millimeter waves and vehicle to vehicle and infrastructure ......ch15-p372533.tex 14/5/2007 18: 10...

© 2021 Robert W. Heath Jr. and Nuria Gonzalez Prelcic

Millimeter Waves and Vehicle to Vehicle and Infrastructure Communications

(V2X) Professor Robert W. Heath Jr.

Joint work with Professor Nuria Gonzalez-Prelcic

Thanks to sponsors including the National Science Foundation ECCS-1711702, CNS-1702800, and CNS-1731658, Toyota, Honda, Samsung and Nokia.Figures used in the presentation are copyright by the authors and may not be reused without permission.

www.6GNC.org

6GNC


2

Our vision at 6GNC www.6GNC.org

V2X


3

Technical expertise in 6GNC

.

.

.

...

....

.

.

.

.

.

LNA

LNA

LNA

Antenna and circuits SP and AI/ML Networking

Developing V2X will require continued collaborations with transportation experts

www.6GNC.org


4

Connectivity as a driver for next generation vehicles


Vehicles are becoming smarter

5[1] 5G-PPP White Paper on Automotive Vertical Sector, October 2015, https://5g-ppp.eu/white-papers/

Self driving

310 2Driver assist

54


V2X for advanced driver assistance systems

Sensors require line-of-sight

Communication can expand sensing range

“See through”

Low latency but modest data rate requirements for alerting driver

High data rate if “see through” capability is included

Both communication and automotive sensors are useful

for collision avoidance



Higher levels of traffic coordination like platooning

Reduces braking shockwaves due to

congestion

Low latency but low rate connectivity may be sufficient

V2X for traffic efficiency

More efficient use of intersections



V2X for fully automated driving

8810

Sharing local sensors information ~ 1Gbps per user

Full automation requires Gbps data rates and ms latencies

Enables cloud control of vehicles through

intersections or congested areas

Exchanging raw sensor data provides information for fully automated safety-

critical functions

Downloading high-definition 3D map data (~Gbyte) for precise

navigation

[1] 5G-PPP White Paper on Automotive Vertical Sector, October 2015, https://5g-ppp.eu/white-papers/


V2X for infotainment

9

Infotainmentapplications increase with

higher levels of automation

Multimedia and gaming100x Mbps - Gbps

Mobile base station for passengers

High rate and low latency Internet access required to keep passengers happy



Fifth generation (5G) cellular communication

10

Multidimensional objectives*

* Recommendation ITU-R M.2083-0, “IMT Vision – Framework and overall objectives of the future development of IMT for 2020 and beyond,” September 2015** “5G empowering vertical industries,” 5GPPP White Paper, Feb. 2016

Lower latency

Higher rates

Mobility

Spectrum efficiency

Userexp.

data rate

Peak data rate

Energyefficiency

Area traffic

capacity

LatencyConnectiondensity

New industry verticals**

Automotive industry provided requirements for the evolution of 5G and

Aut

omot

ive

Med

ia &

Ent

erta

inm

ent

e-H

ealth

Fact

ory

of t

he F

utur

e

Ener

gy


Connectivity options

< 1ms 1ms 10ms 100ms

1 Mbps

10 Mbps

100 Mbps

1 Gbps

throughput

latency

DRIVEN BYTHE CLOUD

DRIVEN BYTHE IN-CAR INFO

DATA

positioninfo

limited sensor

processed sensor

raw sensor

Mixed levels of automation

11

More advanced communication capability is required for future

applicationsmmWave


MmWave bands will enable Gbps V2X data rates

12

12MmWave is the only viable approach for high bandwidth connected vehicles [1]

V2V communication

directional beamformingblockage

V2I

com

mun

icat

ion

Many simultaneous connections allowed

thanks to spatial reuse with narrow beams

High data rates due to high bandwidth

communication channels

Applications to vehicle-to-vehicle and vehicle-to-infrastructure

Ultra low latency easier to support due to smaller packet sizes

[1] Junil Choi, Vutha Va, Nuria González-Prelcic, Robert Daniels, Chandra R. Bhat, and Robert W. Heath Jr, “Millimeter Wave Vehicular Communication to Support Massive Sensing”, IEEE Communications Magazine, vol. 54, no. 12, pp. 160-167, December 2016.


13

Challenges of 5G-NR


Beam-based design

14

3 GHz aperture

30 GHz aperture

TX

RX

Expanded aperture withantenna array

Many antennas are used at higher frequencies to give larger effective aperture

through the virtues of beamforming

5G adopts a beam-based design to support analog beamforming at base

station and user equipment

Requires beam acquisition and tracking


Beam-training incurs a high overhead

15

SS burst size (5ms)

SS burst periodicity (20 ms) SS block

Analog architecture

64 UE antennas

64 gNB antennas

64 SS blocks per burst1.28 sec forInitial Access

64*64/64*20e-3=1.28 seconds

[1] 3GPP “NR - Physical layer procedures for control - Release 15,” TS 38.213, 2018.[2] 3GPP “NR - Physical layer measurements - Release 15,” TS 38.215, 2017.

More than 300 ms with 16 UE antennas


Other architectures are possible

16[1] E. Onggosanusi et al., "Modular and High-Resolution Channel State Information and Beam Management for 5G New Radio," in IEEE Communications Magazine, vol. 56, no. 3, pp. 48-55, March 2018.

Analog beamforming supports single user and single stream

Hybrid beamforming supports few beams and streams / users with more RF complexity

Beam training overhead greatly depends on the MIMO architecture choice

1-bitADCADC

1-bitADCADC

BasebandCombining

RF Combining

RFChain

RFChain

BasebandRFChain

RFainADC


17

AI/ML for beam training in V2X

VuthaVa, Junil Choi, Takayuki Shimizu, Gaurav Bansal, and R. W. Heath, Jr., `` Inverse Fingerprinting for Millimeter Wave V2I Beam Alignment,'' IEEE Trans. on Veh. Tech., vol. 67, no. 5, pp. 4042-4058, May 2018.

V. Va, T. Shimizu, G. Bansal and R. W. Heath, "Online Learning for Position-Aided Millimeter Wave Beam Training," in IEEE Access, vol. 7, pp. 30507-30526, 2019.

N. J. Myers, Y. Wang, N. González-Prelcic, and R. W. Heath, “Deep Learning-Based Beam Alignment in MmwaveVehicular Networks,” in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2020, pp. 8569–8573.


