Geolocation techniques

911 views

Published on

Published in: Technology, Business
0 Comments
4 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
911
On SlideShare
0
From Embeds
0
Number of Embeds
9
Actions
Shares
0
Downloads
0
Comments
0
Likes
4
Embeds 0
No embeds

No notes for slide

Geolocation techniques

  1. 1. Chapter 2Ranging and Localization in HarshMultipath EnvironmentsIn this chapter, we will first introduce the basics of geolocation techniques that arebased on Time of Arrival (TOA), Time Difference of Arrival (TDOA), Angle ofArrival (AOA), and Received Signal Strength (RSS). Then we introduce the majorchallenges to accurate localization: multipath propagation and non-line-of-sightconditions where we will focus on the two most popular ranging techniques, TOAand RSS, and evaluate how the accuracy of localization is affected by thesephysical challenges. We will further highlight the relationship between the accu-racy of estimation and the signal to noise ratio and bandwidth parameters throughthe well-known Cramer-Rao Lower Bound (CRLB) equations. Finally, we willintroduce measurement and modeling of the RSS/TOA ranging that will highlightthe impact of multipath and NLOS on the accuracy of ranging systems.2.1 Basics of GeolocationClassical geolocation techniques (non-survey based) depend on geometrical rela-tionships between the coordinates of the reference points (satellites in GPS technology)and the associated range/angle measurements. Typically, reference points are wirelessdevices with known location information (e.g. x- and y-coordinates) either pre-pro-grammed or obtained through GPS. The mobile device (seeking its own positioninformation) exchanges RF signals with the reference points to estimate the distance orangle to each of the reference points. Equipped with the range measurements and thecoordinates of the reference points, the mobile device can solve for the unknownposition through a variety of techniques (geometrical, optimization, etc.). The accu-racy of the location information is affected by three major factors: the accuracy of thereference points’ position, the accuracy of range/angle estimates, and the geometricalconfiguration of the reference points and the unknown position. The non-surveygeolocation techniques computes location estimates through two steps: range/angleestimation and tri-lateration/angulation. Figure 2.1 illustrates the two-step procedure.C. Gentile et al., Geolocation Techniques, DOI: 10.1007/978-1-4614-1836-8_2, 17Ó Springer Science+Business Media New York 2013
  2. 2. 18 2 Ranging and Localization in Harsh Multipath Environments RF Signal Distance or Angle Position Estimates Estimate r1 (t ) ˆ d 1 or ˆ1 θ = (x, y) ˆ d 2 or ˆ2 T r2 (t ) ˆ ˆ Range/Angle Lateration/ Estimation Angulation Subsystem Subsystem rn (t ) ˆ d n or ˆnFig. 2.1 Classical geolocation system. Range or angle information is extracted from received RFsignals. Location is then estimated by lateration/angulation techniques In this section we will introduce the most popular geolocation techniques: TOA,TDOA, AOA, and RSS and provide an evaluation of the achieved accuracythrough the well-known Cramer-Rao Lower Bound (CRLB) analysis.2.1.1 TOA-Based TechniquesOnce distance/range measurements to at least 3(4) reference points are availablethe 2(3)-dimensional position estimate can be obtained. The set of distancemeasurements from the reference points to the mobile terminal forms a set ofnonlinear equations that can be solved to estimate the position. Here, it is assumedthat the mobile terminals exchanging range measurements are time synchronizedand that they are all in LOS condition (no obstruction between the mobile deviceand the base stations). Figure 2.2 illustrates the basic concept of tri-lateration. The range measurements can be used to estimate the position of a mobile devicethrough several techniques that are generally grouped under Maximum Likelihood(ML) and Least-Squares (LS) Techniques. In ML techniques, the solution is theposition that maximizes the conditional probability density function or À Á ^ ¼ arg max P djh h ^ ð2:1Þ hwhere ^ ¼ ½^; ^ŠT and h ¼ ½x; yŠT are the estimated and true position coordinates, h x y ^respectively. d ¼ d þ w is the measured/estimated distance vector to each base  à ^ ^ ^ ^ Tstation or d ¼ d1 d2 . . . dnB , w is zero-mean Gaussian measurement noiseand nB is the number of base stations. Assuming that the noise measurements areindependent and identically distributed (i.i.d), then the conditional distribution isgiven by
  3. 3. 2.1 Basics of Geolocation 19Fig. 2.2 TOA-based tri-lateration. Rangemeasurements to at least threebase stations make up a set ofnonlinear equations that canbe solved to estimate theposition of a mobile device. ˆ d1 ˆ d2Black points are base stationswith a priori known positioninformation while theintersection of the circles isthe position estimate of themobile device ˆ d3 ( À Á2 ) Á Y 1 À nB ^ di À di ^ P djh ¼ pffiffiffiffiffiffiffiffiffiffi exp À : i¼1 2pr2 i 2r2iwhere r2 is the variance of the ith measurement noise. There are two major iproblems with this ML approach. The first is that conditional PDF requires theknowledge of the exact distances, which is not available in practice. The second isthat solving for the position using the maximization approach requires a searchover all possible locations which is neither practical nor computationally efficient(Guvenc and chong 2009). There are also some variants of the original MLtechnique which are the two-step ML and the approximate ML (AML). Theinterested reader can find more details about the ML techniques in Guvenc et al.(2006), Chan and Ho (1994). The other class of TOA-based localization algorithms is based on the LStechniques. The range measurements to the reference points form a set of nonlinearequations of which the solution is the mobile position. The LS techniques arefurther subdivided into nonlinear LS (NL-LS) and the linearized LS (L-LS). TheNL-LS technique estimates the position by minimizing a residual error function(Caffery and Stuber 1998) or ( ) X À nB Á2 ^ ¼ arg minfR ðhÞg ¼ arg min h ^ bi di À kh À hi k : ð2:2Þ es h h i¼1 Thus the residual, Res ðhÞ, is a measure of error between the measured distances,^di , and the estimated distance obtained from computing the Euclidean distancebetween the reference points and the estimated position, kh À hi k. bi is a weightthat can be used to emphasize range estimates which is proportional to the degreeof confidence in the measurement. The L-LS solution is obtained by linearizing thenonlinear equations formed by the nB distances given by
  4. 4. 20 2 Ranging and Localization in Harsh Multipath Environments 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3 ðx À x1 Þ2 þðy À y1 Þ2 6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 6 7 6 ðx À x2 Þ2 þðy À y2 Þ2 7 FðhÞ ¼ 6 6 7 7 ð2:3Þ 6 . . 7 . 4 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 2 2 ðx À xnB Þ þðy À ynB Þwhere ½xn ; yn Š are the coordinates of the nth base station. The linearization isobtained through the well-known Taylor series expansion around h0 given byFðhÞ % Fðh0 Þ þ Jðh À h0 Þ Kay (1993) where Jðh À h0 Þ is the Jacobian of Fevaluated at h0 and it is given by " ofn #T of1 of2 ox ox . . . oxB J ¼ of1 of2 ofn ð2:4Þ oy oy . . . oyB h¼h0and the L-LS solution (mobile position estimate) is then given by Kay (1993) ^ À ÁÀ1  à ^ h ¼ h0 þ JH J JH d À Fð h 0 Þ ð2:5Þwhere H is the Hermitian operation. Typically, the accuracy of localization isaffected by the accuracy of the base station location; the statistics of the rangemeasurements and the geometry of the base stations with respect to the mobileterminal. The performance of TOA-based localization can be examined by eval-uating the Cramer-Rao Lower Bound (CRLB), which provides the lower bound onthe variance of the estimate or Kay (1993) h i ^ E ðh À hÞ2 ! IðhÞ ð2:6Þwhere IðhÞ is the Fisher Information Matrix (FIM) and E½Š is the expectationoperation. The FIM is given by Kay (1993) 2 # T # o ^ o ^ Á o ln f ðdjhÞ ^ IðhÞ,E ln f ðdjhÞ ¼E ln f ðdjhÞ ð2:7Þ oh oh oh ^ ^where f ðdjhÞ is the joint PDF of d condition on the unknown parameters. Themeasured distances are modeled by ^ d¼dþw ð2:8Þwhere d is the vector containing the actual (exact) distances between the mobiledevice and the BS and w is the zero-mean Gaussian measurement noise. Since thejoint PDF is a function of d which is a function of h, then from the chain rule
