Federated Learning for Reliable mmWave Systems: Vision-Aided Dynamic Blockages Prediction

Full text

Federated Learning for Reliable mmWave Systems:
Vision-Aided Dynamic Blockages Prediction
Mohammad Al-Quraan, Anthony Centeno, Ahmed Zoha, Muhammad Ali Imran, and Lina Mohjazi
James Watt School of Engineering, University of Glasgow, Glasgow, UK
Email: {m.alquraan.1}@research.gla.ac.uk,
{Anthony.Centeno, Ahmed.Zoha, Muhammad.Imran, Lina.Mohjazi}@glasgow.ac.uk
Abstract—Line of sight (LoS) links that use high frequencies
are sensitive to blockages, making it challenging to scale future
ultra-dense networks (UDN) that capitalise on millimetre wave
(mmWave) and potentially terahertz (THz) networks. This paper
embraces two novelties; Firstly, it combines machine learning
(ML) and computer vision (CV) to enhance the reliability and
latency of next-generation wireless networks through proactive
identification of blockage scenarios and triggering proactive
handover (PHO). Secondly, this study adopts federated learning
(FL) to perform decentralised model training so that data privacy
is protected, and channel resources are conserved. Our vision-
aided PHO framework localises users using object detection
and localisation (ODL) algorithm that feeds a multiple-output
neural network (NN) model to predict possible blockages. This
involves analysing images captured from the video cameras co-
located with the base stations (BSs) in conjunction with wireless
parameters to predict future blockages and subsequently trigger
PHO. Simulation results show that our approach performs re-
markably well in highly dynamic multi-user environments where
vehicles move at different speeds, and achieves 93.6% successful
PHO. Furthermore, the proposed framework outperforms the
reactive-HO methods by a factor of 3.3 in terms of latency while
maintaining a high quality of experience (QoE) for the users.
Index Terms—Federated Learning, computer vision, blockage
prediction, ultra-dense networks, network latency.
I. INTRODUCTION
Next-generation wireless networks undergo a substantial
design change when operating in high-frequency bands [1].
Obtaining high data rate services from millimeter wave (mm-
Wave) and terahertz (THz) technologies demands downscal-
ing the communication system, resulting in a new network
paradigm termed ultra-dense networks (UDN) [2]. Moreover,
the use of beamforming enhances the received signal strength
(RSS) by forming line-of-sight (LoS) beams. Nevertheless,
UDNs face critical challenges due to the sensitivity of high-
frequency beams to blockages. These signals suffer high
penetration loss and attenuation, leading to a high RSS drop
each time an obstacle intercepts the LoS communication link.
This problem is aggravated in highly dynamic environments
where many dynamic objects can cause frequent blockages.
In literature, several techniques are adopted to overcome
the connectivity issue. For example, mmWave channel geo-
metry and signal diffraction characteristics have been studied
compared with sub-6GHz to predict whether a mmWave LoS
connection will be blocked [3], [4]. Other solutions rely on
machine learning (ML) and the dual connectivity (DC) to
maintain wireless connectivity and meet the required quality of
experience (QoE) for users [5], [6]. However, such solutions’
limitations vary in practicality, wasting network resources, and
most importantly, they do not solve the link blockage problem
since switching between links is still reactive.
To best solve this problem, UDNs require a sense of the
surrounding environment to move from reactive to proactive
blockage measures. The direct view is essential for UDN
communications and is equally important to computer vision
(CV), where visual information captures only direct visible
objects in the scene, helping to proactively detect blocking
objects. Therefore, leveraging vision information collected
from the served environment is envisioned to aid the operation
of the network rather than relying on wireless information
alone, which fails to address this dilemma [3]–[6]. Images
are rich in detail that can help solve the blockage problem in
UDNs. However, this hinges on two main questions; first, is
it possible to detect objects in the environment and identify
their mobility information? Second, how can wireless users be
distinguished from other passive objects in still images?
In [7], depth images and a deep learning (DL) model
are used to predict a user’s RSS in the next few hundred
milliseconds to assist in handover (HO) decisions. Further,
[8] exploits red-green-blue (RGB) images to train a ResNet-
18 model and then classify the images based on the blockage
status. However, the approaches in [7], [8] do not account
for the associated latency until completing a successful HO
and, therefore, cannot avoid link blockages, which is critical
in highly dynamic UDNs. In our previous work [9] we go
further by providing a simple scenario study that intelligently
detects blockages and performs optimal proactive HO (PHO)
considering the latency required to ensure user HO. The multi-
user environment is considered in [10]; this work adopts CV
and DL to predict whether the beam will be blocked in the
next instance. However, the limited-time prediction will most
likely not avoid beam blockages.
