Position Location for Futuristic Cellular Communications -- 5G and Beyond
aa r X i v : . [ c s . I T ] F e b O. Kanhere and T. S. Rappaport, “Position Location for Futuristic Cellular Communications - 5G andBeyond,” in IEEE Communications Magazine , vol. 59, no. 1, pp. 70-75, January 2021.
Position Location for Futuristic CellularCommunications - 5G and Beyond
Ojas Kanhere and Theodore S. Rappaport
NYU WIRELESSNYU Tandon School of EngineeringBrooklyn, NY 11201 { ojask, tsr } @nyu.edu Abstract —With vast mmWave spectrum and narrowbeam antenna technology, precise position location isnow possible in 5G and future mobile communicationsystems. In this article, we describe how centimeter-level localization accuracy can be achieved, particularlythrough the use of map-based techniques. We show howdata fusion of parallel information streams, machinelearning, and cooperative localization techniques furtherimprove positioning accuracy.
I. I
NTRODUCTION
Precise position location (also called position-ing or localization ) is a key application for thefifth generation (5G) of mobile communicationsand beyond, wherein the position of objects isdetermined to within centimeters. With the rapidadoption of Internet of Things (IoT) devices, avariety of new applications that require centimeter-level precise positioning shall emerge, such asautomated factories that require precise knowledgeof machinery and product locations to within cen-timeters. Geofencing is the creation of a virtualgeographic boundary surrounding a region of in-terest to monitor people, objects, or vehicles, andby using sensors on a moving object, the locationof the object may be continually and adaptively“geofenced” to trigger a software notification im-mediately when the object enters or leaves thevirtual geographic boundary. Position location towithin 1-2 m will enable accurate geofencing, suchthat users entering/leaving a room or equipment andpeople may be tracked in hospitals, factories, withinand outside buildings.Today’s fourth generation (4G) cellular networksrely on LTE signaling and the global positioningsystem (GPS) (which is accurate to within 5 m).However, in indoor obstructed environments, or inunderground parking areas and urban canyons, GPSsignals are attenuated and reflected such that userequipment (UE) cannot be accurately localized.To further refine the positioning capabilities ofGPS indoors and in urban canyons, SnapTrack“ wireless assisted GPS ” (WAG) improved the sen-sitivity of GPS receivers. Additionally, databases of geo-tagged Wi-Fi hotspots have been used bycompanies such as Apple and Google. The UEmay be localized using the known positions of allWi-Fi hotspots that the UE can hear, where theUE position estimate is formed from the weightedaverage of the received signal strengths, providingan accuracy of tens of meters. Although FCCrequirements specify a horizontal localization errorof less than 50 m for 80 percent of enhanced911 (E911) callers, a localization error less than3 m will be required for positioning applicationsof the future. Additionally, FCC requires a verticallocalization error less than 3 m for 80 percent ofE911 callers by April 2021, to identify the caller’sfloor level, which is achievable using barometricpressure sensors present in modern cell phones (seeFCC’s Fifth Report and Order PS Docket 07-114.).In addition to infrastructure-based positioningsystems, other sensor-based technologies such asvision-based localization using cameras (commonlyutilized by drones [1]) can provide accurate posi-tioning capabilities when fused with inertial sen-sors. However, in low-visibility environments, lo-calization systems at cellular frequencies work bet-ter since they are not blocked when visibility ishampered. Ultrasound indoor positioning systemssuch as Forkbeard are able to achieve a preci-sion level of 10 cm within an office environment.Autonomous vehicles utilize light detection andranging (LIDAR) to estimate the relative distancesto other vehicles with sub-millimeter accuracy [2],while factory-based systems using infrared haveshown good accuracy [3].Position location solutions are being developedusing other media such as ultra wideband (UWB),RFID, visible light, and Bluetooth. UWB signals, inthe 3.1-10.6 GHz band, have a bandwidth of morethan 500 MHz. Rapid strides in utilizing UWBfor localization are expected, with the iPhone 11currently carrying UWB chips that are typicallycapable of achieving a ranging accuracy on theorder of centimeters [4].The advent of millimeter-wave (mmWave) com-unications enables a paradigm shift in localizationcapabilities by allowing joint communication andposition location, utilizing the same infrastructure.As shown in this article, the massive bandwidths,coupled with the high gain directional, steerablemultiple-input multiple-output (MIMO) antennasat mmWave frequencies, enable unprecedented lo-calization accuracy in smartphones of the future.We demonstrate how the utilization of cooperativelocalization, machine learning, user tracking, andmultipath enables precise centimeter-level positionlocation.II. F UNDAMENTAL L OCALIZATION T ECHNIQUES
