Content uploaded by Yuyang Wang
Author content
All content in this area was uploaded by Yuyang Wang on May 26, 2019
Content may be subject to copyright.
MmWave Vehicular Beam Training with Situational
Awareness by Machine Learning
(Invited Paper)
Yuyang Wang‡
,Aldebaro Klautau∗
,Monica Ribero‡
,Murali Narasimha§
,and Robert W. Heath Jr.‡
‡Department of Electrical and Computer Engineering, The University of Texas at Austin, USA
∗Federal University of Para (UFPA), Belem, Brazil §Huawei Technologies, Rolling Meadows, USA
Email: {yuywang, rheath, mribero}@utexas.edu, aldebaro@ufpa.br, murali.narasimha@huawei.com
Abstract—Configuring beams in millimeter-wave (mmWave)
vehicular communication is a challenging task. Large antenna
arrays and narrow beams are deployed at transceivers to exploit
beamforming gain, which leads to significant system overhead if
an exhaustive beam search is adopted. In this paper, we propose
to learn the optimal beam pair index by exploiting the locations
and sizes of the receiver and its neighboring vehicles (parts
of the situational awareness for automated driving), leveraging
machine learning tools with the past beam training records.
MmWave beam selection is formulated as a classification problem
based on situational awareness. We provide a comprehensive
comparison of different classification models and various levels of
situational awareness. Practical issues are considered in realistic
implementations, including GPS inaccuracies, out-dated locations
due to fixed location reporting frequencies and missing features
with limited connected vehicles penetration rate. The result shows
that we can achieve up to 86%of alignment probability with ideal
assumptions.
I. INTRODUCTION
With the huge frequency resources available, mmWave
is the only viable approach to support the massive sensor
data sharing in emerging vehicular applications [1]. MmWave
communication employs large antenna arrays to compensate
for the small antenna aperture and the penetration loss in case
of blockage [2]. With the narrow beams used in mmWave
with analog beamforming and high mobility in the vehicular
context, current beam training solutions based on exhaustive
search are inapplicable in most vehicular applications due to
the overhead and latency.
Wireless cellular communication systems have access to
vast data resources – yet untapped – which can make beam
training more efficient. At every base station (BS), sector,
and radio frequency channel, hundreds of pilot signals are
sent between the base station and users in its coverage area
to measure the propagation channel every second. With the
same cadence, the users send feedback to the BS about the
measured channel, for the infrastructure to select transmis-
sion parameters accordingly. Remarkably, this information is
leveraged only over a fraction of a second and then discarded.
This “data collection - discard” mode in the wireless industry
may be explained at lower frequencies due the presence of
more propagation paths in the channel, the variability in phone
handset movement, and limitations in sensing capabilities of
phones.
Data-driven beam training is the right approach for
mmWave vehicular systems. Wireless systems are becoming
too complicated to be fit in stochastic mathematical models.
MmWave channels are more deterministic given environment
geometries. Hence, beam information can be learned from the
geometry of surrounding objects, since they serve as reflectors
and can change directions of the beams. Also, vehicles travel
along roads in predictable ways, following certain patterns
e.g., higher density or slower speed during rush hour, which
makes it easier for mobility modeling. Vehicles are also
being equipped with sensors like satellite navigation, radar,
LiDAR, and cameras which can be used to localize vehicles
and sense the environment [1], [3]. With the availability of
sensors, vehicle locations can be shared to make everything
connected. Lastly, the BS is well equipped with the capability
to accommodate the needs for data-driven solutions. It has
access to cloud and edge computing capability; it is the natural
data fusion hub for data generated by various types of sensors
on a vehicle; it can also monitor current transmission status
and has capacity to cache the useful transmission records [4],
[5].
There are several papers proposing alternative data-driven
mmWave vehicular beam training solutions with side in-
formation, e.g., radar, cameras, etc [3]. In [6], the paper
proposed to recommend mmWave vehicular beams based on
the receiver location, with past transmission histories. The
proposed solution [6] might work well in line-of-sight (LOS)
dominant scenarios. It is very likely to fail, however, when
abundant blockages are present in the dataset, since the beams
could be highly random and large numbers of beams will
be recommended in order to guarantee alignment probability.