  5. 5. 2.1 Basics of Geolocation 21 o ^ od o ^ ln f ðdjhÞ ¼ Á ln f ðdjdÞ: ð2:9Þ oh oh od So (2.7) can be rewritten as T # od ^ Á od Á ln f ðdjdÞ ^ IðhÞ ¼ E Á ln f ðdjdÞ oh oh T # T od o o od ð2:10Þ ¼ E ^ ln f ðdjdÞ ^ ln f ðdjdÞ oh od od oh IðhÞ ¼ JId JTwhere J is the Jacobian given in (2.4) or explicitly 2 xÀxnB 3T xÀx1 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi . . . pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2 6 ðx1 ÀxÞ þðy1 ÀyÞ ðxnB ÀxÞ2 þðynB ÀyÞ2 7 J¼4 yÀy1 yÀynB 5 ð2:11Þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi . . . pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2 2 2 ðx1 ÀxÞ þðy1 ÀyÞ ðxnB ÀxÞ þðynB ÀyÞor alternatively ! cos /1 ... cos /nB J¼ ð2:12Þ sin /1 ... sin /nBwhere /n is the angle between the mobile device and the nth BS. The joint PDF ofthe distance measurements is given by ^ 1 1 ^ T À1 ^ f ðdjdÞ ¼ exp À ðd À dÞ R ðd À dÞ ð2:13Þ ð2pÞnB =2 jRj1=2 2where R is the covariance. Id can then be easily derived and it is given by Id ¼ RÀ1 ¼ diagðrÀ2 ; rÀ2 ; . . .; rÀ2 Þ 1 2 nB ð2:14Þ The CRLB for the mobile device position is then given by h i À ÁÀ1 IðhÞÀ1 ¼ JId JT : ð2:15Þ 2Â2 Another popular metric to characterize the accuracy of localization is theGeometric Dilution of Precision (GDOP) which describes the amplification of theerrors in range measurements to the location error (Patwari et al. 2003) and it isgiven by qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r2 þ r2 x y GDOP ¼ ð2:16Þ rr
  6. 6. 22 2 Ranging and Localization in Harsh Multipath Environmentswhere r2 and r2 the variances of the position estimate and rr is the standard x ydeviation of the range measurement error. An alternative expression for the GDOPcould be derived to emphasize the geometrical relationship between the BSs andthe mobile device Spirito (2001) sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi nB GDOPðnB ; /Þ ¼ À Á2 ð2:17Þ Ri Rj; j [ i sin /ij where /ij is the angle between the ith and jth BSs. Although the CRLB derivations in this subsection assumed single-path idealpropagation (simplified zero-mean Gaussian noise model) it can provide a startingpoint to evaluate the performance and understand the main factors that can affectthe accuracy. Different CRLB derivations that address the NLOS problem can befound in Qi et al. (2006), Dardari et al. (2006), Shen et al. (2007). The accuracy ofthe TOA-based techniques relies heavily on the measurement noise and themultipath condition of the channel. Thus the CRLB will only be meaningful whenthe models are realistic in that they reflect the actual propagation conditions. Inaddition it is common to assume that the BS and the mobile device are synchro-nized, but this is not the case in practice.2.1.2 TDOA TechniquesTime Difference of Arrival (TDOA) technique is based on the idea that theposition of the mobile device can be determined by examining the difference intime at which the signal arrives at multiple reference points (Liu et al. 2007).Adopting this technique is useful in practical scenarios where synchronizationbetween mobile devices is not available. Each TDOA measurement constrains thelocation of the mobile device to be on a hyperboloid with a constant range dif-ference between the two reference points. For two-dimensional position estimationthree reference points are required. Figure 2.3 illustrates the localization techniquebased on TDOA measurements. A TDOA measurement between BS1 and BS2 can be given by Sayed et al.(2005) t21 ¼ ðt2 À t0 Þ À ðt1 À t0 Þ ¼ t2 À t1 ð2:18Þwhere t0 is the clock time of the mobile device, t1 and t2 are the TOA between themobile device and BS1 and BS2 respectively. The equation can be written in termsof distance through speed of light scaling or d21 ¼ ðt2 À t1 Þc. Thus the time dif-ference (or range difference) is d21 ¼ d2 À d1 where d2 ¼ ðx2 À xÞ2 þðy2 À yÞ2 and 2 2 2 2d1 ¼ ð xÞ þ ð yÞ . Without loss of generality, the latter equation is valid with theassumption that the x- and y-coordinates of BS1 are ð0,0Þ. The range difference
  7. 7. 2.1 Basics of Geolocation 23Fig. 2.3 TDOA localization.At least three BS are requiredfor two-dimensional d 2 − d1localization. The time (range)differences d2 À d1 and d3 À BS2d1 form two hyperboloids ofwhich the intersection d2(solution) is the estimated BS1position d1 d3 BS3 d3 − d1equation can be rearranged to d21 þ d1 ¼ d2 . The TDOA equation can then beobtained by squaring both sides or ðd21 þ d1 Þ2 ¼ d2 ¼ x2 þ x2 À 2x2 x þ y2 þ y2 À 2y2 y 2 2 2 ð2:19Þ Using K2 ¼ x2 þ y2 the above equation simplifies to 2 2 2 ðd21 þ d1 Þ2 ¼ K2 À 2x2 x À 2y2 y þ x2 þ y2 2 ð2:20Þwhich can be further rearranged to solve for the unknowns or 1À 2 2 Á Àx2 x À y2 y ¼ d21 d1 þ d À K2 : ð2:21Þ 2 21 Two equations are required to solve for the two unknowns and the secondTDOA equation between BS3 and BS1 can be similarly obtained 1À 2 2 Á Àx3 x À y3 y ¼ d31 d1 þ d31 À K3 : ð2:22Þ 2 The equations can be arranged in matrix form given by Sayed et al. (2005) Hh ¼ d1 a þ b ð2:23Þ ! 2 2 ! Àd21 K2 À d21where a ¼ ,b¼1 2 2 . Solving for h we have Àd31 2 K3 À d31 h ¼ d1 HÀ1 a þ HÀ1 b: ð2:24Þ Extension to more reference points and three dimensions is trivial and moredetails can be found in Sayed et al. (2005). The performance of TDOA-based localization can be similarly examined byevaluating the CRLB. A similar derivation of the CRLB for TDOA localization
  8. 8. 24 2 Ranging and Localization in Harsh Multipath EnvironmentsFig. 2.4 AOA positioning ( x , y)(angulation). The AOAestimate from 2 base stationsto the mobile terminal can beused to estimate the position (x1, y1 ) φ2 φ1 (x2 , y2 )follows from (2.7). In fact it can be shown that the TDOA CRLB is given by Qiet al. (2006) h i À ÁÀ1 ITDOA ðhÞÀ1 ¼ JTDOA ITDOA JTTDOA ð2:25Þ 2Â2where 0 1 cos /1 cos /2 ÁÁÁ cos /nB JTDOA ¼ @ sin /1 sin /1 ÁÁÁ sin /nB A ð2:26Þ 1 1 ÁÁÁ 1 ITDOA ¼ ITOA : ð2:27Þwhere /n is the angle between the mobile device and the nth BS.2.1.3 AOA-Based TechniquesLocalization using angle-of-arrival is simpler than time-based techniques in thatonly two angle measurements are required, as opposed to three range measure-ments, in order to estimate the two-dimensional position. However the challenge ispresented when obtaining accurate angle of arrival estimation using wirelessdevices. In typical scenarios, the base stations are equipped with K antenna arrayelements spaced by D which are capable of estimating the angle of arrival which isthen used to locate the mobile device. Figure 2.4 illustrates the basic concept ofAOA localization. The relationship between the coordinates and the angles is given by y À y1 y À y2 ¼ tanð/1 Þ ¼ tanð/2 Þ ð2:28Þ x À x1 x À x2 These equations can be combined to estimate the position of the mobile ter-minal as Dempster (2006) !À1 ! tan /1 À1 x1 tan /1 Ày1 h¼ ð2:29Þ tan /2 À1 x2 tan /2 Ày2
  9. 9. 2.1 Basics of Geolocation 25 The CRLB for AOA can be similarly obtained from the formulation in (2.6) and(2.7), but with specific models for the angle measurements. In practice, the antennaarray is capable of measuring a function of the angle or Qi et al. (2006) un ¼ un ð/n Þ þ wn ^ ð2:30Þwhere n is the index identifying the BS and wn is a zero mean Gaussian noise witha variance given by Qi et al. (2006) À1 daH ðu Þ dan ðun Þ r2 ¼ 2! Á n n Á w ð2:31Þ dun dunwhere an ðun Þ is the steering vector for a specific antenna array configuration and !is the Signal to Noise Ratio (SNR). For an antenna array with K elements spacedby D the steering vector is given by an ðun Þ ¼ ½ 1 expðiun Þ . . . expðiðK À 1Þun Þ ŠT ð2:32Þwhere un ¼ 2pD cos /n . The variance of the estimation error is then given by Qiet al. (2006) 3 r2 ¼ w : ð2:33Þ KðK þ 1Þð2K þ 1Þ! Given the above model parameters of the AOA localization system the CRLBcan be given by Qi et al. (2006) h i À ÁÀ1 IAOA ðhÞÀ1 ¼ JAOA IAOA JT AOA ð2:34Þ 2Â2where KðK þ 1Þð2K þ 1Þ IAOA ¼ diagð !1 !2 . . . !nB Þ ð2:35Þ 3and JAOA ¼ 2pcD 2 2 2 ! 1 1 1 d1 ðsin /1 Þ d2 ðsin /2 Þ ÁÁÁ dnB ðsin /nB Þ Á À d11 cos /1 sin /1 À d12 cos /2 sin /2 ÁÁÁ À d1 cos /nB sin /nB nB ð2:36Þ The performance of AOA positioning techniques in LOS conditions is satis-factory. However, in severe NLOS multipath conditions the reliability and accu-racy of AOA techniques suffers considerably. As a result in these unfavorablepropagation conditions, TOA- or RSS-based techniques are preferred. Further-more, hybrid positioning techniques can be used to incorporate the advantages oftwo different techniques which usually outperform the individual techniques.