Today’s localization solutions primarily focus ongeometric localization with augmented assistance,wherein the position of the base station (BS) isknown and the UE location is determined based ongeometric constraints such as the BS-UE distancesand physical angular orientations between BS andUE.In angle of arrival (AoA) localization technique,the UE estimates the angle of the strongest re-ceived signal. AoA positioning was conceived forE911 in the early days of cellular [5]. In timeof arrival (ToA) (or time difference of arrival,TDoA) localization techniques, the UE estimatesthe distance (or difference in distance) from theBS by estimating travel time (or differences intravel time) of the reference signal from the BS.The UE may then be localized to the point wherethe circles (or hyperbolas) corresponding to theBS-UE distances intersect. A spatial resolution ofup to 2.44 m and 4.88 m is achievable with 5GNew Radio (NR) waveforms for ToA and TDoAmeasurements, respectively [6]. In addition to uti-lizing GPS for UE localization, 4G (and future 5G)networks implement TDoA localization and utilizethe barometric pressure sensors in UE for altitudeestimation [1]. The operation of AoA, ToA, andTDoA localization techniques is illustrated in Fig.1 and is well understood.
A. Accurate Localization in 5G Networks with Di-rectional Antenna Arrays and Wide Bandwidths
In the 5G era, it is now possible to achieve veryaccurate localization performance with highly di-rectional antenna arrays having narrow beamwidthsand wide bandwidths [7]. The frequency range (FR)2 of 5G NR covers mmWave frequencies rangingfrom 24.25 GHz to 52.6 GHz. Additionally, theIEEE 802.11 ad standard supports the use of the60 GHz mmWave band indoors, from 57 GHz to71 GHz. (cid:1) ,3 (cid:1) (cid:0) ,2 Δ(cid:3) (cid:3) − (cid:3) Δ(cid:3) = (cid:3) − (cid:3) (x ,y )BS 2 (x,y)UE (x ,y )BS 1(x ,y )BS 3d d d hyperbolacircle Fig. 1. The UE may be localized based on ToA (black circles),TDoA (red hyperbola), or AoA (black dotted lines) localizationtechniques [5].
The short wavelength in the mmWave frequencyband allows electrically large (but physically small)antenna arrays to be deployed at both the UE andBS. MmWave BS antenna arrays with 256 antennaelements and 32-element mobile antenna arraysare already commercially available. The frequency-independent half-power beamwidth (HPBW) of auniform rectangular array (URA) antenna withhalf-wavelength element spacing is approximately(102/N)°, where N is the number of antenna ele-ments in each linear dimension of the planar array[8], as seen in Fig. 2.Narrower HPBWs of antenna arrays allow theAoA of received signals to be estimated precisely,and further signal processing provides better accu-racy. For example, the sum-and-difference for aninfrared system technique achieved sub-degree an-gular resolution with two overlapping and slightlyoffset antenna arrays [3], showing it is possible tovery accurately detect precise AoA at UEs or BSs.Although mmWave frequencies suffer fromhigher path loss in the first meter of propagationand experience greater blockage losses comparedto lower frequencies, the greater gain providedby the directional antennas coupled with smallerserving cells (100-200 m radius) compensates forthe additional path loss. Indeed, recent research [9]demonstrates the feasibility of using mmWave foroutdoor localization.Utilization of mmWave frequency bands willenable unprecedented positioning accuracy due tothe ultra-wide bandwidths available, since the largerbandwidths allow finer time resolution of multipathsignals transmitted from the BS to the UE, on theorder of a nanosecond, where a 1 ns time resolutionimplies a spatial resolution of 30 cm before addi- ig. 2. The normalized antenna gain (with respect to boresight -the axis of maximum gain) of URAs with 8 ×
8, 16 ×
16, 32 ×
32, and64 ×
64 array elements. Note the half power beamwidths (HPBWs)are 12.76°, 6.34°, 3.17°, and 1.55° respectively. tional processing that can further improve accuracy.