18

Machine Learning Billboard Collection

(c) Robert W. Heath Jr. and Nuria González-Prelcic 2018

Machine learning-for-everything

(ML4X)!!!…when you need to spice up your research

ML4X

Warnings: Claims made here are based on absolutely no facts. Applying machine learning to your research may not lead to more publications, more funding, or a higher citation index. Using machine learning might cause periodic frustration due to the challenges of getting the software to actually work, the search for large enough data sets, or the constant acquisition of more processing power.

I started working on ML4X and my h-index increased 10 points!!!

Do you still think that ML stands for maximum likelihood?

Are your papers getting the visibility you want?

Is your work missing that energy associated with the next hot topic?

more data, more fun

From R. W. Heath and N. González-Prelcic, "Machine-Learning Billboard Collection [Humor]," in IEEE Signal Processing Magazine, vol. 36, no. 1, pp. 176-176, Jan. 2019.


Vehicular applications of millimeter wave

19

Challenge Use of beam training in 5G and high mobility requires

high overhead to point beams

Irregular motion of user with phone is hard to

exploit

The regular motion in V2X is an opportunity to improve the efficiency of millimeter wave communication systems


20

Huge computation resources @ BS/devices

Edge vehicles

Cloud BSs

Data from previous transmission records

Situational awareness, regularity of environment

20

!

Opportunities in ML4V2X: data-driven solutions

Data-driven solutions are promising for mmWave vehicular beam alignment20


Leveraging side information using machine learning

21

Position& orientation

PositionOrientation

Performance ofbeam pattern

DB

[1] Vutha Va,J. Choi, Takayuki Shimizu, Gaurav Bansal, and R. W. Heath, Jr., ``Inverse Fingerprinting for Millimeter Wave V2I Beam Alignment,'' IEEE Trans. on Veh. Tech., vol. 67, no. 5, May 2018. [2] Vutha Va, Takayuki Shimizu, Gaurav Bansal, and R. W. Heath, Jr., “Online Learning for Position-Aided Millimeter Wave Beam Training,” IEEE Access, 2019.

Recommend a subset of beams based on past performance data


Rates with large arrays

22

Rates keep increasing with the array size

The value of fingerprint database increases with blockage probability

Proposed method can support large arrays at high speeds

Position only

training data

Beam coherence time

No beam training overhead

UPA 16x16

Not enough time to complete training

8x8

16x16

24x24

32x32

Proposed w/ 28 beams


23

Evolving infrastructure A. Ali, N. Gonzalez-Prelcic, R. W. Heath and A. Ghosh, "Leveraging Sensing at

the Infrastructure for mmWave Communication," in IEEE Communications Magazine, vol. 58, no. 7, pp. 84-89, July 2020.

Y. Wang, A. Klautau, M. Ribero, A. C. K. Soong and R. W. Heath, "MmWaveVehicular Beam Selection With Situational Awareness Using Machine Learning," in IEEE Access, vol. 7, pp. 87479-87493, 2019.

M. L. Rahman, J. A. Zhang, X. Huang, Y. J. Guo and R. W. Heath, "Framework for a Perceptive Mobile Network Using Joint Communication and Radar Sensing," in IEEE TAES 2020.


24

Explosion of sensors to support vehicles

https://fingfx.thomsonreuters.com/gfx/rngs/UBER-SELFDRIVING-SENSORS/010061BR2TH/UBER-SELFDRIVING-SENSORS.jpg

DSRC

Camera

Radar

Lidar

Cellular Edgeprocessing

Sensing expands situational awareness, has many applications to automation

Near term: more sensors on vehicles Longer term: sensors collocated with BS

A. Ali, N. Gonzalez-Prelcic, R. W. Heath and A. Ghosh, "Leveraging Sensing at the Infrastructure for mmWave Communication," in IEEE Communications Magazine, vol. 58, no. 7, pp. 84-89, July 2020.M. L. Rahman, J. A. Zhang, X. Huang, Y. J. Guo and R. W. Heath, "Framework for a Perceptive Mobile Network Using Joint Communication and Radar Sensing," in IEEE TAES 2020.


25

Sensors also can aid communication

First measurement location

Second measurement location

APSAPS

A. Ali, N. Gonzalez-Prelcic, R. W. Heath and A. Ghosh, "Leveraging Sensing at the Infrastructure for mmWaveCommunication," in IEEE Communications Magazine, vol. 58, no. 7, pp. 84-89, July 2020.

A. Graff, A. Ali, and N. González-Prelcic, “Measuring radar and communication congruence at millimeter wave frequencies,” in Proc. Asilomar conference on signals, systems, and computers, Nov. 2019, pp. 925–929.

Congruence between radar and

communication channels


Sensing makes the

infrastructure more valuable

26

A. Ali, N. Gonzalez-Prelcic, R. W. Heath, and A. Ghosh, “Leveraging Sensing at the Infrastructure for mmWaveCommunication,” IEEE Communications Magazine, vol. 58, no. 7, pp. 84–89, Jul. 2020


Exploiting sensors for communications

27N. González Prelcic, et. al., ``Millimeter Wave communication with out-of-band information,'' IEEE Com Mag, Dec. 2017.Part of Fig 15.1 from E. Dahlman, S. Parkvall, J. Sköld, and Per Beming, 3G Evolution: HSPA and LTE for Mobile Broadband, Academic Press, 2007.

Motion control

Path planning

Mapping

Localization

Obstacle detection and classification

Ch15-P372533.tex 14/5/2007 18: 10 Page 300

300 3G Evolution: HSPA and LTE for Mobile Broadband

Segmentation, ARQ

Ciphering

Header compression

Hybrid ARQ

MAC multiplexing

Antenna and resource mapping

ModulationAntenna and resource assignment

Modulation scheme

MA

C s

ched

uler

Retransmission control

Priority handling, payload selection

Payload selection

RLC #i

PHY

PDCP #i

MAC

Concatenation, ARQ

Deciphering

Header compression

Hybrid-ARQ

MAC demultiplexing

Antenna and resource demapping

Decoding

Demodulation

RLC

PHY

PDCP

MAC

eNodeB Mobile terminal (UE)

Red

unda

ncy

vers

ion

Coding

Transport channel

SAE bearers

Radio bearers

Logical channels

User #i User #j

IP packet IP packet

Figure 15.1 LTE protocol architecture (downlink).

The header-compression mechanism is based on ROHC [64], a standardizedheader-compression algorithm used in WCDMA as well as several othermobile-communication standards. PDCP is also responsible for ciphering andintegrity protection of the transmitted data. At the receiver side, the PDCP pro-tocol performs the corresponding deciphering and decompression operations.There is one PDCP entity per radio bearer configured for a mobile terminal.