In light of the above discussion, this paper extends our
study in [9] by providing a CV-aided latency-aware PHO
solution that targets practical multi-user UDNs. The proposed
framework utilises the object detection and localisation (ODL)
algorithm in addition to neural network (NN) to accurately
predict blockages and the time when the blockage will occur.
Moreover, we noticed that the studies adopting ML approaches
follow the centralised training mechanism, which raises data
privacy concerns and consumes bandwidth resources. There-
fore, we also adopt federated learning (FL) approach to

collaboratively train the NN model, protect the privacy of the
visual information, and alleviate pressure on radio channels
[11]. The following points summarise the main contributions
of this paper:
•We formulate the CV-aided blockage prediction problem
for multi-user/object UDNs and develop an end-to-end
latency-aware framework that takes advantage of the
RGB images to proactively predict blockages and perform
PHO so that the QoE of users remains as high as possible.
•We consider FL as a distributed learning approach rather
than the conventional centralised learning method to train
the model locally in each small base station (SBS) where
the visual information resides and secure data privacy
while relieving the communication overhead.
•Finally, we validate the accuracy of the proposed frame-
work using modern simulation tools. The simulation
results underpin the importance of our solution in main-
taining seamless connectivity for highly dynamic UDNs.
II. NE TWORK MODEL
This study targets UDNs which are prevalent in smart cities,
where the environment is challenging due to numerous mobile
users and obstacles. We consider an outdoor mmWave system
consisting of one macro base station (BS) and many SBSs,
as depicted in Fig. 1. For clarity, this figure shows only two
SBSs and part of a street as a small portion of a UDN.
Orthogonal frequency division multiplexing (OFDM) with K
subcarriers is adopted as the modulation scheme based on
28GHz. Each SBS has a camera that monitors objects within
its field of view. Moreover, it has an M-element antenna
array that enables beamforming to serve single-antenna mobile
users with beams selected from a predefined beam steering
codebook F={fi}B
i=1, where fi∈CM×1and Bdenotes the
total number of beams.
The network focus is to determine the optimal beam that
achieves the highest RSS at the user side (f⋆). Given this,
we define an area of interest (AoI) as the coverage area that
achieves the optimal RSS when the users are connected to
the corresponding SBS. QoE is the key performance indicator
that this study aims to keep as high as possible; therefore,
we assume that the network hands over the users from one
SBS to another every time they cross the boundaries of AoIs.
Moreover, the geometric mmWave channel model is adopted
since it captures the geometrical distribution of the environ-
ment and is commonly used in practical mmWave systems
[12]. Therefore, the downlink received signal at subcarrier k
is
yk=hT
kf⋆sk+nk,(1)
where hTis the transpose of the downlink channel, sis
the transmitted symbol, and nrepresents the additive white
Gaussian noise (AWGN). In addition, the RSS at the user side
can be determined as follows:
RSS =1
K
K
X
k=1 
hT
kf⋆

2.(2)
AoI 1
SBS2
SBS1
50 80 110 140 170m
Macro BS
x
y
AoI 2
Optimal PHO
boundary
Detection
Region
Failure
Region
Blocked
Region
User A:
User A
Figure 1: The proposed system model: portion of an UDN including
one macro BS and two SBSs each equipped with a vision sensor.
III. PROB LE M FOR MU LATI ON
The beam blockage problem can be formally defined as
follows. The camera captures frames of RGB images, and
image processing is applied to produce flat1RGB (F-RGB)
images focused on the AoI. Each F-RGB image is assumed to
contain Oobjects, and every object o∈Owill be monitored
until it leaves the SBS’s AoI. The F-RGBs are fed to an
ODL algorithm to obtain boundary-boxes information about
every object. This information is then converted to a 6-
dimensional metric vector [x1, y1, xm, ym, x2, y2], where the
subscripts 1,m, and 2indicate the upper left, middle, and
lower right coordinates of the boundary boxes, respectively.
The complete mobility vector (L) of any object ois shaped
by adding it’s movement direction (d) and speed (v) as follows
Lo= [x1o, y1o, xmo, ymo, x2o, y2o, do, vo]O
o=1.