B. Performance of Fundamental Localization Tech-niques in Dense Multipath Environments
ToA, TDoA, and AoA localization techniqueswere designed for line-of-sight (LoS) propagation.In indoor/outdoor non-line-of-sight (NLoS) envi-ronments however, multipath arrives at different an-gles with larger delays, yielding positioning error.Without using any advanced correction techniques,a poor mean error of 10 m was observed with well-known AoA localization based on NLoS indooroffice measurements [10]. Similar enormous meanerrors of 8-10 m inside buildings were observedin NLoS when the localization performance wastested using traditional methods from outdoor E-911 [5] via simulations in NYURay, a 3D mmWaveray tracer [7]. The poor localization accuracy ofknown approaches, in the face of multipath andan obstructed or weak LoS signal, motivates theneed to develop more accurate and robust local-ization approaches that exploit the wide bandwidthand narrow beamwidths of 5G and beyond formultipath-rich NLoS environments.
C. NLoS Mitigation for Accurate Positioning
To combat the poor performance of traditionalToA-, TDoA-, and AOA-based localization tech-niques in NLoS environments, NLoS mitigationtechniques can identify and then discard NLoSsignals to only use the LoS BSs for localization.This subsection describes a variety of techniques toselectively identify and discard the NLoS signals.In [11] the authors observed that with conven-tional WiFi radios operating at 2.4 GHz, the AoAwas stable over small UE movements (5 cm) in LoSenvironments, while in NLoS environments, theAoA varied by more than 5° if the UE was movedby 5 cm. If the AoA of the received power varied by more than 5° when the UE was moved by 5 cm, thesignal was assumed to correspond to an NLoS pathand thus discarded from use in estimating position.By suppressing NLoS multipath and only using theLoS path, a median localization accuracy of 23 cmwas achieved with six 2.4 GHz WiFi access points[11].Estimating the BS-UE distance, a critical stepfor ToA localization, may additionally be utilized todetermine whether the BS-UE link is in NLoS. Therunning variance of the BS-UE distance estimates( σ ) in NLoS is greater than LoS; hence, NLoS BS-UE links may be identified based on the runningvariance observed in real time. The UE can accu-rately be assumed to be in NLoS (and the UE-BSlink is not used for localization) when σ is greaterthan a calibrated threshold γ [12]. The varianceof distance estimates is greater for a mobile userthan for a stationary user due to the change in thetrue BS-UE distance when the UE is in motion. Toaccount for user motion, γ must be increased, andin [12], a constant proportional to the square of thevelocity of the user was added to γ to account foruser motion.Channel features such as maximum receivedpower, root mean square (RMS) delay spread,Rician-K factor, and the angular spread of depar-ture/arrival may be utilized to determine whetherthe UE is in NLoS [13]. NLoS channels typicallyhave lower maximum received power over thepower delay profile (PDP) due to the presence ofobstructions and reflectors. The delay spread ofmultipath components is higher in NLoS environ-ments. The K-factor of a channel is equal to theratio of the square of the peak amplitude of thedominant signal and the variance in the channelamplitude and is known to indicate the degree ofmultipath in a signal [13]. In NLoS channels, dueto the absence of a direct path, the K-factor is closeto 0 dB. The angular spread of NLoS channels iswider since the multipath components arrive fromvaried directions.NLoS classification accuracy is improved whenmultiple channel characteristics are used in tan-dem [13]. A support vector machine (SVM) is apopular classifier capable of classifying data basedon multiple parameters. An SVM utilizes channelcharacteristics to determine a hyperplane, whichdivides data into two classes. For NLoS identifica-tion, the SVM determines the optimal hyperplane todivide data into LoS and NLoS classes. In [13], anSVM was shown to outperform individual channelfeatures, reducing the NLoS identification error ratefrom 10 percent to 5 percent.II. S UB - METER P RECISE P OSITION L OCATION
Identifying and discarding NLoS signals to onlyuse LoS signals for localization wastes multipathsignal energy, and requires dense BS deploymentsince the UE must be in LoS of two or more BSsfor classical LoS positioning techniques to work.However, such over-deployment of BSs may becost-prohibitive. We shall now look at alternativelocalization techniques wherein the UE utilizesinformation from neighboring UEs, and exploitsNLoS BSs, and multipath.