• Radio Link Control (RLC) is responsible for segmentation/concatenation,retransmission handling, and in-sequence delivery to higher layers. UnlikeWCDMA, the RLC protocol is located in the eNodeB since there is only a sin-gle type of node in the LTE radio-access-network architecture. The RLC offersservices to the PDCP in the form of radio bearers. There is one RLC entity perradio bearer configured for a terminal.

• Medium Access Control (MAC) handles hybrid-ARQ retransmissions anduplink and downlink scheduling. The scheduling functionality is located inthe eNodeB, which has one MAC entity per cell, for both uplink and downlink.

LTE protocol architecture Automated vehicle planning

How to connect sensing and communication stacks?

Decision making


28


Exploiting knowledge of vehicle locations

29

(x, y)✓

r

Occupancy gridCartesian Polar

1st lane

2nd lane

Encode the geometry into different coordinates

Set the RX location as coordinate origin

x

y

v = [r, t1, t2, c1, c2]Location of RSU 1st & 2nd lane trucks 1st & 2nd lane cars

Order vehicles based on types and locations

First lane vehiclesTrucks with larger sizeVehicles closer to the RX

Have larger impactson reflections& beam selection

Constrain the number of vehicles in features as Nveh

Cartesian coor., e.g.,

t1 = [xi1 , yi1 , xi2 , yi2 , · · · , xiN , yiN ],

|xi1 | < |xi2| < · · · < |xiN |.

Vehicles on each lane of a certain type are ordered based on the relative distance to RX

Y. Wang, A. Klautau, M. Narasimha, and R. W. Heath Jr., "MmWave beam selection with situational awareness", in preparation for submission.


Smart link configuration

3011

Adding 1st lane trucks

Adding 2nd lane trucksAdding cars

PathsAoA

AoD

Two-lane urban street

Two-types of veh.:Trucks & cars

AoDs

PathsRay tracingoutput

AoAs

One road-sideinfrastructure

Roadside unit

RSU

H[n] =p

NtNr

LpX

`=1

g(nT � ⌧`)ar(�A` )a

⇤t (�

D` )↵`e

j�` ,

0 n Lc � 1,

Situational awareness can be leveraged for fast vehicular beam alignment


Lidar aided beam training

31

3

to 1⇥1, trained with Kera’s Adadelta optimizer [14]. We usedpooling layers and, to mitigate overfitting, regularization anddropout. For beam-selection, the values in (1) below 6 dB fromthe maximum were zeroed and normalized to have unitarysum. For top-M classification, the output layer had a softmaxactivation function and a categorical cross-entropy as lossfunction [14]. For binary classification, the output layer andloss were sigmoid and binary cross-entropy, respectively [14].The number of parameters per network is approximately 105.To promote reproducibility, we share code and data at [18].

As a baseline for comparing with DL applied to the LOSdecision problem, we also evaluated a simple geometric ap-proach: given Pb and Pv , we calculate the line L connectingthem. We denote by d̂ the minimum distance between any pointin C to L. A decision stump classifier [14] uses a threshold �

to decide for NLOS if d̂ < � or LOS otherwise. The intuitionis that if L is far from all obstacles identified by the LIDARin C, the link is potentially LOS.

IV. NUMERICAL RESULTS

A. Simulation methodologyAiming at realistic datasets, we adopted a simulation

methodology using traffic, ray-tracing and LIDAR simulatorsin V2I mmWave communications [16]. We paired the sim-ulations of the mmWave communication system and the LI-DAR data acquisition integrating three softwares: the BlenderSensor Simulation (BlenSor) [19], the Simulation of UrbanMObility (SUMO) traffic simulator [20], both open source, andRemcom’s Wireless InSite for ray-tracing. In the configurationstage, the user provides information about the objects in the3D scenario, lanes coordinates, eletromagnetic parameters, etc.The software execution is based on a Python orchestrator codethat invokes SUMO and converts its ouputs (vehicles positions,orientations, etc.) to formats that can be interpreted by distinctsimulators. The orchestrator then invokes the simulators (LI-DAR and ray-tracing in this case) to obtain paired results.

Fig. 3. a) Urban canyon 3D scenario with vehicles of distinct sizes randomlypositioned. The building color indicates height and corresponds to a rangefrom 0 (blue) to 101 meters (red). b) Corresponding LIDAR point cloud. TheLOS rays between the BS antenna at z = 4 m and vehicle are shown.

Fig. 3a depicts the adopted urban canyon 3D scenario,which is part of Wireless InSite’s examples and representsa region of Rosslyn, Virginia. The study area is a rectangle ofapproximately 337⇥ 202 m2 and the BS antenna array heightis z = 4 m. We placed receivers and LIDARs on top of all

connected vehicles (identified in red) in each scene snapshot.Fig. 3b illustrates an example of the corresponding LIDARpoint cloud. Lines between BS and vehicle are also shown,and suggest a LOS channel.

The ray-tracing simulations used a maximum of L = 25MPCs per transmitter / receiver pair, isotropic antennas,60 GHz carrier frequency, B = 100 MHz, K = 64 subcarriers,and enabled ray-tracing diffuse scattering. Other parametersfollowed the ones in [16].

The downlink mmWave massive MIMO relied on a BSwith a 16⇥ 16 uniform planar array (UPA) and vehicles with4 ⇥ 4 UPAs. When designing Ct and Cr, we first augmentedthe conventional DFT codebook with steered codevectors,linear combinations of codevectors and random vectors fromGrassmannian codebooks. From this large initial set, we keptonly the codevectors that were chosen as [(p, q) more than 100times in the training set. This procedure led to |Ct| = 20 and|Cr| = 12, respectively. Hence, the number of classes for top-M classification is 240.

The LIDAR simulations assumed a Velodyne model HDL-64E2 scanner positioned at a height z = 1 m from the top-center of the vehicle. The angle resolution was 0.1728 degreesand the rotation speed 10 Hz. The experiments adopted bx =by = 6 and bz = 3 bits. We eliminated from C the pointswith small values in the z-axis (< 0.1 m), which correspondto ground reflections (see Fig. 3b), and also the points with adistance from the LIDAR larger than dmax = 25 m.

A preliminary investigation indicated consistent beam-selection results for distinct signal-to-noise ratios and, for sim-plicity, the results here are for a noise-free mmWave channel.But we considered two conditions with respect to positioningaccuracy: noise-free and noisy. The LIDAR noise [19] is as-sumed to have independent components distributed accordingto a zero-mean Gaussian N (0,�2

L/3) with variance �

2L/3.