Assuming the number of wireless users in any F-RGB is U,
U⊆O, and a user u∈Uis identified from one of all objects
(as will be discussed in Section IV-B), then, the Lvector for
that user is represented as Lu= [xmu, ymu, du, vu], given that
the UE is located in the middle of the object and other objects
are possible blockages. Let Su,c represents the combination
of the wireless user and a single blocking object, denoted
as an obstacle (c), Su,c ={Lu,Lc}, u ∈U, c ∈O\{u}.
Therefore, the goal is to classify whether this sample leads
to a possible future blockage b∈ {0,1}, where 0,1 indicate
beam non-blockage or blockage, respectively. Moreover, the
study predicts the remaining time until the user gets blocked
if a link blockage is expected, denoted as TBLK , which could
be defined as:
TBLK =i , b = 1,∀i∈R+
-1, b = 0 (3)
where -1means not applicable when the sample Su,c does
not lead to a future blockage. Thus, su,c ={bu,c, TB LKu,c }is
defined as the labels associated with each data sample Su,c.
The objective of this study is achieved by using an ML
model fΘ(S)that can perform classification and regression in
1Flat term is used to indicate a 2D image that has the same metric width
anywhere.

parallel. It takes in the user-obstacle vectors Sand produces
predictions ˆs. The model predictions are governed by a set of
parameters Θadapted based on dataset of labelled samples
D={Su, su}U
u=1. This dataset trains the ML model to
reach high-fidelity for blockage status and time prediction. The
following mathematical formulas represent the purpose of the
model, which aims to maximise the probability of link status
prediction and reduce the blockage time prediction error.
max
fΘ(S)
U
Y
u=1
Pˆ
bu,c =bu,c |Su,c,∀c∈O\{u}(4)
min
fΘ(S)
U
X
u=1
(|ˆ
TBLKu,c −TBLKu,c |),∀c∈O\{u}(5)
IV. CV-AS SI ST ED DY NAMIC BLOCKAGE
PRE DI CT IO N AN D PHO
A. Key Idea and Schematic Diagram
This study focuses on a practical scenario that considers
multiple dynamic users and objects and extends our previous
work [9] that considers a single user and a stationary blocking
object. The framework is divided into several subtasks, as
illustrated in the schematic diagram in Fig. 2. Initially, the
camera at each SBS captures sequences of time-tagged RGB
images that are processed to zoom in on the respective AoI.
Then, one of the leading-edge ODL algorithms is used to
recognise objects and extract the required augmented inform-
ation, assuming that the vision sensors are unaffected by the
low light and the weather conditions. Next, it is necessary to
differentiate the wireless users from other obstacles to form
Su,c data samples. At this point, the data samples are labelled2
by blockage status and time. The complete dataset is then
stored to train the multi-output model using FL, and when the
model is ready, the unlabelled data samples will be fed directly
to the model for inference. If the predicted TBLK is greater
than the time required by the proposed framework (TF), it is
highly possible to avoid such blockages by requesting a PHO.
The following formulas illustrate the main time parameters of
the proposed solution:
TF=TODL +Tinf +TP H O,(6)
TD≤TBLK −TF,(7)
where TODL is the time associated with using the ODL al-
gorithm on two successive F-RGB images. Tinf is the model’s
inference time. TP HO is the time required for performing
PHO, and TDis the time defined to delay triggering the PHO
to the point that yields the best QoE.
B. Objects Detection and Users/Obstacles Discrimination
ODL algorithms have recently undergone many advance-
ments allowing for super-fast, real-time, and accurate de-
tection. In this study, a state-of-the-art you only look once
2Labels of data samples can be obtained analytically in the absence of
prior information or by observation, which means monitoring and recording
the users blocking status and time.
(YOLO) version 3 algorithm is adopted to detect various
objects in the F-RGB images and produce boundary boxes
indicating the positions of the objects in pixel scale [13]. The
boundary boxes are then converted to metric scale using the
conversion ratio Wm:Wp, where Wmand Wprefer to the
width of F-RGB images in meters and pixels, respectively.
This process is followed by extracting objects’ speed and
direction to build the Lvector for every object. Performing
ODL on two successive F-RGB images is necessary to de-
termine the speed and direction. The direction is determined
by noting the displacement in xlocation, whether to the left
or the right. This offset distance is divided by the difference
of the corresponding time stamps to get the object’s speed.