A. Cooperative Localization
With the introduction of device-to-device (D2D)communication protocols in 5G [1], an excitingavenue for cooperative localization has opened up.UEs may now directly communicate with one an-other instead of communicating with the BS alonein order to achieve localization of all UE.Due to dedicated communication resources al-located for D2D communication in 5G, UEs mayconduct range and angular measurements on eachD2D link. Since UEs are typically located closerto one another than to BSs, the probability of D2Dlinks being LoS and having higher signal-to-noiseratio (SNR) is greater, providing better positioningaccuracy. In a network with N UEs, up to (cid:0) N (cid:1) additional D2D link measurements are possible.The relative UE location information, extractedfrom the D2D link measurements, may be sentto a central localization unit co-located at one ofthe serving BS or a central server (i.e., centralizedcooperative localization). The position of all theUEs in the network is simultaneously determinedby nonlinear least squares (LS) estimation, whereinthe positions of the UEs that jointly minimizethe deviation from the physical angular orientationand distance-based link constraints are determined.Optimization techniques such as the Levenberg-Marquardt algorithm (LMA) [14], which combinesthe Gauss-Newton algorithm and the method ofgradient descent, may be used for nonlinear LSestimation.Centralized cooperative localization in futuredense IoT networks may lead to network conges-tion if all localization messages are routed to acentral server. In distributed algorithms, UEs arelocalized based on local measurements exchangedby neighboring nodes (as is done in centralizedlocalization). The location estimates of the UEsare then iteratively refined until all neighboringUEs reach an agreement [1]. While not as accurateas infrared methods [3], a root mean square errorof 2.5 m and 3 m was achieved in an indoor environment with centralized and distributed coop-erative localization, respectively, over an area ofapproximately 40 m ×
20 m with four BS withknown locations and 13 unknown UE locations [1].
B. Machine Learning for Localization
In contrast to geometry-based localization algo-rithms, machine learning provides a data-centricview of the UE localization problem. Localizationalgorithms that employ machine learning first cre-ate a “fingerprinting database” of the environmentduring the training (offline) phase [9]. A fingerprintis a vector containing channel parameters such asthe received signal strength (RSS), channel stateinformation (CSI), and the AoA of the strongestsignal of all BS links measured a priori at known lo-cations called reference points , distributed through-out the environment. A fingerprinting database isconstructed by storing the fingerprint measured ateach reference point with the coordinates of thereference point.Once the fingerprinting database is constructed,then in the real-time online position location stepthe BS-UE channel is measured by the UE. Thechannel measurements are matched to the finger-printing database (stored in the UE or in thenetwork) to determine the UE position. Matchingmay be done via maximum a posteriori (MAP)estimation.Alternatively, matching may be performed byutilizing a similarity criterion to compare the onlinemeasurements to the fingerprinting database. Acommon similarity criterion is the distance, such asthe Euclidean ( L ) or the Manhattan ( L ) distance,of the online measurements from the channel mea-surement at the reference points. In the k-nearestneighbor (k-NN) algorithm, the user position isthe weighted average of the k “ nearest ” referencepoints.The UE localization problem can be restated asdetermining the nonlinear function that transformsthe channel parameters into a position estimate.A neural network determines the nonlinear func-tion, based on data available in the fingerprintingdatabase. A neural network is a series of multi-levelnonlinear functional transformations of the input,which can be used to approximate a target func-tion. For user localization, the inputs to the neuralnetwork are the measured channel parameters, andthe target function is the positional coordinates ofthe user. Successive layers of a neural network arecombined linearly by weights. The optimal weightsthat transform the inputs (channel parameters) asclose as possible to the target function (user po-sition) are found in the offline training phase