For the noisy condition, we adopted the HDL-64E2 defaultvalue of �L = 0.1 m. Similarly, the accuracy of the GlobalNavigation Satellite System (GNSS) technology is modeledassuming the elements of the position error vector are inde-pendent and identically distributed according to N (0,�2

G/3)

(no bias). Conventional GPS may lead to errors of 3 to 5 m,while sophisticated SLAMs can help to keep the error below50 cm in the horizontal plane [17]. For the noisy condition,we assumed �G = 3 m and �L = 0.1 m.

Beam-selection is harder in NLOS because the predictabilitydecreases considerably when compared to LOS cases. If anexperiment considers both LOS and NLOS channels, theaccuracy of ML will depend on the blockage probability,which is heavily influenced by traffic statistics, large vehicles(potential blockers) and antenna height. Numerical results ofdistinct experiments that used mixed LOS and NLOS areharder to compare and the ML models may be biased by theeasier LOS cases. To avoid this situation, we present separateevaluations of beam-selection for each case. The mmWavedata is composed of NL = 6, 482 LOS and NN = 4, 712NLOS channel examples. The beam-selection experimentsused NL and NN examples in the LOS and NLOS evaluations,respectively, while LOS detection used all NL+NN examples.For all experiments we created disjoint test and training sets

Prediction with 30 pairs

2

antenna arrays with only one radio frequency (RF) chainand fixed beam codebooks. To simulate the channel, we useray-tracing data and combine the ray-tracing output with awideband mmWave geometric channel model as, e. g., in [10].Assuming Rc multipath components (MPC) per transmitter /receiver pair, the information collected from the outputs for ther-th MPC of a given pair is: complex path gain ↵r, time delay⌧r and angles �

D

r, ✓D

r, �A

r, ✓A

r, corresponding respectively to

azimuth and elevation for departure and arrival. The frequency-selective channel model at the time instant corresponding tothe n-th symbol vector is described in detail in [Section III][10], which also includes the definition of the model in thefrequency domain H[k], where k is the subcarrier index.

We assume beam codebooks Ct = {f1, · · · , f|Ct|} andCr = {w1, · · · ,w|Cr|} at the transmitter and the receiversides, with no restriction on the codebook size (e. g., theydo not have to be DFT codebooks). For a given pair (p, q)of vectors, representing precoder fp and combiner wq , thereceived signal at subcarrier k is s[k] = w

H

qH[k]fp, where

H denotes conjugate transpose. The beam-selection is guidedby the normalized signal power

y(p,q) =K�1X

k=0

|wH

qH[k]fp|2 (1)

and the optimum beam pair is [(p, q) = argmax(p,q) y(p,q). Inthis paper, the goal of beam-selection is to recommend a setB = {(pi, qi)}Mi=1 such that [(p, q) 2 B.

III. MACHINE LEARNING USING LIDAR DATA

A. Information exchange protocolWe develop a ML-based beam-selection strategy for V2I

mmWave cellular communication system, assuming that theconnected vehicle is equipped with a LIDAR. The proposedML-based protocol is illustrated in Fig. 1. It is assumed thatthe BS can broadcast its absolute position Pb = (xb, yb, zb)for mmWave V2I beam alignment of incoming vehicles usinga CC provided by, for instance, DSRC signals or as partof the BS CC [6]. A vehicle estimates its position Pv =(xv, yv, zv) using for example, Global Positioning System(GPS) or a simultaneous localization and mapping (SLAM)algorithm [17]. The BS also broadcasts its coverage zoneZ = (xb

z, y

b

z, x

e

z, y

e

z, h), which is a cuboid specified by its

height h, and points (xb

z, y

b

z) and (xe

z, y

e

z) denoting the base.

The ML algorithm is executed at the vehicle and outputs aset B of beam pairs. After the initialization stage (see Fig. 1),the BS transmits using the beams specified by {pi}Mi=1 andthe M pairs of beams are evaluated at the vehicle accordingto (1). The best pair is then fed back to the BS. If beamcorrespondence can be assumed, the same beam pair canbe used for uplink. Once mmWave communication linksare established, the overhead information required by beamtracking can rely on the high data rates of mmWave links.

B. LIDAR-based feature extraction and deep learningWe use ML to tackle two distinct problems. The first is

the use of only LIDAR data for LOS versus NLOS binary

Fig. 1. Timing diagram for the distributed LIDAR-based beam-selectionmethod. The first phase (broadcast and initialization) uses a low-frequencyCC while the second corresponds to mmWave communication.

classification. The second problem is the selection of the top-M beam pairs based on (1) for decreasing the beam-selectionoverhead, which is associated with the protocol explained inthe previous subsection. The raw input data to solve bothproblems is composed of the LIDAR point cloud C collectedby the vehicle, the BS coverage zone Z and positions Pv andPb. The LIDAR cloud C is an array of dimension D ⇥ 3,composed of 3D points indicating the presence of obstacles.Typical values of D are relatively large and using an alternativerepresentation helps to control the computational cost.

Fig. 2. Feature extraction of 3D histogram G from LIDAR data.

We adopt a fixed grid G to represent Z, as depicted in Fig. 2.We use G as a 3D histogram in which a bin corresponds toa fixed region of Z. Each element of G stores the number ofelements of C within the corresponding bin. A large countof occurrences indicates that LIDAR detected many pointswithin the bin. This histogram calculation was implementedas the uniform quantization of elements of C using (bx, by, bz)bits, with Z providing the quantizers’ dynamic ranges. Beforequantization, the relative distances of obstacles provided by Care converted into absolute positions C + Pv . For example,considering only the x-axis for simplicity, if x

b

z= 10 m,

xe

z= 22 m and bx = 2 bits, the four bins are represented

by values 10, 14, 18 and 22. Assuming Pv = (5, y0, z0), anelement (16, y00, z00) in C is mapped to (21, y0 + y

00, z

0 + z00).

The value 21 is then quantized to 22, and this point accountedby incrementing a bin of G that corresponds to the last (4th)position in the x-axis. We discard points that are farther fromPv by more than a certain distance dmax. The ML input featureis then a histogram G with dimension 2bx ⇥ 2by ⇥ 2bz .

For both problems (LOS decision and beam-selection), weadopted neural networks with 13 layers from which 7 are 2Dconvolutional layers with decreasing kernel sizes, from 13⇥13

Lidar point cloud

Urban canyon in ray tracing simulation

Vehicle equipped with lidar

Selection of top beam pairs using deep learning [1]

[1] Aldebaro Klautau, Nuria González-Prelcic and Robert W. Heath Jr. “LIDAR Data for Deep Learning-Based mmWave Beam-Selection”, IEEE Wireless Communication Letters, 2019.[2] Marcus Dias, Aldebaro Klautau, Nuria González-Prelcic and Robert W. Heath Jr. “Position and LIDAR-Aided mmWave BeamSelection using Deep Learning”, in Proc. of SPAWC 2019.