The study in [13] shows that performing ODL on two F-RGB
images requires 102ms, i.e. TODL = 102ms. This time will
be less if edge computing resources are employed in SBSs.
Identifying wireless users:Moving from a single-user [9]
to a multi-user environment necessitates distinguishing each
particular user from other objects in the F-RGB image. This
study uses a mapping technique in which the exact location
of the wireless user in the environment is reflected on the
F-RGB images and compared with all boundary boxes. The
object with a boundary box centre closest to the user’s location
will be considered the wireless user in the F-RGB. Several
techniques are followed to obtain the user’s position in the
wireless environment, such as GPS and RSS triangulation, but
they fail to provide an accurate location. The shift to higher
operating frequency is foreseen to improve the positioning
based on the cellular networks [14]. Moreover, several studies
have considered this research direction and proposed novel
techniques that provide very accurate user localisation [15],
[16]. Based on these developments, this work assumes that
the radio access network adopts one of these highly accurate
methods to provide the location and track the users. Therefore,
this study proposes the dynamic positioning table (DPT) in
each SBS to keep track of the user’s location, which is also
converted and reflected on the pixel scale. With DPT tables,
it is now possible to differentiate wireless users from other
objects and to build the user-obstacle data samples, Su,c.
C. Model Training and Inference: FL Approach
The nature of the defined problem is best solved using a
model that can do both classification and regression simultan-
eously. Hence this work develops a multi-output two-hidden
layer NN model fed by user-obstacle samples to predict the
blockage status and time. In addition, this study adopts FL
rather than centralised learning to protect the privacy of the
data and relieve the pressure on the communication channels.
Training Phase: The proposed framework requires a suf-
ficiently trained model before running. During the FL process,
the NN model is used as the base model to be trained by the
SBSs. The number of clients is set to three. However, the
framework can be easily extended to include many SBSs. A
parameter server (PS) in the macro BS orchestrates the training
process by selecting the number of SBSs participating in each
round and sends them the model to start the training. Each SBS
exploits its dataset to train the model locally and then sends

SBS continuously
captures RGB
images
Data preprocessing 1:
AoI focus
ODL detects different
objects and gives their
classes/pixel location
Data preprocessing 2:
deriving speed and
direction of the
detected objects
Users/obstacles
discrimination
based on DPT
Constructing
inferencing/training
data samples
Multi-output
classification and
regression model
resulted from the FL.
Store data for multi-
output model
training/retraining
in FL environment
BLK event exists?
Break and move to
the detection phase
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑇
𝐵𝐿𝐾 > 𝑇
𝐹?
Wait for 𝑇
𝐷, then
trigger proactive
handover
Yes
Yes
No
No
Figure 2: Schematic diagram of the proposed framework.
Table I: Hyperparameters of the NN model.
the model’s parameters to the PS for aggregation. Furthermore,
an early stopping patience technique is developed to avoid
suboptimal performance or excessive rounds of unnecessary
training.
Inference Phase: After the model has been trained in
the FL environment, it is ready to be used for inference. The
generated user-obstacle data samples can be fed to the trained
model for blockage status and time prediction. Furthermore,
the time associated with model inferencing requires approx-
imately 1 ms, i.e., Tinf =1ms.
D. Optimal PHO Trigger Point
The proposed framework informs the network of the need
for PHO if a blockage is expected. The principal question
is when and at what distance the PHO should be triggered.
Therefore, this study identifies two regions: detection and
failure. The detection region is defined as the region in which
the proposed framework monitors and detects the objects
within it; in this system model, this region is the same as
the AoI. The failure region is where there is no chance of
avoiding link interruption due to insufficient time remaining
to complete the PHO. The failure region is located just before
the blocked region; its width and location vary because of
the environment’s dynamicity and the user’s speed. Fig. 1
illustrates these regions.
Finding the optimal distance to trigger PHO is linked with
maintaining higher values of the user’s QoE. If a BLK event is
detected, the best scenario is to wait for the maximum delay
(Tmax
D) obtained from equation (7) and then perform PHO.
Translating this time to distance gives the optimal distance
(Dopt) defined as follows:
Dopt =v×Tmax
D.(8)
In other words, the optimal PHO must be triggered just
before the failure region in the boundary between the detection
and failure regions. Triggering the PHO anywhere within the
detection region and before the optimal PHO boundary may
also avoid the link blockage but at the cost of affecting the
perceived QoE. In addition, doing early PHO may impact the
balance and the allocation of network resources. Hence, the
objective of the proposed framework is to always trigger the
PHO within the vicinity of the optimal PHO boundary.
