors: )()(1

    N

    nnee

    Note:

    LS estimation is a nonlinear minimization problemLS estimation does not care about the distribution ofdn

    39

    LS (Least Squares) est imation (2/2)

    d4)(4 d

    Basestation 4

    2444 |)(|)( dde

    40

    ML (Maximum Likelihood) estimation (1/2)

    Assume that dn has a known conditional probability densityfunction (pdf) for a given :)(nd

    )|())(|( nnn dpddp

    The likelihood function on =(x, y) is given by

    N

    n

    ndpl1

    )|()(

  • 7/29/2019 Localization11 Aug 09

    11/14

    11

    41

    find which maximizes),( yxML estimation problem [1]

    ML (Maximum Likelihood) estimation (2/2)

    N

    n

    ndpL1

    )|(log)(

    Note:

    ML estimation is a nonlinear minimization problemML estimation does not care about the distribution of

    Log-likelihood function on

    42

    MAP (Maximum A Posteriori) estimation

    Assume that has a knownpdf: p()

    By Bayes theorem

    (|()(

    (|()|( pdp

    dp

    pdpdp n

    n

    nn

    find which maximizes),( yxMAP estimation problem [2]

    N

    nndpMAP

    1

    )(log)(

    Note:

    ML estimation is a nonlinear minimization problem

    43

    A Gauss ian case

    Assume that tn contains a Gaussian distributed error thus dnalso contains a Gaussian distributed error [3]:

    2

    2

    2

    ))(~

    (

    2

    1)(

    ~|( d

    nn dd

    d

    nn eddp

    find which maximizes),( yx

    ML estimation problem

    ML

    N

    n d

    nn Cdd

    L

    1

    2

    2

    2

    ))(~

    ()(

    Therefore,

    find which minimizes),( yx

    )())(~

    ( 2

    1

    eddN

    n

    nn

    LS estimationproblem !

    N

    n d

    nn ddN

    d

    N

    n

    nn eddpl1

    2

    2

    2

    ))((

    1 2

    1))(

    ~|()(

    44

    Concluding remarks

    LS and ML estimations are classical frequentistapproacheswhereas MAP is a Bayesian approach

    The concept of MAP estimation is totally different from that ofML estimation (The Bayesian approach gives a related butconceptually different treatment of the parameter estimation

    problem (LS and ML) [1])

    Only for a Gaussian case, LS estimation becomes equivalentto ML estimation

    The localization performance depends on how accurately wecan knowp(dn|) andp(), that is, the channel propagationcharacteristicand the system layout

    [1] L. Ljung, System Identification: Theory for the User, Prentice-Hall, 1999.[2] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006.[3] N. Patwari, et al., Relative location estimation in wireless sensor networks,

    IEEE Trans. Signal Process., vol.51, no.8, pp.2137-2148, Aug. 2003.

    References

  • 7/29/2019 Localization11 Aug 09

    12/14

    12

    45

    Performance bounds

    Target node location: (x, y, 0) (unknown)Unknown parameter: =(x, y)Anchor node location: (xn, yn, zn) (known) (n=1, , N)Measurement (RSSI, TOA and so on): Mn (L=1)

    True distance between the target and n-thanchor node:

    probability density function (pdf) ofMn given :Log-likelihood function on for the n-thanchor node: L

    n

    ()

    222 )0()()()( nnnn zyyxxd

    ))(|(log)( nnn dMpL

    nd ))(|( nn dMp

    )(nd

    46

    Log-likelihood function on : L()

    Fishers information matrix is defined as [1]

    N

    nnn

    N

    nn dMpLL

    11

    )(~

    |(log)()(

    2221

    1211

    2

    22

    2

    2

    2

    )()(

    )()(

    )(JJ

    JJ

    LyLxy

    Lyx

    Lx

    EJ F

    Taking into consideration:

    0)(

    ~)(

    n

    n

    d

    LE

    Cramer-Rao lower bound for ML estimation (1/6)

    47

    We have

    y

    d

    x

    dJJ

    y

    dJ

    x

    dJ

    nN

    n

    nn

    N

    n

    nn

    N

    n

    nn

    )(~

    )(~

    )(

    )(~

    )(

    )(~

    )(

    12112

    1

    2

    22

    1

    2

    11

    where

    2

    2

    )(~

    )()(

    n

    nn

    d

    LE

    Cramer-Rao lower bound for ML estimation (2/6)

    48

    Defining the inverse matrix ofJFas IF():

    1121

    1222

    211222112221

    1211

    1

    1

    )()(

    JJ

    JJ

    JJJJII

    II

    JI FF

    the Cramer-Rao lower bound on the ML estimation errorvariance is given by

    2

    11

    2

    1

    2

    1 1

    22

    22112

    )(~

    )(~)(

    )(~)(

    )(~)(

    )(~)(

    )(~)(

    )]var()min[var()(

    N

    nn

    n

    n

    nn

    N

    nn

    nn

    N

    nn

    nn

    N

    n

    N

    nn

    nn

    n

    nn

    CRLB

    d

    yy

    d

    xx

    d

    yy

    d

    xx

    d

    yy

    d

    xx

    IIyx

    Cramer-Rao lower bound for ML estimation (3/6)

  • 7/29/2019 Localization11 Aug 09

    13/14

    13

    49

    0

    x

    z

    y

    (x, y, 0)

    (xn, yn, zn)

    (xn, yn, 0)

    n

    n

    From the target node to the n-thanchor node, define the elevationangle and the direction angle on thex-yplane as n and n, respectively

    nd~

    Cramer-Rao lower bound for ML estimation (4/6)

    50

    The Cramer-Rao lower bound can be re-written as

    N

    i

    N

    ijji

    ji

    ji

    N

    nn

    n

    CRLB

    1 1 22

    2

    1 2

    2

    coscos

    )(sin)()(

    cos

    1)(

    )(

    Note:

    when the target and anchor nodes are on a line(1=2=, , =N)

    )(2CRLB

    Cramer-Rao lower bound for ML estimation (5/6)

    51

    For L measurements case, the Cramer-Rao lower boundcan be re-written as

    1

    1 1 22

    2

    1 2

    2

    coscos

    )(sin)()(

    cos

    1)(

    )(

    L

    LN

    i

    N

    ijji

    ji

    ji

    N

    nn

    n

    CRLB

    Cramer-Rao lower bound for ML estimation (6/6)

    52

    Assuming we know a priori knowledge on the location ofthe target node asp(), the priori information matrix iscalculated as

    )(log)(log

    )(log)(log

    )(

    2

    22

    2

    2

    2

    py

    pxy

    p

    yx

    p

    xEJ P

    Bayesian Cramer-Rao lower bound for MAP estimation (1/2)

  • 7/29/2019 Localization11 Aug 09

    14/14

    14

    53

    the Bayesian Cramer-Rao lower bound on the MAPestimation error variance is given by

    DefiningJT()=JF()+JP() and its inverse matrix as IT():

    2221

    12111

    ''

    '')()(

    II

    IIJI TT

    22112 '')]var()min[var()( IIyxBCRLB

    Note:

    When no priori knowledge on the targets location is available,the BCRLB (MAP estimation) becomes equivalent to theCRLB (ML estimation), because JP()=0

    [1] H.L.VanTrees, Detection, Estimation, and Modulation Theory, Part I, JohnWiley & Sons, 1968.

    References

    Bayesian Cramer-Rao lower bound for MAP estimation (2/2)