TopM-30 LOS TopM-30 NLOS0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Accu

racy

LIDAR DistributedLIDAR CentralizedPosition (penetration=100%)Position (penetration=90%)Position (penetration=50%)

LOS NLOS


Need hybrid models that trade interpretability, accuracy and speed

32

Obtaining meaningful data sets is a challenge

Drones PedestriansVehicles

5G system deployment is too late to start collecting data for 6G

Need to obtain communication and sensing data at the appropriate spatial / frequency / temporal -scales with enough detail but not a computational burden

Some data sets like Raymobtime are available, including multi-modal, used in ITU challenge

https://www.lasse.ufpa.br/raymobtime/https://research.ece.ncsu.edu/ai5gchallenge/

A. Klautau, P. Batista, N. González Prelcic, Yuyang Wang, and R. W. Heath, Jr., ``5G MIMO Data for Machine Learning: Application to Beam-Selection using Deep Learning,'' (invited) Proc. of the Information Theory and Applications, San Diego, California, February 11-16, 2018.

Figure from Prof. Klautau

https://www.lasse.ufpa.br/raymobtime/

https://research.ece.ncsu.edu/ai5gchallenge/


33

ITU Challenge AI/ML for 5GSite specific-channel estimation with hybrid mmWave MIMO systems

Prof. Nuria González Prelcic

ML5G-PHY[channel estimation]

Joint work with Prof. Robert Heath, Prof. Aldebaro Klautau and Prof. Ismail Guvenc

https://research.ece.ncsu.edu/ai5gchallenge/


34

Combining radar and communications

P. Kumari, J. Choi, N. González-Prelcic, and R. W. Heath, “IEEE 802.11ad-Based Radar: An Approach to Joint Vehicular Communication-Radar System,” IEEE Transactions on Vehicular Technology, vol. 67, no. 4, pp. 3012–3027, Apr. 2018.

P. Kumari, S. A. Vorobyov, and R. W. Heath, “Adaptive virtual waveform design for millimeter-wave joint Communication–Radar,” IEEE Transactions on Signal Processing, vol. 68, pp. 715–730, 2020.

M. L. Rahman, J. A. Zhang, X. Huang, Y. J. Guo and R. W. Heath, "Framework for a Perceptive Mobile Network Using Joint Communication and Radar Sensing," in IEEE TAES 2020.


Milestones for mmWave systems in civilian applications

35

Year

[1] H. H. Meinel and Juergen Dickmann. "Automotive radar: From its origins to future directions." Microwave Journal, 56(9), 2013.[2] H. H. Meinel, "Commercial applications of mmWaves: history, present status, and future trends." IEEE Trans. on MTT 43, no. 7, pp. 1639-1653, 1995.

1974 20081998 20092005 2016

AEG-Telefuken for

distance warning radar

Mercedes-Benz(MB) ACC

radar

MB distronic plus using long- and

short- range radars

Wilocity chipset802.11adWiGig

WLAN

WirelessHDusing CMOS

WPAN

1979

OKI radio for telephone/video

5G cellular

Soli technology in Pixel 4 for Motion Sense

2012

Tesla autopilot 8 using radar as primary sensor

2020

Bosch long-range radar using SiGe

Figure not drawn to scale

Rad

arC

omm

unic

atio

n

2018

Joint communication-radar for smart

devices and vehicles

Traditionally, mmWave communications and radars developed independently


Many types of joint radar and communication

36Fig. 1. Proposed System Model.

algorithms. It reconstructs clutter using simple recur-sive computation, and allows separation of signals withlargely separated Doppler frequencies. We also provideclosed-form expressions to show how the reconstructionperformance and noise are related to the parameters inthe recursion equation. This method is not only capableof removing clutter but also has the potential of dividingmultipath signals into different groups according to theirDoppler shift values.

The rest of this paper is organized as follows: InSection II, we introduce the system platform for the per-ceptive mobile network. In Section III, we provide math-ematical models for the sensing problems. In Sections IVand V, the direct and indirect sensing schemes are pre-sented, respectively. Section VI presents the backgroundsubtraction method for clutter suppression. In Section VII,simulation results are provided to validate the effective-ness of the proposed framework and sensing algorithms.Section VIII concludes the paper.

Notations: (·)H , (·)T and (·)c denote the Hermitiantranspose, transpose and conjugate of a matrix/vector, re-spectively. | · · · | denotes the element-wise absolute value,(A)n,m denotes the (n, m)th element of the matrix A, (A)·,mand (A)m,· denotes the mth column and row of A, respec-tively, {an} denotes a vector with elements an, diag{an}denotes a diagonal matrix with diagonal elements an.

II. SYSTEM PLATFORM FOR THE PERCEPTIVE MOBILENETWORK

Our proposed system platform aligns with the specifica-tion of the evolution of mobile networks, such as 5G. In thissection, we describe the system model, the supported sens-ing operations and the required modifications to existingmobile communication infrastructure.

A. System Model

We assume a cloud-radio-access network (CRAN) ar-chitecture using multiuser-MIMO and OFDMA technolo-gies. Fig. 1 shows the CRAN architecture-based systemmodel of the proposed perceptive mobile network. In thismodel, cooperative remote radio units (RRU), are denselydistributed and synchronized in clock. Signal processing

for both cellular communication and radio sensing basedon collected signals from these RRUs is done centrally inCRAN central, which includes the baseband unit (BBU)pool for communication and the sensing processing unit.We assume that cooperative RRUs are within the signalcoverage area of each other. All RRUs’ clocks are synchro-nized, typically via GPS. A typical communication sce-nario is as follows: several RRUs work cooperatively toprovide connections to mobile stations (MSs), using mul-tiuser MIMO techniques over the same subcarriers. Whilewe consider CRAN it could work for a standalone base sta-tion (BS) too. So hereafter we will use CRAN central andBS without differentiating them.

We focus on the case where radio sensing is conductedin the BS, although MS-side sensing is also possible. Com-pared to MS, BS has advantages of networked connection,flexible cooperation, large antenna array, powerful compu-tation capability, and known and fixed locations.

B. Supported Sensing Operations

In the perceptive mobile network, the transmitted sig-nal from BSs or mobile stations (MSs) is used for bothcommunication and sensing. The signal may be optimizedjointly for the two functions, and one example is availablefrom [15]. We define uplink and downlink sensing, to beconsistent with uplink and downlink communications. Inuplink sensing, the used sensing signal is from MSs. Indownlink sensing, the sensing signals are from BSs. Thedownlink sensing is further classified as Downlink ActiveSensing and Downlink Passive Sensing, for the cases whena RRU collects the echoes from its own and from otherRRUs transmitted signals, respectively.