E. PHO Latency
The final required parameter for the proposed framework
is TP HO . In a conventional network employing beamforming,
1) Actual 𝑇
𝐵𝐿𝐾 < 𝑇
𝐹230 ✓
2) Predicted 𝑇
𝐵𝐿𝐾 <= Actual 𝑇
𝐵𝐿𝐾 2695 ✖
3) Predicted 𝑇
𝐵𝐿𝐾 > Actual 𝑇
𝐵𝐿𝐾 2075 ✓
𝑇
𝐵𝐿𝐾 categories
Samples Blockage?
Category
Figure 3: Classification and regression model performance.
if a user loses the connection with the SBS, it will undergo
certain steps to reconnect again. The complete steps are
beam failure detection, beam failure recovery, cell search,
and contention-based/free random access [10]. According to
the 3GPP specifications, each step is associated with a time
duration until completion and the total delay time is ∼312.2
ms, indicating the time associated with performing reactive
HO when contention-based random access is assumed. Us-
ing proactive blockage prediction significantly minimises this
time, and the TP HO will be equal to 80ms [9], assuming
contention-based radio access since this study targets urban
areas with dynamic wireless environments. Consequently, the
proposed framework’s TFis now determined and equals
183ms. If the predicted TBLK is greater than 183ms, the
framework has a high chance of avoiding link interruption.
Otherwise, a link interruption will happen.
V. PE RF OR MA NC E EVALUATI ON A ND RE SU LTS
This section first discusses how the NN model is developed
and then delves into describing the simulation setup and
discussing the simulation results.
A. FL-based Multi-Output Model Development
The NN model is trained under the FL setup using the feder-
ated averaging algorithm. The complete information about the
model structure and the selected hyperparameters are shown
in Table I. Model performance is tested using ten thousand
samples forming 50% blocking and 50% nonblocking. Fig.
3 displays the testing results in which the confusion matrix
demonstrates the near-optimal classification accuracy of the
model, while the table divides the TBLK into three categories
and gives the blockage status for each category.
To give a better understanding, the PHO success rate is
defined as SRP H O =Ns/NT, where Nsdenotes the number

Table II: PHO success rate versus percent shift.
Pshift % 0 1 3 5 7 9 11 13 15
SRP H O % 54 77.1 91 93.4 93.6 93.3 92.8 92.4 92
of samples with successful PHO and NTindicates the total
number of blocking samples. Therefore, we can conclude from
the table in Fig. 3 that the success rate is unsatisfactory with
54%. However, this result can be improved by making a trade-
off between the SRP H O and the QoE. Therefore, we introduce
a new parameter called the percent shift (Pshift ) defined to
reduce the predicted TBLK by this percent. This parameter
aims to move as many samples as possible from the third to
the second category to enhance the SRP H O at the cost of
a slight drop in the QoE. Table II shows how changing the
values of the Pshift affects the SRP HO , and the best value is
7%. Accordingly, Fig. 4 illustrates the cumulative distribution
function (CDF) of the samples with successful PHO versus
TDoffset, which indicates how far is the predicted TD(ˆ
TD=
ˆ
TBLK −TF) from the actual one and defined as:
TDoff set =Tmax
D−ˆ
TD
Tmax
D
×100%,∀ˆ
TD≤Tmax
D.(9)
The optimal PHO point is when the TDoff set is zero, and the
closer the samples are to this point, the better the performance.
Moreover, moving away from this point means an earlier PHO
which may affect the QoE. However, performing earlier PHO
with some QoE drop is better than losing the connection and
establishing it again, which will incur much overhead and
increase the network’s latency. Finally, the framework is now
ready to be used under the considered scenario that will be
discussed subsequently.
B. Simulation Setup
The overall performance of the proposed framework is
evaluated by considering a practical outdoor environment. The
scenario considered in this study is inspired by the vision
wireless (ViWi) ASU downtown scenario (ASUDT1) [17],
which has a very similar system model to the one adopted for
this study. ASUDT1 comprises two mmWave SBSs operating
at 28GHz and located 60 m apart on opposite sides of the
street. Each has an antenna array that forms LoS beams
to serve 60 users moving in straight trajectories. Users are
UEs placed in the center of vehicles of different sizes, such
as cars, buses, and trucks, moving at different speeds and
directions, and each can be seen as a potential obstacle for
other users. At each time instance (also known as a scene),
the ASUDT1 provides raw data for every user uconsisting
of a 4-tuple of concurrent information (user location, RGB
images, mmWave channel, link status) from each SBS, which
helps in evaluating the performance of the proposed solution.