In [7], the paper presented a 5G-MIMO dataset of vehicular
channel and beam statistics with temporally-correlated vehicle
trajectories. An initial investigation was provided to solve the
mmWave beam selection problem when all the vehicle ge-
ometries instead of only the receiver location are incorporated
in the feature. In addition to beam pair index prediction, it
was shown in [8] that vehicles’ situational awareness can also
be used to predict the beam power to reduce overheads and
automate vehicular beam training.
In this paper, we propose a comprehensive investigation
of an efficient machine learning framework with situational
awareness in Section II. Aside from using the receiver lo-
cation, we propose to locations and sizes of vehicles in
the environment to predict the optimal beam pair index. In
Section II-C and II-D, we apply three different approaches, in
Cartesian coordinates, polar coordinates and occupancy grids,
with appropriate feature ordering. In Section IV, we compare
different classification methods to predict the optimal beam
pair index based on the situational features. Furthermore, we
show the significant improvement of prediction accuracy when
a richer set of vehicles locations are included in the prediction.
Since the model relies on the vehicle locations, location errors
in realistic implementations are considered, including GPS
inaccuracies, vehicle location reporting frequency and different
penetration rates of connected vehicles.
II. BEAM SELECTION USING MACHINE LEARNING
A. Problem formulation
In mmWave vehicle-to-infrastructure (V2I) communica-
tions, road-side units (RSU) are deployed at a low height
(generally collocated with the lamp) on the road side, to
support data showering for the passing vehicles. Multiple
reflections might happen on the objects in the environment,
e.g., the road-side buildings, vehicles, pedestrians, etc, before
the beams reach the receiver. Among the different types
of reflecting objects, buildings (or roads) are stationary and
pedestrians are small in size which have negligible impacts
on the channel. This makes the vehicles the most important
dynamic factors in affecting the channel. We divide the set of
environment information Eto several categories: 1) stationary
objects in the environment as S(e.g., buildings, roads, other
infrastructures), 2) mobile vehicles as V, and 3) other objects
small in size (e.g., pedestrians, foliage) as N. There exists
a function f(·)that can map the set of the full environment
information Eto the beam information B, with some negligible
error η, i.e.,
∃f(·), s.t., f(E) = f(S,V,N) = B+η, (1)
where ηis due to some random disturbances in the model and
is very small.
Assuming the effect of Non the beam is negligible, and
Sis identical as S=S0for different data samples for the
current urban canyon, (1) can be further simplified to
∃f(·), s.t., f (V,S=S0,N)≈fS0(V)≈ B.(2)
Hence, we propose to learn the optimal beam pair index by
leveraging all vehicles’ locations and sizes. Specifically, the
beam selection will be formulated as a classification problem.
In Section II-B - II-D, we will respectively explain the details
of channel modeling, optimal beam index calculation, and how
to encode vehicle geometries with appropriate feature format.
B. Data collection and channel model
Since there are no testbeds available for mmWave V2I
communications, we utilize a ray tracing simulator to establish
the dataset. Ray tracing simulation has been widely adopted in
both industry and academia for channel measurements [9], [10]
and modeling [11]. We use a commercial ray tracing simulator
- Wireless Insite, from Remcom [12]. We consider a two-lane
straight street in urban canyon as shown in Fig. 1. Buildings
modeled by cubes are located on the two road sides with
random sizes. We consider two types of vehicles, respectively
large trucks, with length, width, height =T`, Tw, Thand
low height cars (sedans) of size C`, Cw, Ch. All vehicles
are modeled as cubes with material exteriors and randomly
dropped on the two lanes based on some certain distribution.
Ray tracing simulation outputs the Lstrongest paths from
the given configurations of geometry, transceiver, etc. Since
the path directions span a three-dimensional space, the output
angles are formatted in a spherical coordinate. In particular,
each path includes: 1) arrival azimuth angle φA, 2) arrival ele-
vation angle θA, 3) departure azimuth angle φD, 4) departure
elevation angle θD, 5) path gain αejφ with φas the phase, and
6) time of arrival τ. Due to the sparsity of mmWave channel,
we assume that parameters above are enough to approximate
the channels. Let’s denote the number of transmit antenna as
Ntand that of receive antenna as Nr. The wide-band channel
H[n]with Lctaps can be calculated as below [6]
H[n] =pNtNr
L
X
`=1
g(nT −τ`)ar(φA
`)a∗
t(φD
`)α`ejφ`,
0≤n≤Lc−1,(3)
where g(·)is the pulse shaping factor, at(·)and ar(·)are
the steering vectors, and Tis the sampling period. Uniform
planar arrays (UPA) are applied at both the transmitter and the
receiver. We apply DFT codebook Wand Ffor the precoder
and combiner and assume Nt=Nr=Ny×Nx= 4 ×2=8.