It is important to note that in a distributed antenna sys-tem such as CRAN, sensing is for the environment sur-rounding a specific transmitter and receiver, and hence it isseparately done for each node (RRU in CRAN), althoughsome joint processing is possible.

1) Downlink Active Sensing: We refer to downlink ac-tive sensing as the case that a RRU uses reflected downlinkcommunication signals from its own transmitted signal forsensing. In this case, similar to a monostatic radar, transmit-ter, and receiver are co-located although they may have twoindependent antennas separated in space. This will enablea RRU to sense its surrounding environment.

2) Downlink Passive Sensing: Passive sensing is typ-ically referred to the case when a third receiver outside thecommunication system exploits the communication signalfor sensing. Here we use downlink passive sensing for thecase where a RRU uses the downlink communication sig-nals received from other RRUs for sensing. Depending onthe distance between RRUs, reflected signals from otherRRUs or the RRU itself may arrive at different time seg-ment or overlapped. Downlink passive sensing senses theenvironment among RRUs.

3) Uplink Sensing: BS uses the uplink communi-cation signal from MS transmitters for uplink sensing.Uplink sensing estimates relative, instead of absolute, time

1928 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 3 JUNE 2020

Authorized licensed use limited to: University of Texas at Austin. Downloaded on November 25,2020 at 00:16:51 UTC from IEEE Xplore. Restrictions apply.

MIMO radar


MmWave JCR system challenges

37Introduction

BeamformingWaveform Experimental evaluation

PreambleMmWave WLAN

or 5G cellular standard

T

t

f

f0

f0 + B Proprietary FMCW waveform

Exploit antennas for more flexible beamforming design

Data

Radar

Communication

Use high data rate waveform for radar detection/estimation

LRRSRR

V2V

Radar

Communication

*LRR: Long-range radar *SRR: Short-range radar *V2V: Vehicle-to-vehicle communication

Validate theory in implementation

10 MSps with 12-bit ADC radar

3GSps with 12-bit ADC

communication

Radar

Communication


IEEE 802.11ad WLAN-based JCR

38

IEEE 802.11ad is a mmWave WLAN standard at 60 GHz, which provides a good baseline for designing a future vehicular standard at higher mmWave frequencies[1]

Single-carrier physical layer IEEE 802.11ad frame structure

Data block Pilot

Preamble BLK…H OptionalSubfields

BLKP P

Header

Preamble consists of Golay complementary sequences that has good auto-correlation properties for radar

As a next step, advanced adaptive mmWave JCR algorithms were developed to achieve high radar performance, at the cost of minimal data rate reduction

s

TX antenna array

RX antenna array

Direction of cruise

V2V

101110100010

Radar Echo

RX antenna array

DestinationSource

[1] P. Kumari, J. Choi, N. G. Prelcic, and R. W. Heath, “IEEE 802.11ad-based radar: An approach to joint vehicular communication-radar system,” in IEEE TVT, 2018.[2] P. Kumari, N. Gonzalez-Prelcic, and R. W. Heath Jr, “Investigating the IEEE 802.11ad Standard for Millimeter Wave Automotive Radar,” in Proc. VTC, 2015.

Introduction


JCR70 MIMO mmWave JCR hardware platform

39

JCR70 full-duplex RX

JCR70 TX

Motor

SIMO hardware platform with sliding motor Experimental setup for JCR measurements

JCR70 Chassis

Our fully-digital wideband MIMO JCR at 73 GHz with a software-defined architecture

Corner reflector

CommRX

Slider with total length 20 cm

JCR testbed

Wall

INRASRadarbook


IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. XX, NO. XX, XXX 2020 10

-50 0 50

Azimuth (deg)

0

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0

(a) JCR70 traditional processing

-50 0 50

Azimuth (deg)

0

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0

(b) Radarbook traditional processing

-50 0 50

Azimuth (deg)

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0

(c) JCR70 GAMP processing

-50 0 50

Azimuth (deg)

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0

(d) Radarbook GAMP processing

Fig. 11. Estimated radar channels for the two-target scenario using corner reflectors of 4.3 inch and 3.2 inch edge length with traditional and advancedprocessing using our testbed (left) as well as the Radarbook (right). The channel estimates in (c/d) with EM-BG-GAMP have reduced sidelobes than (a/b)with traditional processing. The channel estimates in (a/c) using our JCR70 testbed have recovered the two-target channel response better than the channelestimates in (b/d) using the Radarbook.

target experiment using emulations on the data collected fromour testbed. Fig. 10(a) shows the estimated NMSE variationwith different ADC resolutions at -5 dB, using the tradi-tional algorithm, EM-BG-GAMP, and EM-GM-GAMP. Theestimated NMSE decreases marginally till 2 bit ADC. Thegap between the estimated NMSEs of any two consecutiveADC resolution is highest between the 1-bit ADC and 2-bitADC. Fig. 10(b) depicts the estimated NMSE variation withdifferent SNRs for 12-bit, 4-bit, 3-bit, 2-bit, and 1-bit ADCsusing the traditional FFT-based algorithm and EM-GM-GAMPtechnique. The gap between high-resolution ADCs and low-resolution ADCs increases with SNR. The gap between theestimated NMSEs of traditional processing and sparse GAMPtechnique decreases with SNR. From Figs. 10(a) and (b), wesee that 2-bit ADCs perform very closely to the high-resolutionADCs at low and medium SNRs, whereas 3-bit ADCs performvery closely at all considered SNR values.

B. Radar: Two-target scenarios

We performed a two-target experiment with two cornerreflectors of 4.3 inch and 3.2 inch edge length in the indoorlab using our fully-digital SIMO testbed and the Radarbook.The two targets are closely placed in range and angle domain.Figs. 11(a) and (b) shows the estimated radar channels withtraditional processing algorithm using our testbed and theRadarbook, while Figs. 11(c) and (d) show the estimatedradar channels using EM-BG-GAMP algorithm with reducedsidelobes and noise. Fig. 12 shows the experimental set-up

Fig. 12. Experimental set-up to evaluate the radar performance of our testbedfor a two-target scenario using two corner reflectors in the indoor lab.

for the two-target scenario. We use 30 steps to emulate SIMOusing our testbed. In Fig. 11(a), we observe two scatteringcenters corresponding to the two corner reflectors used, un-like the Radarbook due to higher bandwidth and number ofsynthesized antennas. The GAMP processed image in (c) hasrecovered the amplitudes of the two corner reflectors betterthan the one in (d).