The simulation experiments are based on Python programs,
and the key performance metrics are the PHO success rate,
the network latency, and the perceived QoE.
C. Simulation Results
Several aspects are considered to examine the efficacy of
the proposed framework. First, given the dynamicity of the
Figure 4: The distribution of the TD offset of the samples with
successful PHO.
considered environment, the impact of vehicles’ speed on
performing a successful PHO is studied. The speed is set in a
range of 1.5 to 20 mph, and a new parameter called relative
speed is introduced. For every blocking sample in Su,c, the
relative speed parameter is defined as the sum of the user’s
and obstacle’s speeds if they are moving towards each other
or the difference in their speeds if they are moving in the
same direction. This parameter measures how fast a blockage
occurs and will be used to investigate its impact on performing
successful PHO. Consequently, the relative speed from all
blocking samples of the testing dataset is divided into three
categories, slow, medium, and fast. Fig. 5(a) demonstrates the
results of this study. It can be observed that the PHO success
rate is high when the relative speed is low and medium. At the
same time, it decreases as the relative speed increases, which
is expected due to reducing the TBLK for a blockage, thus
reducing the probability of a successful PHO.
Then, the latency associated with performing HO on both
reactive- and proactive-based approaches is studied. Section
IV-E shows that the latency associated with reactive HO is
about 312.2 ms. However, the PHO requires only 80 ms,
assuming contention-based random access. Following a similar
approach from [10], the average HO latency for 5000 users is
calculated as:
ζ={ρ×U} × 80 + {(1 −ρ)×U} × 312.2
U,(10)
where Usignifies the total number of users and ρ∈[0,1] is
the percentage of users who successfully perform a PHO. Fig.
5(b) shows the average latency improvement of the proposed
PHO framework compared to the reactive mechanism. The
average latency is 94.8 ms for our CV-aided PHO solution,
which outperforms the reactive HO approach by a factor of 3.3,
vital to maintaining connectivity for real-time applications.
The final study adopts a similar approach from [9] of
considering moving users running an RTP-based application
and measures the average QoE/mean opinion score (MOS)
of a group of users with prospect blockages. This study
takes advantage of the mmWave channel information provided
by the ASUDT1 scenario. Since ASUDT1 provides plentiful
information represented as scenes for every location point,
this study spotlights on the portion of the street between

(a) (b) (c)
Figure 5: Simulation results: (a) The impact of different relative speeds on the PHO success rate, (b) Comparison of the average latency
between the reactive HO and the proposed CV-aided PHO, (c) The RSS percentage drop due to performing reactive and PHO, and how
much this drop affects the QoE measured through MOS.
the two SBSs and only focuses on the blockages within the
scene interval from 680 to 980. For every blocking, the RSS
percentage drop when a user is handed over to another SBS
is recorded and is done for all users who encounter blockages
in between the two SBSs. Then, the average percentage drop
of the RSS is calculated and mapped to the corresponding
value of the MOS. However, for the reactive-HO approach,
no proactive measures are taken and the percentage drop in
RSS is also measured to find the average percentage drop in
RSS. Fig. 5(c) illustrates the outcome of this study in which
the proposed framework can keep the MOS at a high level
despite the small drop in the average RSS. Reactive HO failed
to keep users at high MOS during interruption time. This
result confirms the potential of the proposed framework for
improving the reliability of high-frequency wireless networks
and making them suitable for latency-sensitive applications.
VI. CO NC LU SI ON S
This study explored the potential of leveraging vision in-
formation to improve the reliability of high-frequency net-
works by predicting dynamic blockages in advance and taking
measures to perform PHO. A NN multi-output model is
developed that, combined with CV technology, propose a
novel framework capable of accurately predicting blockages
and the time needed before the user reaches the blocked
region. Moreover, the model is trained using FL to protect data
privacy and conserve bandwidth resources. Simulation results
indicated that our framework achieves a high PHO success
rate of 93.6%, outperforms the reactive-HO approaches by a
factor of 3.3 in terms of latency, and maintains the QoE at
higher levels. These results highlight a promising solution for
beam blockages in multi-user mmWave/THz networks.