Therefore, there is a total of NB=NtNr= 64 different
beam pairs in our dataset. The received power of all the NB
beam pairs in the k-th sample is represented as a vector yk.
Assuming the i-th beam pair, (wi,fi), i ∈ {1,2,· · · , NB}, is
selected from the codebook, the received power of the i-th
beam pair yk[i]can be calculated as
yk[i] =
Lc−1
X
n=0
w∗
iH[n]fi
2.(4)
and the optimal beam index skcan be obtained by
sk= arg max
i∈{1,2,··· ,|W×F |}{yk[i]},(5)
Fig. 1. Illustration of the deployment of ray tracing simulation. Receivers are
mounted on top of the low vehicles (denoted by the yellow boxes) and the
RSU is denoted by the green box.
Fig. 2. Illustration for encoding the environment with a Cartesian coordinate.
The origins lies at the left boundary of the simulation road side at the bottom.
And the each feature of the vehicle location is encoded as [horizontal location,
lane index].
Fig. 3. Illustration for encoding the environment with a polar coordinate.
The origins lies at RSU. Vehicle location is encoded as (r, φ), where φis the
angle between the point and the x-axis.
where |·| denotes the cardinality of the set.
C. Encoding vehicle geometry
To properly encode the locations of multiple vehicles,
we propose to format the features in Cartesian, polar, and
occupancy grids.
1) Cartesian coordinate: The first approach is by represent-
ing the simulation area in a Cartesian coordinate, as shown in
Fig. 2. We are considering a two-lane street and the cars are
assumed to be traveling in the center of each lane. Categorical
variables 1 or 2 are used to denote the lane index. Each
vehicle’s location is represented by (horizontal location, lane
index).
2) Polar coordinate: Cartesian coordinate gives a way to
encode the environment in a lossless way. It includes the
accurate location of every vehicle in the environment. Another
lossless option is to use polar coordinate to encode. Beam
training is in essence a problem to find the beam angles. In the
LOS links, polar coordinates give direct ways of interpreting
angles of the paths. To capture this angle correlation, the origin
is set at the RSU, as shown in Figure 3. And each location is
given by (distance to the origin, angular coordinate).
3) Occupancy grid: Another approach is to encode the
environment in either two-dimensional or three-dimensional
occupancy grids. This approach, though encoding the data in
a lossy way due to the quantization, can be more flexible in
including a wide category of urban canyons with different sizes
of vehicles and could be more resilient to potential feature
noises. The idea is to quantize the locations in a regular-
sized grid with a certain granularity dg, which is quantified
Fig. 4. Illustration for encoding the environment in regular-sized grids. The
origin lies at the left bottom of the simulation area. The x-axis direction is
quantized to different grids to encode the data.
by the distance between two adjacent grids. The size of the
simulation area horizontally is Dh. In our model, there are
in total of 3types of target vehicles, respectively neighboring
trucks,neighboring cars, and the receiver car, represented as
1, 2, 3. An illustration is provided in Fig. 4. The grids on the
first and second lanes will be represented by two vectors g1
and g2of length Dh
dg. If nothing lies inside the i-th grid of
the j-th lane (j= 1,2), we indicate the value gj[i] = 0; if
part of the truck lies inside the grid, gj[i] = 1; if part of a
car (excluding the receiver) lies inside the grid, gj[i]=2; if
the receiver lies inside the grid, gj[i] = 3. Finally, the vehicle
geometry vector in the occupancy grid gis concatenated by
the two vectors on the first and second lanes into one vector
g= [g1,g2].