Figs. 13(a) and (b) show the estimated radar NMSE vari-ation with different ADC resolutions and SNRs for the two-target experiment using emulations on the data collected fromour testbed. Fig. 13(a) shows the estimated NMSE variationwith different ADC resolutions at -5 dB, using the tradi-tional algorithm, EM-BG-GAMP, and EM-GM-GAMP. Theestimated NMSE decreases marginally till 3 bit ADC. Thegap between the estimated NMSEs of any two consecutiveADC resolution is highest between the 1-bit ADC and 2-bit ADC. The gap between traditional FFT-based processing


-50 0 50

Azimuth (deg)

0

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

0

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4R

ange (

m)

-60

-50

-40

-30

-20

-10

0










0 2 4 6 8

ADC Resolution (bit)

-20

-18

-16

-14

-12

-10

-8

No

rma

lize

d M

SE

TraditionalEM-BG-GAMPEM-GM-GAMP

(a) NMSE versus ADC resolution

-15 -10 -5 0 5

SNR (dB)

-25

-20

-15

-10

-5

0

No

rma

lize

d M

SE

12-bit4-bit3-bit2-bit1-bit

Traditional

EM-GM-GAMP

(b) NMSE versus SNR

Fig. 13. Estimated radar NMSE for the two-target scenario for different ADCresolutions at SNR equal to -5 dB (left) and versus SNRs for 12-, 4-, 3-, 2-, aswell as 1- bit ADC resolutions (right). The radar channel estimate with 2-bitADCs perform closely to the high-resolution ADCs at low SNRs, whereaswith 3-bit ADCs perform closely at all considered SNR values.

and advanced GAMP algorithms is smaller as compared tosingle-target scenarios. Fig. 13(b) depicts the estimated NMSEvariation with different SNRs for 12-bit, 4-bit, 3-bit, 2-bit,and 1-bit ADCs using traditional FFT-based algorithm andEM-GM-GAMP technique. The gap between high-resolutionADCs and low-resolution ADCs increases with SNR. The gapbetween the estimated NMSEs of traditional processing andsparse GAMP technique decreases with SNR for all consideredADC resolutions except 1-bit ADC, where it first decreasesand then increases with SNR. From Figs. 13(a) and (b), wesee that 2-bit ADCs perform very closely to the high-resolutionADCs at low SNRs, and 3-bit ADCs perform very closely atall considered SNR values.

C. Radar: Extended target scenarios

We used a bike in the indoor lab to evaluate the radarperformance in the extended target scenario using our fully-digital SIMO testbed and the Radarbook. Figs. 14(a) and (b)shows the estimated radar channels with traditional process-ing algorithm using our testbed and the Radarbook, whileFigs. 14(c) and (d) show the estimated radar channels usingEM-BG-GAMP algorithm with reduced sidelobes and noise.Fig. 15 shows the experimental set-up for the extended targetscenario using a bike. We use 50 steps to emulate SIMO usingour testbed. In Fig. 14(c), we observe that multiple scatteringcenters corresponding to different parts of the bike. Due tohigher bandwidth and number of synthesized antennas, the

resolution of the bike image using our testbed is much higherthan the INRAS Radarbook.

Figs. 16(a) and (b) show the estimated radar NMSE vari-ation with different ADC resolutions and SNRs for the bikeexperiment using emulations on the data collected from ourtestbed. Fig. 16(a) shows the estimated NMSE variation withdifferent ADC resolutions at -5 dB, using the traditionalalgorithm, EM-BG-GAMP, and EM-GM-GAMP. The esti-mated NMSE decreases marginally till 4 bit ADC. The gapbetween the estimated NMSEs of any two consecutive ADCresolution is highest between the 1-bit ADC and 2-bit ADC.The gap between traditional processing and advanced GAMPalgorithms is smaller as compared to single- and two-targetscenarios. Fig. 16(b) depicts the estimated NMSE variationwith different SNRs for 12-bit, 4-bit, 3-bit, 2-bit, and 1-bitADCs using traditional FFT-based algorithm and EM-GM-GAMP technique. The gap between high-resolution ADCs andlow-resolution ADCs increases with SNR. The gap betweenthe estimated NMSEs of traditional processing and sparseGAMP technique for high-resolution ADCs decreases withSNR, whereas it first decreases and then increases with SNRfor the low-resolution ADCs. From Figs. 16(a) and (b), we seethat 2-bit ADCs perform very closely to the high-resolutionADCs at low SNR, 3-bit ADCs perform very closely for themedium SNR, and 4-bit ADCs perform very closely at allconsidered SNR values.

Additionally, we also conduct outdoor joint communicationand radar experiments with 50 steps using Subaru Crosstrekas the vehicle target on Speedway parking garage in UTAustin, as shown in Fig. 17. From Fig. 17 and Fig. 18(b), wesee that the JCR transmitter and the communication receiverwere separated by 7.51 m. The vehicle target was placed inbetween the JCR transmitter and the communication receiver.The directivity of the horn antenna reduced the reflectionsfrom the railings and poles around our set-up. Fig. 18(a) showsthe estimated radar channel around the car reflections using thetraditional FFT-based algorithm with an interpolation of factor4, while Fig. 18(b) shows the estimated radar channel using theadvanced EM-BG-GAMP technique. We see that Fig. 18(b)has reduced sidelobes and noise as compared to Fig. 18(a).The range-spread of the car is much wider in range than thatof the bike in Fig. 14. Due to the high-resolution of our testbed,we can also see multiple scattering centers corresponding tothe car which makes the car radar image look quite differentthan the bike.

D. Joint communication-radar

The performance of our fully-digital SIMO widebandtestbed is also evaluated for the simultaneous communicationand radar modes at 73 GHz. We conducted JCR experimentswith 15 steps in the same indoor lab as in Fig. 8.