by ig. 3. Map generation on-the-fly and seeing through walls usingnarrow beam antennas and multipath. minimizing the closeness of the output of the neuralnetwork to the target function at the referencepoints.Machine-learning-based localization algorithmsrequire the availability of a dense fingerprintingdatabase, the creation of which is a time-intensiveprocess. The localization accuracy of fingerprint-ing algorithms depends on the distance betweenreference points, with the localization accuracytypically on the order of the distance betweenthe reference points. Additionally, changes in theenvironment such as the addition of new furniturerequire the fingerprinting database to be re-created.Transfer learning may be leveraged to reduce theamount of data required. Theoretical radio wavepropagation models are leveraged to replace datacollection partially by ray tracing. The ray tracer,once calibrated to the environment based on thelimited measurements conducted, may be used topredict channel parameters at the reference points.Minor changes to the propagation environment maybe quickly incorporated into the environment maputilized by the ray tracer, expediting the process ofcreating (and updating) the fingerprinting database.A neural network may be first trained on thesynthetic data generated by the ray tracer, with theweights of the neural network refined by furthertraining on real-world measurements. C. User Tracking and Data Fusion
Localization accuracy of a stationary target maybe improved by averaging the position estimate,reducing the variance of the estimate. For mobiletargets, the location must be estimated in a shorterperiod of time, which can be achieved via usertracking. User tracking refers to continuously es-timating the position of a mobile UE, due to whichsudden changes in the user’s apparent positionfrom one sampling instant to another, caused bypositioning errors, may be smoothed out. Modern cell phones are equipped with a varietyof sensors. UEs possess an inertial measurementunit (IMU), consisting of a gyroscope to measurerotation, an accelerometer to measure acceleration,and a magnetometer to measure the magnetic fieldintensity. Given the initial position of the user, thecurrent user position may be obtained by integrat-ing the measured acceleration twice to get the userposition. However, errors in IMU measurementsgrow with time - a constant offset in accelerationmeasurement leads to a quadratic error in position.Data from the sensors may be fused with channelmeasurement data using a Kalman filter/ extendedKalman filter (KF/EKF) to correct the drift inIMU measurements. A KF is a recursive linearestimator of the state (position and velocity) ofa user. The current state of the user is modeledas a linear transformation of the state of the userat the previous time instant, based on kinematicequations derived from Newton’s laws of motion,whereas sensor measurements are modeled as alinear transformation of the current state of the user.A KF is the optimal estimator of a linear process,given the mean and variance of the noise. If therelation is not linear, an EKF may be used to locallylinearize the process via Taylor series expansion[1]. The KF/EKF minimizes the mean square errorof the position estimate based on measurementsobtained from all sensors up to the current timeinstant. When new information is obtained by theuser in the form of new channel measurements ornew sensor data, the KF/EKF recursively updatesthe position estimate based on the old positionestimate and the new data.
D. Localization Algorithms Exploiting Multipath
As discussed earlier, multipath components areconventionally thought to be a hindrance to ac-curate localization. However, in conjunction witha map of the environment, multipath componentsprovide additional vital useful information regard-ing the location of the UE. For example, with a mapof the environment available (Fig. 3), “forbiddentransitions” of a UE wherein the UE moves throughwalls or from one floor to another in consecutivetime steps may be detected and discarded.Multipath components from the BS may arrive atthe UE via a direct path or via indirect paths alongwhich the source ray suffers multiple reflectionsor scattering. Virtual anchors (VAs) are successivereflections of the BS on walls in the environment[1], which are treated as an LoS BS in place ofthe physical NLoS BS. Future wireless deviceswill exploit real-time ray tracing [10] for multipathpropagation prediction in order to determine the