  10. 10. 26 2 Ranging and Localization in Harsh Multipath Environments2.1.4 Received Signal Strength LocalizationLocalization using Received Signal Strength (RSS) is very similar to TOA-basedtechnique in that the distances to nB base stations are used in a tri-laterationapproach to estimate the position. The difference is the method in which thedistance is estimated. For a mobile device and nB base stations, the unknownlocation can be estimated using the LS method similar to that of the TOA pre-sented in (2.5) or ^ À ÁÀ1  à ^ h ¼ h0 þ JH J JH dRSS À Fðh0 Þ ð2:37Þ The difference between (2.37) and (2.5) is the estimated distance vector. For RSS-based localization the distance can be estimated through the power–distance rela-tionship that is very well known for wireless propagation in different environments.The RSS between the mobile device and the nth base station is modeled by PdBm ¼ À10c log10 dn þ Sn r ð2:38Þwhere c is the pathloss exponent (governing the rate of power decay with dis-tance), Sn is the log-normal shadow fading component with variance r2n and dn is Sthe distance between the mobile devices and the nth base station. The ML estimate ^of the distance is given by dn ¼ 10ðÀPr Þ=ð10cÞ Patwari et al. (2003). Then the dis-tance vector in (2.37) is given by dRSS ¼ ½d1 ; d2 ; . . .; dnB ŠT . The CRLB for RSS- ^ ^ ^ ^based localization can be similarly derived from (2.6) to (2.7) (Qi et al. 2006) h i À ÁÀ1 IRSS ðhÞÀ1 ¼ JRSS IRSS JT RSS ð2:39Þ 2Â2where IRSS ¼ diagð rÀ2 S1 rÀ2 S2 ... rÀ2 Þ S nB ð2:40Þand 0 cos /nB 1 cos /1 cos /2 10cc @ d1 d2 ÁÁÁ d nB JRSS ¼ Á sin / A: ð2:41Þ sin /2 sin /nB ln 10 1 ÁÁÁ d1 d2 d nB2.2 The Multipath ProblemThe presence of multipath fading in harsh propagation environments can have asignificant impact on the performance of TOA-, RSS-, or AOA-based ranging andlocalization systems. Multipath is the reception of multiple copies of the trans-mitted signal—each arriving from different propagation paths—which combine in
  11. 11. 2.2 The Multipath Problem 27 (a) (b) (c)Fig. 2.5 LOS multipath channels. a Outdoor open space—single bounce model, b urban LOS,c indoor LOSeither a constructive or destructive manner that distorts the received signal. Thetransmitted signal undergoes reflections and diffractions along different propaga-tion paths to the receiver. At the receiver, replicas of the transmitted signal arriveattenuated, phase-shifted, and time-delayed. For RSS-based systems, multipathcauses the well-known fast fading phenomenon, where the received power in agiven location fluctuates significantly due to constructive and destructive inter-ference of incoming multipath signals. For TOA-based systems, the multipathimpacts the distance estimation directly by adding a random bias to the estimation.In this section, we will introduce the multipath problem and highlight its impact onRSS- and TOA-based ranging/localization systems. In order to appreciate the impact of multipath, it is important to analyze it inLOS environments, since multipath is the major error contributor. LOS propaga-tion can behave drastically different based on the environment. For example,performance in outdoor open-field LOS, outdoor urban LOS, and indoor LOS canexhibit different TOA estimation behavior. In outdoor open-field LOS, the directpath between the transmitter and receiver is unobstructed and there is at least asingle ground reflection at the receiver. In urban LOS or indoor LOS, there may bemany signals arriving at the receiver that were reflected or diffracted from thesurrounding buildings or objects. Figure 2.5 illustrates different possible LOSmultipath scenarios. In outdoor open space, the multipath structure is mainly composed of the directpath signal and a single-bounce ground reflection (see Fig. 2.5a). In urban LOS,reflections from the surrounding buildings make up the multipath environment (seeFig. 2.5b). The density of the buildings and surrounding obstacles will dictate thestructure of the multipath environment. Finally, in indoor LOS environments,
  12. 12. 28 2 Ranging and Localization in Harsh Multipath Environmentsthe multipath structure can be significantly different as there are reflections fromthe many cluttering objects and also reflections from walls, doors, and windowswith closer interarrival of multipath components at the receiver (see Fig. 2.5c).This creates an environment that is very different from the urban environment. Formally, the multipath can be modeled by Lp X hðsÞ ¼ ak ej/k dðt À sk Þ ð2:42Þ k¼1where Lp is the number of MPCs, ak and, /k and sk are amplitude, phase andpropagation delay of the kth path, respectively (Pahlavan and Levesque 2005;Rappaport 1996). The received waveform is then given by rðtÞ ¼ hðtÞ Ã sðtÞ wheresðtÞ is the transmitted signal waveform and (*) is the convolution operator.2.2.1 TOA-Based Ranging in LOS Multipath ChannelsThe basic idea behind TOA-based ranging is to estimate the distance between atransmitter and a receiver through measuring the signal propagation delay. For atransmitter at location ðx1 ; y1 Þ and a receiver at location ðx2 ; y2 Þ the Euclidean qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffidistance is given by d ¼ ðx1 À x2 Þ2 þ ðy1 À y2 Þ2 . In practice the distance can becalculated from the speed of light/propagation delay relationship given by c ¼ d=s,where c is the speed of signal propagation (in free space c ¼ 3e8 m/s) and s is thepropagation delay. But in realistic applications, the propagation delay estimates arealways corrupted by noise—additive white Gaussian noise (AWGN). Thus, the ^measured distance can be written as d ¼ c  s þ w ¼ d þ w. Here w is a zero-meanGaussian noise. In practice, the delay can be estimated using two methods: one-wayTOA ranging or two-way TOA ranging. The latter requires no synchronization and itis the basic ranging technique proposed in IEEE 802.15.4a (IEEE 802.15.TG4a).The former requires strict synchronization since the distance is estimated from thereceived waveform. This is practically challenging for two reasons. The first is thatextracting the TOA of the first path arrival is difficult (Lee and Scholtz 2002; Guvencand sahinoglu 2005). The second is that synchronization of wireless devicesin multipath environments is very difficult to achieve and is in fact an open researcharea. The main challenges are due to the clock drift over time and the effect oftemperature and humidity on the accuracy of clock frequency (Sundararaman et al.2005). Two-way TOA ranging techniques are the most popular due to the fact thatthey do not require synchronization and the protocols are very simple. For treatmentof one-way TOA ranging further details can be found in (Guvenc and sahinoglu2005). Two-way TOA ranging is achieved by noting the time that the rangingreference signal is sent out with the time it takes to receive it. Figure 2.6 illustrates anexample where Device 1 is attempting to estimate the distance to Device 2.