D. Ordering the location features
Now each vehicle is encoded in a certain format by the
three coordinates discussed in Section II-C, properly ordering
the multiple vehicles in the feature array is still required in
Cartesian and polar coordinates. Similarly as Section III-A in
[8], vehicles are ordered based on the relative distance of the
vehicle to the receiver, the type of vehicle, the lane the vehicle
is located on, etc. In brief, the feature vin each of the data
sample is a one-dimensional vector and can be represented by
v= [r,t1,t2,c1,c2].(6)
In (6), ris the location of the RSU (in the Cartesian coordi-
nate) or the location of the receiver (in the polar coordinate).
The vector tand crepresent the set of locations of trucks
and the low cars, where the subscripts 1and 2denote the lane
index. In each t1,t2,c1and c2, the vehicles are ordered based
on their relative distance to the receiver, where the details are
referred to [8].
III. PRACTICAL IMPLEMENTATIONS
The proposed beam training approach relies on a “fully
connected vehicles” assumption, i.e., we assume all vehicles
are equipped with communicating transceivers, corresponding
protocols and high-resolution GPS, which enable them to
report real-time GPS locations to infrastructures. Each time
when a vehicle initializes a beam training request, the infras-
tructure will be able to construct a real-time map of all vehicles
in the site and formulate the corresponding feature to predict
the beam based on the trained model after offline learning.
In reality, however, a lot of factors could lead to inaccurate
or missing vehicle locations in the feature. In particular,
we exemplify three important sources for noisy features as
follows.
A. GPS inaccuracy
GPS is a satellite-based radio navigation system, where
GPS satellites broadcast their signals in space with a cer-
tain accuracy. Generally, the accuracy depends on a lot of
additional factors, e.g., satellite geometry, signal blockage,
atmospheric conditions, etc [13]. Particularly in the vehicular
urban canyon with frequent signal blockages and reflections,
the GPS inaccuracy can be a big issue.
Since the vehicles are traveling along the center of the lanes
and lane index is simply a categorical variable, the lane index
is assumed to correct all the time. Hence, the GPS location
error is modeled by a Gaussian distribution with zero mean,
and variance of σ2. Consistently, the GPS inaccuracies will be
modeled in both the training datasets and the testing datasets,
affecting all vehicles in the features.
B. Location updating frequency
In addition to the inherent inaccuracy with the vehicle
GPS, the randomness of vehicle mobility can introduce extra
errors to the vehicle locations. Generally, vehicles report
their locations to the infrastructures every ∆ttime. With
mobility, however, the locations of vehicles stored in the
infrastructure soon become out-dated. Vehicles hence have
to report locations to the infrastructures frequently, so as to
guarantee a low level of vehicle location errors. Particularly,
given some initial setting of velocity v(t)and a(t)at t, in
the time interval [t, t + ∆t], the maximum location error is
¯x=v(t)∆t+1
2a(t)∆2
t. For simplicity, we assume the same
velocities and accelerations for all vehicles in the simulation.
The location error ¯xis applied to both training and testing
datasets.
One alternative is to pack more information in the ve-
hicles’ report to infrastructure. Connected vehicles are built
around the SAE protocol J2735 basic safety message (BSM),
which incorporates information of vehicle locations, and other
mobility related information, such as velocities and acceler-
ations [14]. With full BSM feedback, vehicle locations can
be predicted based on simple equations of motion, given the
previous state of the mobility parameters. For example, given
the vehicle’s location, velocity and acceleration at time of tas
x(t), v(t), a(t), at any time during ˜
t∈[t, t + ∆t], the updated
location can be approximated by
x(˜
t)≈x(t) + v(t)(˜
t−t) + 1
2a(t)(˜
t−t)2.(7)
As long as the reporting interval ∆tis small enough, during
which the acceleration remains stable, (7) will keep track of the
vehicle’s location precisely during the time interval [t, t +∆t].
Hence, the key is simply to capture the appropriate time
interval within which the acceleration remains. The required
reporting interval is now converted to the scale of transporta-
tion, which is several orders of the time scale of that in
communication. Also, it should be noted that exhaustive beam
search, in most of the times, cannot capture beam changes on
time, since the beam training only happens at fixed time slots
t=τB,2τB,· · · (τBis the interval of conducting beam search
in traditional mmWave systems). Our model will be flexible
enough to predict any potential changes of beams, since now
we have continuous interpretation of vehicle locations and
tracking of the beams based on the locations.