ABLE I. Summary of the different position location techniques
Position Location Method Description BS Density Deployment Cost Accuracy
Fundamental Techniques Use uplink and downlinkAoA, ToA, TDoA measurementsto calculate position via geometry High Low [10] Low [10]Cooperative Localization Use side-link (UE-UE) measurements tocomplement BS-UE measurements Low Low [1] Medium [1]Machine Learning Channel features mapped tovalues stored in fingerprint database Medium High [9] High [9]User Tracking Refine position estimateof fundamental techniques,predict user trajectory with sensor data Medium Low [1] Medium [1]Multipath Exploiting Techniques Extract position informationembedded in multipath components Low Medium [7], [14] High [7], [14]
VA locations. If the user’s location is continuouslytracked with an EKF, each multipath componentreceived by the UE may be associated with a VAbased on the previously estimated UE location.Once the correspondence between each multipathcomponent and the VAs is known, any of thefundamental localization techniques (AoA, ToA, orTDoA) may be used to localize the UE.With large bandwidths and narrow beamwidths atmmWave frequencies, more multipath componentsare resolvable, which makes the task of associatingthe multipath components with the VA more diffi-cult. Ray tracing may be used to take advantage ofNLoS multipath components arriving at a UE, pro-viding single-shot user location estimation withoutuser tracking. With knowledge of the AoA at theBS, the ToA of the source rays, and a map of thesurrounding environment, the BS may determinethe location of the UE via ray tracing each mul-tipath component. Since it is not known whetherthe signal is reflected or transmitted through eachobstruction along the traced signal path, two possi-ble locations are recursively stored as “candidatelocations” at each obstruction encountered whileray tracing a multipath component. A majority ofthe candidate locations will be clustered near thetrue UE location, so the user may be localized to thecentroid of the largest cluster of candidate locations[7].In the absence of a map, with the assumption thateach multipath component is reflected or scatteredat most one time, the problem of determiningthe location of a UE can be reformulated into anonlinear LS estimation problem [14]. The scat-terer/reflector positions and the UE position andorientation are estimated by jointly finding the scatterer and user locations where the expected dis-tances and angles (geometrically calculated) matchthe measured distances and angles most closely inthe least squared sense. Optimization techniquessuch as particle swarm optimization (PSO) andthe LMA [14] may be used for nonlinear LSoptimization.IV. C
ONCLUSION AND F UTURE R ESEARCH
This article has provided an overview of existingand emerging localization techniques, illustratinghow utilizing the wide bandwidths at mmWavefrequencies could lead to unprecedented localiza-tion accuracies. The narrow antenna beamwidthsat mmWave frequencies require smart beam man-agement, while optimal localization requires anexploration of multipath components arriving fromall directions, for which a detailed study of jointcommunication and localization is required. TableI provides a summary of the different positionlocation methods.Looking into the future, we predict that a combi-nation of machine learning, data fusion of measure-ments from multiple sensors, and cooperative lo-calization will be used for robust, accurate positionlocation. The wireless systems will need to seam-lessly transfer the localization responsibility fromone wireless technology (e.g., WiFi access pointsindoors) to another (e.g., cellular BSs outdoors),similar to handovers in current cellular networkswhen a user moves in and out of BS coverage cells.With centimeter-level localization accuracy infuture cellular networks, privacy will become agrowing concern. Users must be allowed to opt outof tracking if they so desire, and any user locationdata stored in the network must be protected fromackers. Additionally, the localization solution mustbe robust to interference from malicious users, whocould, for instance, attempt to replicate the refer-ence signals transmitted by the cellular network inorder to gain unauthorized access to user locationinformation.The computing capabilities of UEs will enablemapping and ray tracing in real time. We envisagethat cell phones