  13. 13. 2.2 The Multipath Problem 29Fig. 2.6 Two-way TOA Device 1 Device 2ranging. Devices 1 and 2exchange transmit and tTX1receive time information. tRX 2With these four time stamps,the propagation delay tTX 2(distance) between thedevices can be estimated tRX1 Device 1 initiates the two-way ranging by sending a ranging packet (signal) toDevice 2 and noting the time as tTX1 . Device 2 receives the signal at tRX2 and preparesits own ranging signal (after a processing delay) and sends out a response rangingsignal at time tTX2 . Finally Device 1 receives the response at tRX1 . Given that Device2 shares the time stamp information tRX2 and tTX2 with Device 1 it is now possible toestimate the propagation delay (distance) between the two devices by ttotal À troundÀtrip ðtRX1 À tTX1 Þ À ðtTX2 À tRX2 Þ s¼ ¼ ð2:43Þ 2 2where ttotal ¼ 2s þ troundÀtrip is the total time it takes for the two-way ranging andtroundÀtrip is the round-trip delay at Device 2. The assumptions regarding this two-way TOA ranging are overly simplistic and not valid in practice. In reality, theclocks of the two devices are not synchronized and not perfect. This means thatwith time the clocks will drift and the delay estimation will not be accurate(biased). Recently, researchers have investigated this problem and proposed somepractical techniques to estimate the delay in non-ideal scenarios (clock drift andbias) (Zheng and Wu 2010; Wu et al. 2011). The performance of TOA estimation in single path AWGN ideal scenario isusually analyzed using the Cramer Rao Lower Bound, which is a statisticalapproach to quantifying the variance of TOA estimation. Essentially any algo-rithm, in theory, can achieve the CRLB given that both the CRLB and algorithmfollow the same assumptions (for example LOS single path model and same noisevariance). The variance of TOA estimation r2 is bounded by the CRLB given TOAby Gezici et al. (2005), 1 r2 ! TOA ð2:44Þ 2 B2 8p2 !TBf0 1 þ 2 12f0where T is the signal observation time, ! is the SNR, f0 is the frequency ofoperation, and B is the system bandwidth. This relationship highlights that theaccuracy of TOA estimation can be improved by either increasing the SNR—sincehigher signal level will enable the estimation of the direct path signal with greateraccuracy—or increasing the system bandwidth—since higher system bandwidth
  14. 14. 30 2 Ranging and Localization in Harsh Multipath Environments Multipath error Relative Amplitude Time delay (s)Fig. 2.7 Power delay profile highlighting the multipath error corrupting TOA-based rangeestimatesmeans higher time–domain resolution leading to better range estimates.The increase in time–domain resolution of the channel can be attributed to narrowertime–domain signals/pulses. This makes it possible to discriminate or resolve thedifferent multipath arrivals and improve the TOA estimation. Multipath signals(especially in dense cluttered environment) tend to arrive fairly close to the directpath. If the interarrival time between the multipath components is much smaller thanthe time–domain resolution of the system (low bandwidth systems) then at thereceiver those multiple signals will combine to create a new cluster. The TOAestimate (from the receiver’s point of view) will then be the peak of the cluster. Inorder to clarify this phenomenon, Fig. 2.7 illustrates a power delay profile exampleand the resulting envelope. A power delay profile is a representation of the channelimpulse response where the power from different arrival paths can be measured andanalyzed. In the figure there are ten multipath components where the first multipathcomponent is the strongest and, in this case, is the LOS or direct path. The multipathcomponents arriving after the direct path fall in close proximity to each other(because of the nature of the propagation environment). For this narrowband system,the multipath components arrive and combine (due to low time–domain resolution)and appear at the receiver as four multipath components (the peaks of the blueenvelope). As a result, the peaks will ultimately be detected as path arrivals. The firstpath arrival will be estimated as the LOS path and thus used for distance estimation.It is clear in this case that the actual TOA is not equal to the estimated TOA. Thisdifference in estimation is the multipath error. For higher system bandwidths, the multipath error in LOS environments isusually smaller. For example, Fig. 2.8 illustrates a measured channel impulseresponse (measurement systems will be described in detail later in the chapter) for200 MHz bandwidth in a typical LOS office environment.
  15. 15. 2.2 The Multipath Problem 31Fig. 2.8 Measured powerdelay profile highlightingTOA estimation in indoorLOS office at 200 MHzbandwidth As can be seen from the figure, the actual/expected TOA is very close to thepeak of the measured TOA of the first path. Also, note that there are about 15multipath components for a noise threshold of –85 dBm. These non-direct mul-tipath components can originate from wall reflections, furniture diffractions, andscattering from other objects in the office. One way to assess the performance of TOA-based ranging is to analyze theranging error. In LOS environments, the ranging error could be attributed to bothmultipath and measurement noise. Let aDP and sDP denote the DP amplitude and 1 1propagation delay, respectively. The distance between the transmitter and thereceiver is dDP ¼ m  sDP , where m is the speed of signal propagation. Then 1ranging error which is defined as the difference between the estimated and theactual distance or, ^ e ¼ d À dDP ð2:45Þ In a general LOS multipath environment, the ranging device will experiencevarying error behavior depending on the structure of the propagation environmentand the system bandwidth. In LOS, the distance estimate can be modeled by ^ dDP ¼ dDP þ eDP ðBÞ þ w ð2:46Þ ~where eDP ¼ bm ðBÞ is a bias induced by the multipath and it is a function of thesystem bandwidth and w is a zero-mean additive measurement noise. As we willlater discuss, the statistics of the multipath bias can be modeled differently. Onepopular approach is to model it spatially as a zero-mean Gaussian (Alavi andPahlavan 2003). This means that an ensemble of LOS measurements in a givenLOS environment will generally result in a Gaussian distribution. The variance ofthe distribution will be directly related to the variation in the multipath structure in
  16. 16. 32 2 Ranging and Localization in Harsh Multipath Environmentsa given environment. For example, the spatial variance in an indoor office envi-ronment is typically higher than the variance in an outdoor, flat terrain. An analytical treatment of the performance of TOA estimation in multipathenvironments can be found in (Dardari et al. 2009) where Ziv-Zakai Bounds areintroduced for realistic propagation environments.2.2.2 RSS-Based Ranging in LOS Multipath EnvironmentsUnlike TOA-based ranging, RSS-based ranging depends on an a priori power–distance relationship or pathloss model. The power–distance relationship has beeninvestigated extensively in wireless communications for different technologies(Pahlavan and Levesque 2005; Rappaport 1996). In many of the experimentalfindings, the distance is related to the power law. For a narrowband transmittedsignal in free-space with transmitted power Pt , the received signal power Pr isgiven by Pahlavan and Levesque (2005) k 2 Pr ¼ Pt Gt Gr ð2:47Þ 4pdwhere Gt and Gr are the transmitter and receiver antenna gains, respectively. k isthe wavelength of the transmitted signal and d is the distance between the trans-mitter and receiver. A reference received power at distance d ¼ 1 m is usuallydefined as P0 ¼ Pt Gt Gr ðk=4pÞ2 then the distance–power relationship in free spacecan be given by P0 Pr ¼ : ð2:48Þ d2 RSS ranging is based on models which assume an a priori relationship betweenthe distance and the received power (or pathloss of the signal). A popular model inLOS channels relates the received power to the transmitted power by the followingequation log10 Pr ¼ log10 P0 À 10c log10 d ð2:49Þwhere c is the pathloss exponent that determines the rate of power loss withincreasing distance. Note that this is equivalent to (2.48) for c ¼ 2. If we definepathloss to be the ratio of received power to transmitted power then the abovepower–distance relationship can be rewritten in terms of pathloss L given by L ¼ L0 þ 10c log10 d ð2:50Þwhere L0 ¼ 10 log10 Pt À 10 log10 P0 and L ¼ 10 log10 Pt À 10 log10 Pr . In order tomodel the power–distance relationship more accurately, a random component thatmodels the shadow (slow) fading is included or
  17. 17. 2.2 The Multipath Problem 33 L ¼ L0 þ 10c log10 d þ S ð2:51Þwhere S is a normally distributed random variable in the log domain and it modelsthe fluctuation of the signal away from the median pathloss. This fluctuation stemsfrom the presence of different obstructions between the transmitter and receiverwhich ‘‘shadow’’ the signal. RSS-ranging is mainly affected, however, by fast-fading (Pahlavan and Levesque 2005; Rappaport 1996). At the receiver, theattenuated and phase shifted replicas of the transmitted signal combine eitherconstructively or destructively. The effect is a fast fluctuation of power at a givendistance. One way to deal with this fast fading problem is to collect more RSSmeasurements and ‘‘average out’’ the fluctuations by taking the mean of themeasurements. Then, for a given pathloss exponent and P0 , the Maximum Like-lihood Estimate (MLE) of the distance between a transmitter and a receiver can beestimated from the measured received power as Patwari et al. (2003) dMLE ¼ 10ðP0 ÀPr Þ=ð10cÞ ^ ð2:52Þ The major weakness with RSS-based distance estimation is the assumption thatthe pathloss exponent (pathloss model) is known a priori when in fact the exponentchanges between multipath environments—and even within the same environ-ment. Furthermore, the accuracy of the range estimate cannot be improved byaveraging the received signal power alone. Averaging of the RSS prior to esti-mating the distance will only remove the small-scale fading (fast fading) due to themultipath but not the shadow fading (which is more common in NLOS environ-ments). Typical values for the pathloss exponent in LOS multipath environmentsrange between 1 and 2 (Pahlavan and Levesque 2005). There are approaches thatattempt to estimate the pathloss exponent prior to the localization stage, but thatapproach presents some challenges as well (Li 2006). The statistical performanceof RSS ranging can be analyzed through the well-known CRLB given by Qi andKobayashi (2003) ðln 10Þ2 r2 d 2 r2 ! RSS S ð2:53Þ 100cwhere r2 is the variance of the shadow fading term. This relationship indicates that SRSS-based ranging estimation is affected by the pathloss exponent and the vari-ance of the shadow-fading in addition to the distance. As the distance increases,RSS estimation degrades. More importantly, as the variance of the shadow fadingincreases, the variance of RSS ranging also increases. This basic, yet powerfulrelationship highlights the challenges of RSS-based ranging. In typical multipathenvironments, the shadow fading variance is significant and thus reliable estima-tion of the distance can be difficult. In addition, the inverse dependency on thepathloss exponent indicates that performance of RSS ranging in LOS environ-ments (lower pathloss exponent *1–2) is expected to be much better than NLOSenvironments (typical pathloss exponents *3–5). These challenges to RSS-basedranging make it a more practical, but inaccurate option for localization.