REF ER EN CE S
[1] M. Agiwal, A. Roy, and N. Saxena, “Next generation 5G wireless
networks: A comprehensive survey,” IEEE Commun. Surv. Tuts., vol. 18,
no. 3, pp. 1617–1655, Feb. 2016.
[2] M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks: A
survey,” IEEE Commun. Surv. Tuts., vol. 18, pp. 2522–2545, May 2016.
[3] J. Bao, T. Shu, and H. Li, “Handover prediction based on geometry
method in mmWave communications-a sensing approach,” in Proc. IEEE
Int. Conf. Commun. Workshops (ICC Workshops), Kansas City, MO,
USA, May 2018, pp. 1–6.
[4] L. Yu et al., “Long-range blockage prediction based on diffraction fringe
characteristics for mmWave communications,” IEEE Commun. Lett.,
Apr. 2022.
[5] K. Qi et al., “Dual connectivity-aided proactive handover and resource
reservation for mobile users,” IEEE Access, vol. 9, pp. 36100–36 113,
Feb. 2021.
[6] C. Wang et al., “Deep learning-based intelligent dual connectivity for
mobility management in dense network,” in Proc. IEEE 88th Vehic.
Technol. Conf. (VTC-Fall), Chicago, IL, USA, Aug. 2018, pp. 1–5.
[7] T. Nishio et al., “Proactive received power prediction using machine
learning and depth images for mmWave networks,” IEEE J. Sel. Areas
Commun., vol. 37, no. 11, pp. 2413–2427, Aug. 2019.
[8] M. Alrabeiah, A. Hredzak, and A. Alkhateeb, “Millimeter wave base
stations with cameras: Vision-aided beam and blockage prediction,”
in Proc. IEEE 91st Vehic. Technol. Conf. (VTC2020-Spring), Antwerp,
Belgium, May 2020, pp. 1–5.
[9] M. Al-Quraan et al., “Intelligent blockage prediction and proactive
handover for seamless connectivity in vision-aided 5G/6G UDNs,”
arXiv preprint arXiv:2203.16419, Feb. 2022. [Online]. Available:
http://arxiv.org/abs/2203.16419
[10] G. Charan, M. Alrabeiah, and A. Alkhateeb, “Vision-aided 6G wireless
communications: Blockage prediction and proactive handoff,” IEEE
Trans. Vehic. Technol., vol. 70, no. 10, pp. 10193–10 208, Aug. 2021.
[11] M. Al-Quraan et al., “Edge-native intelligence for 6G communications
driven by federated learning: A survey of trends and challenges,”
arXiv preprint arXiv:2111.07392, Nov. 2021. [Online]. Available:
http://arxiv.org/abs/2111.07392
[12] Q. C. Li, G. Wu, and T. S. Rappaport, “Channel model for millimeter-
wave communications based on geometry statistics,” in Proc. IEEE
Globecom Workshops (GC Wkshps), Austin, TX, 2014, pp. 427–432.
[13] J. Redmon and A. Farhadi, “YOLOv3: An incremental improvement,”
arXiv preprint arXiv:1804.02767, Apr. 2018. [Online]. Available:
http://arxiv.org/abs/1804.02767
[14] D. Dardari, P. Closas, and P. M. Djuri ´
c, “Indoor tracking: Theory,
methods, and technologies,” IEEE Trans. Vehic. Technol., vol. 64, no. 4,
pp. 1263–1278, Feb. 2015.
[15] Rastorgueva-Foi et al., “User positioning in mmW 5G networks using
beam-RSRP measurements and Kalman filtering,” in Proc. IEEE 21st
Int. Conf. Inf. Fusion (FUSION), Cambridge, UK, July 2018, pp. 1–7.
[16] M. Koivisto et al., “Joint device positioning and clock synchronization
in 5G ultra-dense networks,” IEEE Trans. Wireless Commun., vol. 16,
no. 5, pp. 2866–2881, Mar. 2017.
[17] M. Alrabeiah et al., “ViWi: A deep learning dataset framework for
vision-aided wireless communications,” in Proc. IEEE 91st Vehicu-
lar Technology Conference (VTC2020-Spring), Antwerp, Belgium, May
2020, pp. 1–5.