C. Penetration rate
Another important issue in practical implementations is
the non-ideal penetration rate of the connected vehicles.
The penetration rate defines the percentage of the vehicles
with connected devices, whose locations are known to the
infrastructures. Fully connected vehicles are not a realistic
assumption in the current state of art. Low or not 100%
penetration rate might have significant impacts on our model
since the features have to be properly selected and ordered
correspondingly based on the locations as shown in Section
II-C and II-D. In this case, the vehicle locations might be
missing and the features might be misaligned. In particular, it
is assumed that only the low cars connected will initiate beam
training request within the proposed framework. Therefore,
among the data samples, we randomly select trucks as being
connected based on Bernoulli distribution with probability ppr,
where ppr is the penetration rate pre-defined. Trucks who are
not connected will be invisible to the infrastructure and the
locations will be missing in the feature.
IV. PER FO RM AN CE EVAL UATION
In this section, we provide a comprehensive evaluation
of the proposed beam selection approach. We mainly focus
on the two performance metrics: alignment probability and
the achieved throughput. Let max{·} denote the maximum
element of the vector (·), and (·)is the indicator function.
Given the real power y1,· · · ,ym, the real optimal beam
pair index s1, s2,· · · , smand the predicted beam pair index
ˆs1,ˆs2,· · · ,ˆsmof the msamples, the alignment probability
can be formulated as
PA=1
m
m
X
i=1
(ˆsi=si),(8)
and the achieved throughput ratio RTis
RT=Pm
i=1 log2(1 + yi[ˆsi])
Pm
i=1 log2(1 + max{yi}),(9)
We start with comparing the beam selection performance
with different machine learning models. Then we evaluate
the prediction performance with different levels of situational
awareness by adding in different numbers of vehicle in the
feature. Lastly, some realistic issues of noisy features in
implementations are discussed as in Section III.
A. Learning models
For fair comparison, we start with applying a simple feature
set encoded in a Cartesian coordinate as in Section II-C, with
TABLE I
ALIGNMENT PROBABILITY AND ACHIEVED THROUGHPUT RATIO OF BEAM
SELECTION WITH DIFFERENT CLASSIFIERS.
PA(%)RT(%)
Naive-Bayes 59.31 91.14
AdaBoost 45.80 75.05
RBF-SVM 55.89 89.32
Gradient Boosting 69.05 96.49
Random forest 85.14 98.32
no GPS errors. The used classifiers are listed in Table I. From
the result, it is shown that random forest outperforms the
other classifiers, and it achieves 85.14%alignment probability.
Another important observation is that the achieved through-
put does not scale with the alignment probability. Despite
the fact that random forest achieves much higher alignment
probability compared to the classifiers such as Naive-Bayes or
gradient-boosting, there are no significant differences among
their achieved throughput. The reason is the strongest several
beams’ powers are close. Therefore, even if the classifier
cannot predict the exact optimal beam pair index, as long
as the selected one is among one of those top beams, the
throughput will not be too bad. The conclusion reveals the fact
that alignment probability might not be the best performance
metric for evaluation. Even though the model cannot guar-
antee 100% alignment probability, it could still be accurate
enough to identify those “good beams”. This is important
for mmWave vehicular system design, since it is beneficial
to sacrifice some optimality to achieve low overheads. Some
sub-optimality is introduced in the prediction accuracy, but
the overheads can be largely reduced which can still provide
significant improvement over the system performance.
B. Different levels of situational awareness
In this part, we evaluate and compare the system per-
formance with different levels of situational awareness. The
red curve in Fig. 5 plots the alignment probability with
different number of vehicles modeled in the feature. The
vehicle locations are ordered correspondingly based on relative
distance to the receiver as in Section II-D. When the feature
only includes the receiver location, an alignment probability
of around 62%. is achieved. After adding in first lane truck
locations, the alignment probability is improved to 82%. From
the curve, it can be concluded that most of the enhancement is
contributed by the closest trucks’ (around only 2-3) locations.
Similar observations can be obtained by adding in second
lane trucks. When including low vehicles into the features,
the alignment probability remains stable which implies that
low vehicles have negligible impacts in mmWave vehicular
channels and are not very useful in predicting the beams.