  18. 18. 34 2 Ranging and Localization in Harsh Multipath Environments2.3 The NLOS ProblemThis section introduces the NLOS problem and describes the impact of NLOSchannels on TOA- and RSS-based ranging. For the former, NLOS affects theestimation of the direct path signal. Since in most cases the direct path will not bedetectible, ranging is achieved through non-direct path components which biasTOA-based estimation. For the latter, NLOS introduces the problem of shadowfading, where RSS is attenuated randomly as the mobile device moves from onearea to the other.2.3.1 TOA-Based Ranging in NLOS Multipath EnvironmentsIn the previous section, the basics of TOA-based ranging in LOS environmentswere introduced. A natural extension of the LOS case is a more challenging andcomplex situation where the transmitter and receiver experience an NLOS mul-tipath channel. Specifically, when considering NLOS cases, there is an obstructionin the path of the transmitter and receiver. Depending on the type of obstructionand the relative distances of the transmitter/receiver to the obstruction, the channelbehavior can vary significantly. There are two specific NLOS cases that occur intypical obstructed environments. The first is when the direct path (DP) signal isattenuated but detected (albeit weak SNR). This situation can arise naturally whenthe transmitter and receiver are separated by ‘‘light’’ obstructions such as a glass ora wooden door. Indeed, in this scenario TOA estimates can be obtained with goodaccuracy due to the detection of the DP signal. The second NLOS case is whenthere is a ‘‘heavy’’ or severe obstruction between the transmitter and receiver,where the direct path is severely attenuated and ‘‘buried’’ under the noise floor ofthe receiver, making it undetectable. The first non-Direct path (NDP) component isthen used for TOA estimation. This results in a significant bias that corrupts theTOA estimation and ultimately the position estimate. In this severe NLOS con-dition, the variance of TOA estimation with time is usually large due to the factthat the estimated first arrival path varies significantly due to the shadowingproblem. For a quasi-static channel, the first path can be detected. However, whensome perturbation is introduced to the multipath structure (another person movesaround/close to the TX-RX path), then the estimation of the first path arrival willfluctuate significantly. It is clear, then, that NLOS does not only introduce a bias,but also introduces significant TOA estimation perturbations that can degrade thereal-time distance estimation. Formally stated, in the absence of the DP, ranging is achieved using the amplitudeand propagation delay of the first Non-Direct Path (NDP) component—denoted asaNDP and sNDP respectively—resulting in a longer distance dNDP ¼ m  sNDP , where 1 1 1dNDP [ dDP . In order for the receiver to successfully identify the DP, the ratio of thestrongest multipath component to that of the DP, given by
  19. 19. 2.3 The NLOS Problem 35 Lp max jai ji¼1 j1 ¼ ; ð2:54Þ aDPmust be less than the receiver dynamic range, j, and the power of the DP must begreater than the receiver sensitivity, u. These constraints are given by j1 j ð2:55Þ PDP [ u ð2:56Þ À DP Áwhere PDP ¼ 20 log10 a1 . In an indoor environment the mobile device will experience varying errorbehavior depending on the availability of the DP and, in the case of its absence, onthe characteristics of the DP blockage. It is possible to categorize the error basedon the following ranging states (Alsindi et al. 2009). In the presence of the DP,both the constraints above are met and the distance estimate is accurate, yielding ^NLOS dDP ¼ dDP þ eNLOS þ w ð2:57Þ DP eNLOS ¼ bpd þ ~m ðBÞ DP b ð2:58Þwhere ~m is the zero-mean random bias induced by the multipath, bpd is the bias bcorresponding to the propagation delay caused by NLOS conditions and w is azero-mean additive measurement noise. It has been shown that ~m is indeed a bfunction of the bandwidth and signal to noise ratio (SNR) (Pahlavan et al. 1998),while bpd is dependent on the medium of the obstacles (Gentile and Kik 2007). Inthe more severe case, the DP is completely attenuated and the requirement thatj1 j is not met because the DP is shadowed by some obstacle, burying its powerunder the dynamic range of the receiver. In this situation, the ranging estimateexperiences a larger error compared to the LOS condition. Emphasizing thatranging is achieved through the first arriving NDP component, the estimate is thengiven by ^NLOS dNDP ¼ dDP þ eNLOS þ w ð2:59Þ NDP eNLOS ¼ ~m ðBÞ þ bpd þ bNDP NDP b ð2:60Þwhere bNDP is a deterministic additive bias representing the nature of the blockage.Unlike the multipath biases, and similar to biases induced by propagation delay,the dependence of bNDP on the system bandwidth and SNR has its own limitations,as reported in Pahlavan et al. (1998). Figure 2.9 illustrates the two specific con-ditions occurring in NLOS environments. An example of the measured channel profiles in the NLOS conditions is shownin Figs. 2.10 and 2.11. It is clear from the figures that in NLOS channel conditions large ranging errorsare possible, highlighting the major limitation to deploying accurate geolocation
  20. 20. 36 2 Ranging and Localization in Harsh Multipath Environments (a) (b)Fig. 2.9 Indoor NLOS multipath channels. a ‘‘Light’’ NLOS—the DP is attenuated but can bedetected b Severe NLOS—the DP is not detectedFig. 2.10 Measurement of a‘‘light’’ NLOS channel—theDP is attenuated but can bedetectedsystems in urban and indoor environments. The impact of NLOS range mea-surements on the localization performance can be evaluated through CRLB-typebounds. Given that the statistics of the NLOS biases are available then it is pos-sible to derive the Generalized-CRLB which integrates the statistical information.The analytical treatment of this problem can be found in Qi et al. (2006).2.3.2 RSS-Based Ranging in NLOS Multipath EnvironmentsIn the previous section, RSS-based ranging in LOS multipath environment wasintroduced and it was illustrated how a simple pathloss model can be used toestimate the distance. Besides the limitation due to the unknown parameters of the
  21. 21. 2.3 The NLOS Problem 37Fig. 2.11 Samplemeasurement of a SevereNLOS multipath channel—the DP is not detectedpathloss model, the challenge of RSS ranging in NLOS is exacerbated by the factthat obstructions between the transmitter and receiver can further complicate thedistance–power relationship, making it difficult to directly estimate the distanceaccurately. For example, consider a mobile station moving away from a basestation in a typical LOS environment. The pathloss model for this scenario is atypical LOS propagation model with pathloss exponent around 1–2 and minimalshadowing variance. However, as the mobile moves behind a wall, cabinet, or evenan elevator, the power suddenly fluctuates and severe attenuation perturbs the LOSdistance–power relationship. It then becomes very difficult to achieve accuratedistance estimation in light of this problem. Although Li (2006) proposed atechnique to estimate the pathloss exponent in real-time, the limitations still affectthe accuracy and practicality of this approach. As a result, numerous researchefforts have focused instead on an alternative RSS-based localization technique,namely fingerprinting-based localization, an approach to which Chap. 4 of thisbook is completely dedicated. In NLOS environments the pathloss model introduced earlier for LOS envi-ronments can be further extended L ¼ L0 þ 10cNLOS log10 d þ SNLOS ð2:61Þwhere cNLOS and SNLOS is the pathloss exponent and shadow fading parameters forNLOS. Usually cNLOS [ cLOS , with cNLOS ranging between 3 and 6 (Pahlavan andLevesque 2005). The NLOS pathloss model will be significantly different whenconsidering the type and number of obstructions separating the transmitter andreceiver. For example in indoor NLOS environments, the number of walls betweenthe transmitter and receiver can significantly change the pathloss behavior. An
  22. 22. 38 2 Ranging and Localization in Harsh Multipath Environmentsadditional parameter to incorporate the wall effect has been modeled in the lit-erature as X N L ¼ L0 þ 10cNLOS log10 d þ SNLOS þ Wn ð2:62Þ n¼1where Wn is the attenuation specific to a type of wall (Durantini and Cassioli 2005).It is clear that, in practice, it is very difficult to have an accurate pathloss model thatcan be used to estimate the distance accurately for all the environments.2.4 Empirical Evaluation of the Multipath and NLOS ProblemsIn order to understand the impact of the propagation channel on the effectiveness ofexisting TOA-based and RSS-based algorithms and to appreciate the limitations thatthey face, it is necessary to empirically characterize the radio propagation channel forthe ranging- or geolocation-specific application. The TOA- and RSS-specificpropagation studies help to shed light on the fundamental aspects of the rangingtechnique and the parameters that control its performance. In this section, we willprovide an overview of the measurement techniques, results, and modeling effortsthat have been carried out for TOA- and RSS-based ranging. The aim of this section isto introduce the reader to the methodologies used to measure and characterize thewireless channel for geolocation applications. This will serve as a foundation throughwhich it is possible to understand the limitations facing some of the popular rangingand localization techniques that will be introduced in the later chapters.2.4.1 Channel Measurement SystemsIn order to characterize the behavior of TOA- or RSS-based ranging in multipathenvironments, the channel impulse response (CIR) or the power delay profile ofthe channel must be measured. The CIR is the time-delay characterization of themultipath and it provides the amplitude/delay relationship of the arriving multipathcomponents. In practice the CIR can be measured directly by either using a time–domain measurement system or indirectly by using a frequency–domain mea-surement system. For geolocation-specific measurements and modeling, eithersystems can be used to extract relevant information for TOA-based ranging.Specifically, the measurement systems can be used to measure the large-scale,spatial characteristics of the direct path, mainly the ^DP and the ^DP , which can be a1 s1used to examine the ranging coverage (pathloss characterization) and accuracy,respectively. In the absence of the DP, it is possible to measure the first detected