Fig. 19(a) shows the estimated indoor radar channel in therange-azimuth domain using the traditional processing, whileFig. 19(b) shows the estimated communication channel in theoutdoor setting. In the estimated radar channel image, weobserve the full-duplex effect unlike the estimated commu-nication channel image. The delay-spread in the radar channel

Algorithms and experiments

40

JCR70 Platform INRAS Radarbook NMSE Vs. ADC resolutions

Good channel amplitude recovery with 30 steps and 2 GHz BW

Poor channel amplitude recovery with 29 virtual elements and 1 GHz BW

3-bit ADC perform very closely to the high-resolution ADCs


-50 0 50

Azimuth (deg)

0

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

0

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4

Ra

ng

e (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4R

an

ge

(m

)-60

-50

-40

-30

-20

-10

0










-50 0 50

Azimuth (deg)

0

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

0

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0


-50 0 50

Azimuth (deg)

1

2

3

4

Range (

m)

-60

-50

-40

-30

-20

-10

0










0 2 4 6 8

ADC Resolution (bit)

-20

-18

-16

-14

-12

-10

-8

Norm

aliz

ed M

SE

TraditionalEM-BG-GAMPEM-GM-GAMP

(a) NMSE versus ADC resolution

-15 -10 -5 0 5

SNR (dB)

-25

-20

-15

-10

-5

0

Norm

aliz

ed M

SE

12-bit4-bit3-bit2-bit1-bit

Traditional

EM-GM-GAMP

(b) NMSE versus SNR

Fig. 13. Estimated radar NMSE for the two-target scenario for different ADCresolutions at SNR equal to -5 dB (left) and versus SNRs for 12-, 4-, 3-, 2-, aswell as 1- bit ADC resolutions (right). The radar channel estimate with 2-bitADCs perform closely to the high-resolution ADCs at low SNRs, whereaswith 3-bit ADCs perform closely at all considered SNR values.

and advanced GAMP algorithms is smaller as compared tosingle-target scenarios. Fig. 13(b) depicts the estimated NMSEvariation with different SNRs for 12-bit, 4-bit, 3-bit, 2-bit,and 1-bit ADCs using traditional FFT-based algorithm andEM-GM-GAMP technique. The gap between high-resolutionADCs and low-resolution ADCs increases with SNR. The gapbetween the estimated NMSEs of traditional processing andsparse GAMP technique decreases with SNR for all consideredADC resolutions except 1-bit ADC, where it first decreasesand then increases with SNR. From Figs. 13(a) and (b), wesee that 2-bit ADCs perform very closely to the high-resolutionADCs at low SNRs, and 3-bit ADCs perform very closely atall considered SNR values.

C. Radar: Extended target scenarios

We used a bike in the indoor lab to evaluate the radarperformance in the extended target scenario using our fully-digital SIMO testbed and the Radarbook. Figs. 14(a) and (b)shows the estimated radar channels with traditional process-ing algorithm using our testbed and the Radarbook, whileFigs. 14(c) and (d) show the estimated radar channels usingEM-BG-GAMP algorithm with reduced sidelobes and noise.Fig. 15 shows the experimental set-up for the extended targetscenario using a bike. We use 50 steps to emulate SIMO usingour testbed. In Fig. 14(c), we observe that multiple scatteringcenters corresponding to different parts of the bike. Due tohigher bandwidth and number of synthesized antennas, the

resolution of the bike image using our testbed is much higherthan the INRAS Radarbook.

Figs. 16(a) and (b) show the estimated radar NMSE vari-ation with different ADC resolutions and SNRs for the bikeexperiment using emulations on the data collected from ourtestbed. Fig. 16(a) shows the estimated NMSE variation withdifferent ADC resolutions at -5 dB, using the traditionalalgorithm, EM-BG-GAMP, and EM-GM-GAMP. The esti-mated NMSE decreases marginally till 4 bit ADC. The gapbetween the estimated NMSEs of any two consecutive ADCresolution is highest between the 1-bit ADC and 2-bit ADC.The gap between traditional processing and advanced GAMPalgorithms is smaller as compared to single- and two-targetscenarios. Fig. 16(b) depicts the estimated NMSE variationwith different SNRs for 12-bit, 4-bit, 3-bit, 2-bit, and 1-bitADCs using traditional FFT-based algorithm and EM-GM-GAMP technique. The gap between high-resolution ADCs andlow-resolution ADCs increases with SNR. The gap betweenthe estimated NMSEs of traditional processing and sparseGAMP technique for high-resolution ADCs decreases withSNR, whereas it first decreases and then increases with SNRfor the low-resolution ADCs. From Figs. 16(a) and (b), we seethat 2-bit ADCs perform very closely to the high-resolutionADCs at low SNR, 3-bit ADCs perform very closely for themedium SNR, and 4-bit ADCs perform very closely at allconsidered SNR values.

Additionally, we also conduct outdoor joint communicationand radar experiments with 50 steps using Subaru Crosstrekas the vehicle target on Speedway parking garage in UTAustin, as shown in Fig. 17. From Fig. 17 and Fig. 18(b), wesee that the JCR transmitter and the communication receiverwere separated by 7.51 m. The vehicle target was placed inbetween the JCR transmitter and the communication receiver.The directivity of the horn antenna reduced the reflectionsfrom the railings and poles around our set-up. Fig. 18(a) showsthe estimated radar channel around the car reflections using thetraditional FFT-based algorithm with an interpolation of factor4, while Fig. 18(b) shows the estimated radar channel using theadvanced EM-BG-GAMP technique. We see that Fig. 18(b)has reduced sidelobes and noise as compared to Fig. 18(a).The range-spread of the car is much wider in range than thatof the bike in Fig. 14. Due to the high-resolution of our testbed,we can also see multiple scattering centers corresponding tothe car which makes the car radar image look quite differentthan the bike.

D. Joint communication-radar

The performance of our fully-digital SIMO widebandtestbed is also evaluated for the simultaneous communicationand radar modes at 73 GHz. We conducted JCR experimentswith 15 steps in the same indoor lab as in Fig. 8.

Fig. 19(a) shows the estimated indoor radar channel in therange-azimuth domain using the traditional processing, whileFig. 19(b) shows the estimated communication channel in theoutdoor setting. In the estimated radar channel image, weobserve the full-duplex effect unlike the estimated commu-nication channel image. The delay-spread in the radar channel

(a) Traditional

(b) EM-BG-GAMP

(a) Traditional

(b) EM-BG-GAMP

3.2″ target

4.3″ target


41

Questions


Can mmWave really work with high mobility?

42

Beams are too hard to configure

Doppler is too high

Nothing is known about the channel

Many misconceptions surround mmWave communications in mobile

channels

Ongoing research to establish and maintain mmWave links in vehicular scenarios


Channel coherence time and directional reception

43

Long term beamforming can be used

Beam training not Doppler

Beams should be narrow but not too “pointy”

*V. Va, J. Choi, and R. W. Heath Jr. The impact of beamwidth on temporal channel variation in vehicular channels and its implications, IEEE TVT, vol. 7, no. 5, 2018

Optimum beamwidthis a tradeoff between

pointing error and Doppler

Mathematical expression relating coherence time and

beamwidth

millimeter waves and vehicle to vehicle and infrastructure ......ch15-p372533.tex 14/5/2007 18: 10...

Documents