Furthermore, we reduce the feature set by only keeping the
locations of vehicles that improve the alignment probability on
the red curve. Specifically, the reduced feature set becomes:
[receiver location, closest three trucks on the first lane, closest
0 5 10 15 20 25 30 35 40 45
Number of Vehicle Locations in the Feature
0.6
0.65
0.7
0.75
0.8
0.85
0.9
Alignment Probability
Increasing feature
Reduce feature
Adding low-height carsAdding 2nd lane truck
Adding 1st lane trucks
(Nveh = 10)
Only RX location
Fig. 5. Alignment probability adding in different number of vehicle locations
in the feature. We start with only using the receiver location of the feature and
add on the first lane truck, second lane truck, first lane cars, second lane cars
correspondingly based on the vehicle order in Section II-D. The blue curve
is the alignment probability by only including part of the vehicle locations in
the feature, e.g., the closest several trucks and cars on each lane.
012345678910
Variance of Gaussian error
0.55
0.6
0.65
0.7
0.75
0.8
0.85
Alignment probability
0.89
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
Throughput
Cartesian
Polar
Grids
Fig. 6. Comparison of alignment probability and achieved throughput with
different distributions of GPS location error. Results of Cartesian coordinates,
polar coordinates and occupation grid are compared in the figure.
truck on the second lane]. Blue curve shows the alignment
probability of around 86%when this reduced feature set is
used.
C. Noisy features
In this section, noisy location features are included in the
model. The inaccuracies of the location can come from three
parts: 1) inherent GPS inaccuracy, 2) out-dated locations due
to vehicle mobility and fixed location reporting interval, 3)
missing features due to limited penetration rate.
Alignment probability and achieved throughput are com-
pared in Fig. 6, with different GPS inaccuracies and encodings
of vehicle geometries. It is shown that location inaccuracy can
make a big difference on the prediction accuracy, even when
GPS error variance is very small. When the GPS error grows
large, the prediction accuracy falls below 60%. Therefore, one
limitation of our model is that it heavily relies on exact feed-
back of vehicle locations provided by GPS, whose accuracy
cannot be guaranteed, at least in the state-of-art. Also, it can
0 0.2 0.4 0.6 0.8 1
Penetration rate
0.6
0.65
0.7
0.75
0.8
0.85
0.9
Alignment probability
0 penetration rate (no situational awareness)
Full penetration rate
(full situational
awareness)
Fig. 7. Comparison of alignment probability with different penetration rates
with Cartesian coordinates.
0 100 200 300 400 500 600 700 800 900 1000
Reporting interval (ms)
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Alignment probability
a = 0, 2, 5, 10 m/s2
v = 30km/h
v = 60km/h
v = 120km/h
Fig. 8. Comparison of alignment probability with different vehicle velocities,
accelerations and location reporting intervals. Here we assume identical
uniformly accelerated motion for all vehicles and there are no GPS errors
modeled in either the training or the testing datasets for fair comparison.
be observed that Cartesian and polar coordinates achieve better
performance than the occupancy grids.
Fig. 8 demonstrates how the performance is affected when
a longer location reporting interval is applied with vehicle
mobility. We assume uniformly accelerated rectilinear motion
for all vehicles. A combination of parameters v= 30,60,120
km/h and a= 0,2,5,10 m/s2is used and a fixed location error
¯x=v∆t+1
2a∆2
tis added to all features. For fair comparison,
no GPS errors are included in the datasets. It is shown
that if only the locations are reported to the infrastructures,
the alignment probability degrades very fast with a longer
reporting interval, even at a relatively low speed. Particularly,
using reporting interval of around 200 ms, the alignment
probability decreases from 86%to around 70%at v= 30km/h.
The different values of acceleration rates, however, do not have
a significant impact on the alignment probability. Therefore, if
the velocities can be fed back to the infrastructure along with
the locations, the updated locations can largely be predicted
and the model’s accuracy can be guaranteed, with negligible
extra overheads of feedback required.
Fig. 7 plots the alignment probability with different pene-
tration rates (of trucks). Since it is assumed that all the low
vehicles report locations to the infrastructure and only trucks
might be disconnected from the network, 0penetration rate
indicates the case without situational awareness, while full
penetration rate is equivalent to the full situational awareness
case as discussed in Section IV-B. It is shown that alignment
probability scales proportionally with the penetration rate and
even at very low penetration rate of around 20%, there is still
prominent improvement of the performance compared to the
no situational awareness case.