  23. 23. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 39 PA LNA Signal Digital Sampling Channel Generator Oscilloscope Laptop Trigger SignalFig. 2.12 Time domain measurement system block diagrampath, ^NDP , and analyze the probability of blockage and the error statistics in this s1condition. These TOA-based parameters can be extracted directly from the mea-sured CIR.2.4.1.1 Time Domain SystemsOne way to capture the channel multipath profile is through the well-known timedomain measurement system. The channel is captured by transmitting a knownwaveform (with special autocorrelation properties) and post-processing thereceived waveform by cross-correlation with the known template. Sincethe arriving waveform will be a superposition of shifted and attenuated replicas ofthe original signal, then the output of the cross-correlation will contain ‘‘peaks’’ atthe delay values of the multipath components. A typical time domain measurementsystem is depicted in Fig. 2.12. Typically, the template waveform can be either pulses or PN-sequences,employed in direct-sequence spread spectrum systems (Ciccognani et al. 2005).After amplification, the received waveform is captured by a digital samplingoscilloscope and stored for post-processing (Cassioli et al. 2002). Depending onthe waveform type, the multipath profile can be extracted from the receivedwaveform. In the case of the PN-sequence waveform, the received signal is cor-related (after demodulation) with a replica of the transmitted sequence (Janssenand Vriens 1991). Note that for this measurement system the signal generator mustbe ‘‘synchronized’’ with the digitally sampling oscilloscope. That is a triggersignal is typically used to trigger the events for correlation purposes.2.4.1.2 Frequency Domain SystemsOne of the most popular and practical methods to measure the wireless channel isthrough the use of the frequency-domain measurement system. For such mea-surement systems a generic vector network analyzer (VNA) can be used. Fre-quency-domain measurement techniques have been previously employed tocharacterize the channel impulse response (Ghassemzadeh et al. 2004); Chong and
  24. 24. 40 2 Ranging and Localization in Harsh Multipath Environments RX LNA Vector Network TX Analyzer PAFig. 2.13 Frequency-domain measurement system block diagram Calibration IFT (Chirp-Z) Peak Detection H U ( f )3−6GHz H ( f )3−6GHz h (τ )Fig. 2.14 Measurement post-processing and CIR generationYong (2005), Pahlavan and Levesque (2005), Howard and Pahlavan (1990) but formodeling the communication channel—characterizing RMS delay spread andpower-distance relationships. The frequency measurement system captures thechannel transfer function (CTF) and the time domain CIR can then be obtained bythe inverse Fourier Transform (IFT). The core of the measurement system is the VNA, which is used to sweep thefrequency spectrum of a desirable system bandwidth with a certain samplinginterval. The CTF can be captured by measuring the S21 S-parameter on the VNAwhich are samples of the frequency domain of the channel. Figure 2.13 illustratesan example measurement system setup. Further details of the measurement systemcan be found in Ghassemzadeh et al. (2004), Pahlavan and Levesque (2005) andAlsindi et al. (2009).
  25. 25. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 41 H( f ) Δf f1 f2 B IFT h( ) Δt maxFig. 2.15 Frequency domain measurement system—parameter relationships between thefrequency and time domain signalsTable 2.1 Common B ¼ f2 À f1 ¼ Nf Dffrequency/time domain smax ¼ tNt À t1 ¼ Nt Dtdefinitions and relationships 1 1 B / Dt and smax / Df The CIR is then obtained by an IFT process and Fig. 2.14 highlights the systemblock diagram of the post-processing stage. The uncalibrated measured CTF from the VNA is passed through a post-measurement calibration process that removes the channel response of the cables,LNA and PA. The CIR is then obtained by the IFT or the Chirp-Z transform whichhas a signal processing ‘‘zooming’’ capability. The time and amplitude of themultipath delays are then extracted by passing the raw estimated CIR through apeak detection algorithm, that essentially identifies the peaks in the profile that aregreater than a certain noise threshold (typically -120 to -110 dBm). The frequency domain measurement parameters are related to the time domainchannel impulse response. The parameters that can be controlled in the VNA whenmeasuring the frequency response are the swept frequencies (bandwidth), thenumber of samples, and the transmitted power. The frequency spacing is deter-mined by the number of samples in a given bandwidth. For a CTF measurement
  26. 26. 42 2 Ranging and Localization in Harsh Multipath EnvironmentsHðf Þ ¼ Hðf1 ; f2 Þ the VNA can be configured to measure a certain bandwidthbetween f1 and f2 , or B ¼ f2 À f1 . Selecting the number of points will dictate thefrequency spacing Df . The relationship between the number of measured fre-quency samples, Nf , and the frequency spacing, Df is Nf ¼ ðf2 À f1 Þ=Df . Thefrequency samples on the VNA directly affect the time domain CIR. The measuredbandwidth B controls the time domain resolution Dt and the frequency spacing Dfcontrols the maximum time delay, smax , that can be measured. Figure 2.15 andTable 2.1 illustrate and summarize the relationship. The collected measurement data can be then used to extract the TOA or RSSparameters for analysis. In the next subsectionss we introduce some of the modelsdeveloped for the indoor environment.2.4.2 Alavi ModelsOne of the earliest TOA-based ranging measurements and modeling was con-ducted by Alavi and Pahlavan (2006). The focus of the measurement and modelingwas to characterize the impact of multipath on the accuracy of range estimation.The measurements and modeling provided an analysis of the impact of systembandwidth on the multipath-induced error. In addition, the TOA-specific mea-surements errors were analyzed under different NLOS conditions. Specifically, inthis work, ranging error was referred to as Distance Measurement Error (DME)and it is given by ^ eB ðdÞ ¼ dB À d ð2:63Þ ^where d is the ground-truth distance, dB is the measured distance, and its depen-dence on system bandwidth is explicitly given by the subscript B. As a result, theerror is a function of the distance between the transmitter and receiver andthe bandwidth. Furthermore, depending on the condition of the indoor channel, theerror can be significantly different: in LOS environments, multipath is the domi-nant source of error while in NLOS the absence of the DP—also known asUndetected Direct Path (UDP)—dominates the error. UDP is essentially severeNLOS where the DP cannot be detected due to a large obstruction between thetransmitter and receiver which causes the DP path to be buried under the receivernoise floor. The models were obtained by conducting frequency domain mea-surements using the VNA described in the previous subsections. Figures 2.8 and2.11 illustrate LOS versus NLOS with undetected DP. By comparing the two measured profiles, it is clear that the error in UDPconditions contains a combination of the multipath error and a ‘‘UDP’’ error,which is essentially a bias in the time delay estimation. Note from the figure thatthe direct path is severely attenuated and lies below the noise threshold, whichmakes its detection very difficult. Based on the measurements in an indoor