in the future shall generate a mapof the environment on the fly or have maps loadedwithin, enabling map-based localization algorithmsthat exploit real-time multipath propagation. Theaugmentation of human and computer vision willallow users to see in the dark and see throughwalls [7]. Cell phones in the future will have thecapability to either download or generate a mapof the environment on the fly and “see in the dark”[7]. The UE will behave like a radar, measuring thedistances of prominent features in the environment,such as walls, doors, and other obstructions. Ad-ditionally, reflections and scattering off walls willenable cell phones to view objects around cornersor behind walls [7], as illustrated in Fig. 3.For ranging measurements, a radar operates inthe pulsed radar mode, wherein the radar transmitsa single pulse, switches from transmit to receivemode, and waits for the echo back from the objectthat is to be range-estimated. However, due toconstraints on switching speed, only objects at asufficient distance from the user may be ranged.For example, an mmWave phased array with aTX-RX switching time of ∼
100 ns cannot rangeobjects closer than 50 ft (electromagnetic wavestravel 1 ft/ns). To range closer objects, a UEmust simultaneously transmit and receive the radarsignal, operating in the full duplex mode, requiringTX-RX isolation [15].V. A
CKNOWLEDGMENTS
This work is supported by the NYU WIRELESSIndustrial Affiliates Program and National ScienceFoundation (NSF) Grants 1702967, 1731290, and1909206. R
EFERENCES [1] T. Pedersen and B. H. Fleury, “Whitepaper on New Local-ization Methods for 5G Wireless Systems and the Internet-of-Things,”
COST Action CA15104, IRACON , Apr. 2018.[2] H. Guan, J. Li, S. Cao, and Y. Yu, “Use of mobile LiDAR inroad information inventory: A review,”
International Journalof Image and Data Fusion , vol. 7, no. 3, 2016, pp. 219–242.[3] C. D. McGillem and T. S. Rappaport, “A beacon navigationmethod for autonomous vehicles,”
IEEE Transactions on Ve-hicular Technology , vol. 38, no. 3, Aug. 1989, pp. 132–139.[4] C. Falsi, D. Dardari, L. Mucchi, and M. Z. Win, “RangeEstimation in UWB Realistic Environments,” in , June 2006, pp. 1–6. [5] T. S. Rappaport, J. H. Reed, and B. D. Woerner, “Positionlocation using wireless communications on highways of thefuture,”
IEEE Communications Magazine , vol. 34, no. 10, Oct.1996, pp. 33–41.[6] 3GPP, “E-UTRA; Requirements for support of radio resourcemanagement (Release 15),” TS 36.133 V15.9.0, Jan. 2020.[7] T. S. Rappaport et al. , “Wireless Communications and Appli-cations Above 100 GHz: Opportunities and Challenges for 6Gand Beyond,”
IEEE Access , vol. 7, June 2019, pp. 78 729–78 757.[8] C. A. Balanis,
Antenna theory: analysis and design , 4th ed.Hoboken, New Jersey: John Wiley & Sons, 2016.[9] J. Gante, G. Falc˜ao, and L. Sousa, “Deep Learning Architec-tures for Accurate Millimeter Wave Positioning in 5G,”
NeuralProcessing Letters , vol. 51, no. 1, Aug. 2020, pp. 487–514.[10] O. Kanhere and T. S. Rappaport, “Position Locationing forMillimeter Wave Systems,” in
Proc. IEEE GLOBECOM 2018 ,Dec. 2018, pp. 1–6.[11] J. Xiong and K. Jamieson, “ArrayTrack: a fine-grained indoorlocation system,” in
USENIX, NSDI , Apr. 2013.[12] J. Schroeder, S. Galler, K. Kyamakya, and K. Jobmann,“NLOS detection algorithms for Ultra-Wideband localization,”in , March 2007, pp. 159–166.[13] C. Huang et al. , “Machine Learning-Enabled LOS /NLOSIdentification for MIMO Systems in Dynamic Environments,”in
IEEE Transactions on Wireless Communications , vol. 19,no. 6, June 2020, pp. 3643-3657.[14] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados,and H. Wymeersch, “Position and Orientation EstimationThrough Millimeter-Wave MIMO in 5G Systems,”
IEEETransactions on Wireless Communications , vol. 17, no. 3,March 2018, pp. 1822–1835.[15] J. Zhou, N. Reiskarimian, J. Diakonikolas, T. Dinc, T. Chen,G. Zussman, and H. Krishnaswamy, “Integrated full duplexradios,”
IEEE Communications Magazine , vol. 55, no. 4, Apr.2017, pp. 142–151. B IOGRAPHIES O JAS K ANHERE received the B.Tech. andM.Tech. degrees in electrical engineering from IITBombay, Mumbai, India, in 2017. He is currentlypursuing the Ph.D. degree in electrical engineeringwith the NYU WIRELESS Research Center, NewYork University (NYU) Tandon School of Engi-neering, Brooklyn, NY, USA, under the supervisionof Prof. Rappaport. His research interests includemmWave localization and channel modeling.T
HEODORE
S. R