V. ACKNOWLEDGMENT
This research was partially supported by a gift from Huawei
through UT Situation-Aware Vehicular Engineering Systems
(UT-SAVES), by the U.S. Department of Transportation
through the Data-Supported Transportation Operations and
Planning (D-STOP) Tier 1 University Transportation Center
and the Communications and Radar-Supported Transportation
Operations and Planning (CAR-STOP) project funded by the
Texas Department of Transportation.
REFERENCES
[1] J. Choi, V. Va, N. Gonzalez-Prelcic, R. Daniels, C. R. Bhat, and R. W.
Heath, “Millimeter-wave vehicular communication to support massive
automotive sensing,” IEEE Commun. Mag., vol. 54, no. 12, pp. 160–
167, 2016.
[2] Y. Wang, K. Venugopal, A. F. Molisch, and R. W. Heath, “Blockage
and coverage analysis with mmwave cross street BSs near urban
intersections,” in Proc. Int. Conf. Commun. (ICC), pp. 1–6, IEEE, 2017.
[3] N. Gonz´
alez-Prelcic, A. Ali, V. Va, and R. W. Heath, “Millimeter-wave
communication with out-of-band information,” IEEE Commun. Mag.,
vol. 55, no. 12, pp. 140–146, 2017.
[4] Y. Wang, K. Venugopal, R. W. Heath, and A. F. Molisch, “Mmwave
vehicle-to-infrastructure communication: Analysis of urban microcellu-
lar networks,” IEEE Trans. on Veh. Technol., pp. 1–1, 2018.
[5] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, “Heterogeneous cloud radio
access networks: A new perspective for enhancing spectral and energy
efficiencies,” IEEE Wireless Commun., vol. 21, no. 6, pp. 126–135, 2014.
[6] V. Va, J. Choi, T. Shimizu, G. Bansal, and R. W. Heath, “Inverse
multipath fingerprinting for millimeter wave V2Ibeam alignment,” IEEE
Trans. Veh. Technol., vol. 67, pp. 4042–4058, May 2018.
[7] A. Klautau, P. Batista, N. G. Prelcic, Y. Wang, and R. W. Heath, “5G
MIMO data for machine learning: Application to beam-selection using
deep learning,” in Proc. Inf. Theory and Appl. Workshop (ITA), pp. 1–6,
Jan. 2016.
[8] Y. Wang, M. Narasimha, and R. W. Heath, Jr, “MmWave Beam
Prediction with Situational Awareness: A Machine Learning Approach,”
ArXiv e-prints, May 2018.
[9] S. Hur, S. Baek, B. Kim, J. Park, A. F. Molisch, K. Haneda, and M. Peter,
“28 GHz channel modeling using 3D ray-tracing in urban environments,”
in Proc. European Conf. Antennas and Propagation (EuCAP), pp. 1–5,
IEEE, 2015.
[10] K. Haneda, L. Tian, H. Asplund, J. Li, Y. Wang, D. Steer, C. Li,
T. Balercia, S. Lee, Y. Kim, et al., “Indoor 5G 3GPP-like channel models
for office and shopping mall environments,” in Proc. Int. Conf. Commun.
Workshops, pp. 694–699, IEEE, 2016.
[11] V. Degli-Esposti, F. Fuschini, E. M. Vitucci, M. Barbiroli, M. Zoli,
L. Tian, X. Yin, D. A. Dupleich, R. M¨
uller, C. Schneider, et al., “Ray-
tracing-based mm-wave beamforming assessment,” IEEE Access, vol. 2,
pp. 1314–1325, 2014.
[12] https://www.remcom.com/wireless-insite-em- propagation-software/.
[13] https://www.gps.gov/.
[14] Y. Zhang, G. W. Gantt, M. J. Rychlinski, R. M. Edwards, J. J.
Correia, and C. E. Wolf, “Connected vehicle diagnostics and prognostics,
concept, and initial practice,” IEEE Trans. on Reliability, vol. 58, no. 2,
pp. 286–294, 2009.