  27. 27. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 43environment, Alavi introduced a model that incorporates the different rangingconditions. Specifically, the error is modeled as eB ðdÞ ¼ e þ eM;B ðdÞ þ nðBÞeU;B ðdÞ ð2:64Þwhere eM;B ðdÞ is the multipath error, eU;B ðdÞ is the UDP error or bias, and nB ðBÞ isa random variable that takes the value of ‘‘1’’ when a UDP condition occurs and‘‘0’’ otherwise. The model also includes e, which is an error that models theinaccuracies occurring during measurement of the actual distance between thetransmitter and receiver. Typically, this error can be assumed zero-mean Gaussianwith a variance that depends on the accuracy of the measurement error. Since ecannot be separated from the multipath error, it is assumed thate þ eM;B ðdÞ % eM;B ðdÞ, which simplifies the model to eB ðdÞ ¼ eM;B ðdÞ þ nB ðBÞeU;B ðdÞ: ð2:65Þ The multipath error eM;B ðdÞ can been modeled by eM; B ðdÞ ¼ XðmM; B ; rM; B Þ logð1 þ dÞ ð2:66Þwhere XðmM; B ; rM; B Þ is a Gaussian random variable with mean mM; B and standarddeviation rM; B . The UDP error component was similarly modeled as GaussianXðmU;B ; rU;B Þ. As a result, the overall model is given by ^ d ¼ d þ MDME þ nB ðdÞUDME ð2:67Þ ¼ d þ XðmM;B ; rM;B Þ logð1 þ dÞ þ nB ðdÞXðmU;B ; rU;B Þ The random variable nB ðdÞ can be modeled as À Á fnW ðyÞ ¼ 1 À pU;B ðdÞ dðyÞ þ pU;B ðdÞdðy À 1Þ: ð2:68Þ The proposed models have been verified to fit actual data in Alavi and Pahlavan(2006). The work in Alavi and Pahlavan (2006) also investigated the impact of thesystem bandwidth on the DME. Basically, as the system bandwidth increases,the error decreases due to enhanced time resolution. The finding further supportsthe idea that one way to mitigate the multipath problem is to increase the systembandwidth. This observation was also highlighted in Gentile and Kik (2007).2.4.3 Alsindi ModelsAs stated earlier, the Alavi models were the first models developed for TOA-basedranging that analyzed the impact of LOS/NLOS and system bandwidth on theaccuracy. The results of these models highlighted the fundamental limitations andchallenges facing TOA-based ranging in harsh multipath environments. The
  28. 28. 44 2 Ranging and Localization in Harsh Multipath EnvironmentsFig. 2.16 Firefighter/soldier localization scenario in hostile indoor environmentsmeasurements and models, however, were limited in scope since they were basedon a single floor/office of an indoor environment and thus lacked comprehensiveanalysis in different buildings/environments. In addition, the models do not pro-vide any indication of the coverage aspect of the ranging systems. As a continu-ation of the modeling efforts, Alsindi’s work focused on developing models forUltra-Wideband (UWB) TOA-based systems that characterize in detail some ofthe fundamental parameters such as ranging coverage, ranging error in LOS,NLOS-presence of DP, and NLOS-absence of DP (Alsindi et al. 2009). UWB isdefined as any system operating with a bandwidth of 500 MHz or with a band-width exceeding 20 % of the center frequency. The objective of the measurement campaign was to develop models for fire-fighter/soldier TOA-based ranging/localization in hostile indoor environments. Insuch scenarios, beacons or anchors were placed surrounding a given building inorder to aid firefighters/soldiers to localize and navigate themselves in an indoorenvironment through cooperative localization using wireless sensor networks(WSN). Cooperative localization is dealt with in Chaps. 6 and 7 where centralizedand distributed techniques will be discussed in more detail. Figure 2.16 illustratesthe localization scenario that was considered for the measurement campaign. In order to develop reliable systems operating in these challenging environ-ments, it is necessary to understand the propagation characteristics that impactranging and localization accuracy. It is clear from the figure that three distinctranging scenarios are possible: Indoor-to-Indoor (ITI), Outdoor-to-Indoor (OTI)and Roof-to-Indoor (RTI). In addition four different building types were investi-gated: old office (Atwater Kent—AK), new office (Fuller Labs), residential(Schussler) and manufacturing floor (Norton). All the buildings are in Worcester,MA, USA. From this application point of view, it is then interesting to investigatethe following issues:
  29. 29. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 45Fig. 2.17 NLOS challengesfacing the firefighterlocalization application• For the outdoor beacons (OTI RTI), how far can the devices reliably provide TOA-based ranging estimates? What is the ranging coverage?• What is the probability of DP blockage in NLOS environments?• What are the ranging error characteristics in ITI, OTI and RTI?• How is the ranging/localization performance impacted for different building types: residential, office, etc.? For the firefighter/soldier localization scenario, the multipath and NLOSproblems can be difficult challenges that will impact the accuracy of the locali-zation directly. Figure 2.17 highlights the NLOS challenges facing OTI/RTI andITI scenarios. For OTI/RTI scenarios, the signal propagating through the external walls of thebuilding typically undergoes significant attenuation because the walls are usuallythick in construction and are composed of brick and steel material. As a result, theranging coverage can be limited significantly and, in most cases, is much less thanthe ITI scenarios. For ITI scenarios the ranging coverage, although higher thanOTI/RTI, is significantly different for LOS and NLOS scenarios. Alsindi’s models focused on characterizing the ranging coverage and rangingerror in these different scenarios and environments. For the former the distance–power relationship of the Direct Path (DP) signal provides an empirical evaluationof the ranging coverage which is the maximum distance where the DP can bedetected. For the latter the spatial distribution of the ranging error in differentscenarios and environments provides an empirical evaluation of the physicallimitation facing indoor geolocation. In indoor environments, the distance-dependence of the received power, whichcan be used to determine the communication coverage, is usually predicted from
  30. 30. 46 2 Ranging and Localization in Harsh Multipath EnvironmentsTable 2.2 Summary of TOA-based ranging error conditionsLOS NLOS-DP NLOS-NDP/UDP^dDP ¼ dDP þ eDP ðBÞ þ w ^NLOS dDP ¼ dDP þ eNLOS þ w ^NLOS dNDP ¼ dDP þ eNLOS þ w DP NDP eDP ðxÞ ¼ bm ðBÞ eNLOS ¼ bpd þ bm ðBÞ DP eNLOS ¼ bm ðBÞ þ bpd þ bNDP NDPexperimental pathloss models of the total signal energy in different environmentsand scenarios (Durgin et al. 1998; Molisch 2005; Ghassemzadeh et al. 2004).Similarly, the distance-dependence behavior of the power of the DP can be used todetermine the ranging coverage. Unlike communication coverage which is relatedto the received power of all the multipath components at a given distance, rangingcoverage is related to the received power of the DP component. For a given systemdynamic range, j, ranging coverage, Rr , is defined as the distance in which themaximum tolerable average pathloss of the DP is within j (Alsindi et al. 2009).This is represented by n o max LDP ¼ 10c log10 ðRr Þ j p ð2:69Þwhere LDP is the average pathloss of the DP and c is the pathloss exponent. The ppathloss behavior of the DP is distance-dependant, but because of the attenuationand energy removed by scattering, its intensity decreases more rapidly with dis-tance compared with the total signal energy (Siwiak et al. 2003). This means thatfor typical indoor multipath scattering environment, communication coverage isgreater than ranging coverage, Rc [ Rr . Operating out of ranging coverage causeslarge TOA estimation errors and performance degradation (always in NLOS-NDPcondition). The characterization of ranging error in different scenarios has beenintroduced earlier in the chapter and it is summarized in Table 2.2 forconvenience.2.4.3.1 Modeling the Pathloss: Ranging CoverageUsing the same established pathloss modeling approach used in the literature,(Ghassemzadeh et al. 2004; Pahlavan and Levesque 2005), Alsindi characterizedthe distance–power dependence of the measured DP (Alsindi et al. 2009) andcompared it to the distance–power relationship of the total received power (RSS).The pathloss exponent is determined from measurement data through least-square(LS) linear regression. The pathloss relationship is provided in (2.61) but anadditional factor attributed to the power loss due to penetration through walls canbe incorporated as LX (which is depending on the ranging scenario OTI, RTI, etc.).Thus the modified expression is given by LðdÞ ¼ L0 þ LX þ 10c log10 ðd=d0 Þ þ S; d ! d0 : ð2:70Þ
  31. 31. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 47Fig. 2.18 Pathloss scatter plots in Fuller ITI LOS at 3 GHz bandwidth All the parameters of the model in (2.70) are a function of the building type/propagation environment. Figures 2.18, 2.19 and 2.20 show sample measuredscatter plots of the pathloss as a function of TX-RX separation for differentbuildings and ranging scenarios. The pathloss model parameters are summarized in Table 2.3. Several observations can be made from the table and the figures. The first is thatfor all the measurement data the pathloss exponent is higher for the DP relative tothe total signal power, which is consistent with the modeling approach. Second,the DP power experiences greater fluctuations around the mean pathloss ascompared with the total signal counterpart. This observation makes sense becausesmall variations on the transmitter location affect the DP power more than the totalpower. Third, LX changes for the different penetration scenarios. In ITI scenariosSchussler NLOS suffers 6 dB penetration loss due to the walls compared to 7.5 inAK. Norton ITI measurements are a mixture of LOS/NLOS because the manu-facturing floor contained scattered machines. The impact can be clearly seen on thepathloss exponent when the bandwidth increases, hence higher attenuation. Resultsof OTI measurements show that Fuller and AK exhibit the largest penetration lossmainly because the signal had to penetrate a thicker building construction whencompared with Norton and Schussler. In addition, the pathloss exponents in AKare large mainly because the measurement locations were conducted inside a metalshop on the edge of the building and between concrete corridors and rooms. AK ingeneral imposes a very challenging environment for ranging because of the
  32. 32. 48 2 Ranging and Localization in Harsh Multipath EnvironmentsFig. 2.19 Pathloss scatter plots in Norton OTI at 500 MHz bandwidthFig. 2.20 Pathloss scatter plots in AK RTI at 500 MHz
  33. 33. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 49Table 2.3 Pathloss modeling parametersScenario Environment LX(dB) Direct Path Total signal 500 MHz 3 GHz c S (dB) c S (dB) c S (dB)ITI Fuller (LOS) 0 3.2 8.9 3.3 7.1 2.4 5.5 Norton (Mixed) 0 3.5 8.5 4.5 9.1 2.6 3.4 Schussler (NLOS) 6 3.4 7.9 4.0 8.4 3.0 4.6 AK (NLOS) 7.5 5.4 6.2 5.6 8.5 3.6 6.2OTI Fuller 14.3 3.4 13.7 3.7 14.1 2.2 7.7 Norton 8.7 3.9 7.8 5.0 10.1 3.3 4.4 Schussler 7.6 4.1 10.5 4.2 11.1 3.2 6.1 AK 10 4.6 8.7 5.1 8.9 3.1 3.2RTI AK 24.5 4.3 7.6 5.3 8.8 2.9 1.7building material and dense cluttering. RTI measurements experienced the largestpenetration loss and high pathloss exponent. Finally, note that the harsher theindoor environment, the higher the pathloss exponent difference when moving to ahigher system bandwidth. This is mainly due to the fact that larger systembandwidths provide better time domain resolution at the cost of reduced power permultipath component. This implies that the advantage of higher time domainresolution comes at a cost of shorter ranging coverage.2.4.3.2 Modeling the Ranging ErrorThe spatial characteristics of the ranging errors are determined through thebehavior of the biases, which are random due to the unknown structure of theindoor environment and the relative location of the user to them. Since the errorsare highly dependent on the absence or the presence of the DP, the modelsintroduced by Alsindi are based on the classification in Table 2.2. Further, in orderto model and compare the behavior in different building environments and sce-narios, the normalized ranging error was modeled instead as À Á e ^ dÀd w¼ ¼ : ð2:71Þ d d The range error observed in an indoor environment can then be modeled bycombining the conditions in Table 2.2 through the following expression w ¼ wm þ G wpd þ XwNDP ð2:72Þwhere wm is the normalized multipath error that exists in both the presence andabsence of the DP. wpd is the normalized propagation delay-induced error, andwNDP is the normalized error due to DP blockage. In order to distinguish between
  34. 34. 50 2 Ranging and Localization in Harsh Multipath EnvironmentsTable 2.4 Probabilities of the presence and absence of the DPScenario Environment 500 MHz 3 GHz pðf1 Þ pðf2 Þ pðf1 Þ pðf2 ÞITI Fuller 0.1 0.90 0.2 0.98 Norton 0.96 0.4 0.83 0.17 Schussler 0.89 0.11 0.87 0.13 AK 0.39 0.61 0.32 0.68OTI Fuller 0.42 0.58 0.39 0.61 Norton 0.57 0.43 0.24 0.76 Schussler 0.77 0.23 0.60 0.40 AK 0.40 0.60 0.22 0.78RTI AK 0.58 0.42 0.37 0.63the error behavior in LOS and NLOS, a Bernoulli random variable, G was used.That is, 0, LOS G¼ ð2:73Þ 1, NLOSwhere pðG ¼ 0Þ ¼ pðLOSÞ is the probability of being in LOS and pðG ¼ 1Þ ¼pðNLOSÞ is the probability of being in NLOS. Similarly, X is a Bernoulli randomvariable that models the occurrence of DP blockage and is given by 0; f1 X¼ ð2:74Þ 1; f2where pðX ¼ 0Þ ¼ pðf1 Þ denote the probability of detecting a DP, whilepðX ¼ 1Þ ¼ pðf2 Þ denotes the probability of the occurrence of blockage. It isimportant to emphasize that Alsindi’s modeling approach focuses on the DP andnot the traditional definition of NLOS used for communications. This means that amobile station and a base station separated by a wall, for instance, is consideredNLOS, but does not necessarily imply the absence of the DP. In the remainder ofthe chapter, ranging error, bias, and normalized error will be used interchangeably. The results of the measurement and modeling also revealed a significant dif-ference in the probability of DP blockage among the different environments, whichis highlighted in Table 2.4. Several observations can be concluded. First, a positive correlation between thesystem bandwidth and the blockage probability pðf2 Þ exists due to lower energyper MPCs in higher system bandwidths. Second, as expected, DP blockageincreases from ITI, to OTI, and RTI. Attenuation due to penetration from exteriorwalls and ceiling results in higher pðf2 Þ. Third, blockage is highly correlated withthe building type. In residential environments, blockage probability is low sincethe interior is composed of wooden structures with few metallic objects (e.g. afridge, laundry room, etc.). Office buildings, however, pose harsher conditionswith thicker walls, metallic beams, vending machines, metallic cabinets, shelves,
  35. 35. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 51Fig. 2.21 Norton ITI at 500 MHz bandwidth: confirming the normality of the biases in LOSconditionsand elevator shafts, resulting in a substantial blockage up to 90 %, see Fuller andAK (ITI/OTI). Also, ITI measurements on the manufacturing floor highlight theimpact of occasional clutter of machineries. Finally, it is worth mentioning thatthese results were measured using a 120 dB dynamic range provided by theexternal amplifiers and LNA extending the measured range. In realistic UWBsystems, unfortunately, this would be prohibitively high in terms of implementa-tion expense, which means that the results here can be seen as a lower bound. The models also analyze the behavior of ranging error in the presence and in theabsence of the DP. The measurement results of the ranging error in LOS scenariosrevealed that the impact of the multipath can be modeled through a normal dis-tribution since the DP is available and the error deviates in both directions relativeto the actual distance. In addition, normality of the ranging error in this conditionhas been reported in Alavi and Pahlavan (2003, 2006). The error distribution canthen be explicitly modeled as, # 1 ðw À lLOS Þ2 f ðwjG ¼ 0Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp À ð2:75Þ 2pr2 LOS 2r2 LOSwith mean lLOS and standard deviation rLOS specific to the LOS multipath-induced errors. Figure 2.21 further confirms the normality of errors in thiscondition. A similar observation of the multipath effect in indoor LOS environments hasbeen reported through measurements (Alavi and Pahlavan 2006). In NLOS
  36. 36. 52 2 Ranging and Localization in Harsh Multipath EnvironmentsFig. 2.22 Schussler ITI NLOS—mean of biases is larger than LOSscenarios, when the DP is present, the amount of propagation delay and multipathdue to obstructing objects such as wooden walls causes the biases to be morepositive. The results show (see Fig. 2.22) that the spatial characteristics retain thestatistics of the LOS counterpart but with a higher mean and standard deviation. According to these results, the normalized ranging error is modeled similar to(2.75), but with emphasis on the condition. This is given by, # 1 ðw À lNLOSÀDP Þ2 f ðwjG ¼ 0; X ¼ 0Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp À ð2:76Þ 2pr2 NLOSÀDP 2r2 NLOSÀDP The subscripts in (2.76) specify the contributing error factors. Table 2.5 pro-vides the modeling parameters of all the scenarios and environments in thepresence of the DP. The results show a positive correlation between the statistics of the normaldistribution with the complexity of environment and/or ranging scenario. Negativecorrelation can be seen between the statistics and the system bandwidth due toreduction of multipath error in higher bandwidths. The ranging error behavior in the absence of the DP is significantly different.The shadowing of the DP impacts the error behavior in several ways. First, onlypositive errors occur, since the blockage induces a higher positive bias thatdominates compared to the multipath counterpart. Second, there are occasionallylarge positive range errors that occur due to heavier indoor constructions such aselevator shafts, clustering of cabinets, or even metallic doors. Third, the diversityof blocking material in indoor environments means that the spatial distribution of
  37. 37. 2.4 Empirical Evaluation of the Multipath and NLOS Problems 53Table 2.5 DP normal distribution modeling parameters for normalized ranging errorScenario Environment 500 MHz 3 GHz lLOS rLOS lLOS rLOSITI Fuller (LOS) 0 0.028 0 0.006 Norton (LOS) 0 0.022 0 0.007 lNLOSÀDP rNLOSÀDP lNLOSÀDP rNLOSÀDP Fuller (NLOS) 0.058 0.028 0.003 0.01 Schussler 0.029 0.047 0.014 0.016 AK (NLOS) 0.023 0.020 0.009 0.004OTI Fuller 0.015 0.017 0.002 0.011 Norton 0.019 0.029 0.002 0.015 Schussler 0.041 0.045 0.011 0.013 AK 0.034 0.023 0.012 0.004RTI AK 0.029 0.041 0.012 0.012Fig. 2.23 Schussler OTI at 3 GHz bandwidth—confirming the lognormality of the measurednormalized ranging errorerrors will in general exhibit a heavier positive tail. By examining the PDF of theerrors in this condition, it is observed that different subsets of the data showedvarying tail behavior. The ‘‘heaviness’’ of the tail depended on the ranging envi-ronment and scenario. Thus harsher blockage conditions, i.e., higher number ofblocked MPCs, exhibited heavier tails. As a result, the ranging error in this con-dition was modeled as log-normally distributed. The lognormal model is thengiven by,

×