Situational Collective Perception: Adaptive and Efﬁcient Collective

Perception in Future Vehicular Systems

Ahmad Khalil

1 a

, Tobias Meuser

1 b

, Yassin Alkhalili

1 c

, Antonio Fernandez Anta

2 d

Lukas Staecker

3 e

and Ralf Steinmetz

1 f

Multimedia Communications Lab, Technical University of Darmstadt, Darmstadt, Germany

IMDEA Networks Institute, Madrid, Spain

Stellantis, Opel Automobile GmbH, R

usselsheim, Germany

Keywords:

Collective Perception, Vehicular Networks, Intelligent Transportation Systems, V2X, Federated Learning.

Abstract:

With the emerge of Vehicle-to-everything (V2X) communication, vehicles and other road users can perform

Collective Perception (CP), whereby they exchange their individually detected environment to increase the

collective awareness of the surrounding environment. To detect and classify the surrounding environmental

objects, preprocessed sensor data (e.g., point-cloud data generated by a Lidar) in each vehicle is fed and classi-

ﬁed by onboard Deep Neural Networks (DNNs). The main weakness of these DNNs is that they are commonly

statically trained with context-agnostic data sets, limiting their adaptability to speciﬁc environments. This may

eventually prevent the detection of objects, causing safety disasters. Inspired by the Federated Learning (FL)

approach, in this work we tailor a collective perception architecture, introducing Situational Collective Percep-

tion (SCP) based on dynamically trained and situational DNNs, and enabling adaptive and efﬁcient collective

perception in future vehicular networks.

1 INTRODUCTION

The number of vehicles on the roads is expected to

reach 1.8 billion worldwide by 2035 (L

ofﬂer, 2021),

therefore more crowded roads will continue to be a

common phenomenon in the future. Crowded roads

not only increase the probability of having accidents

but also reduce the overall trafﬁc efﬁciency. This

highlights the necessity to develop and deploy more

intelligent systems in the vehicles, to assist the vehi-

cle’s driver. Supporting intelligent systems help in-

creasing trafﬁc safety and efﬁciency, by making the

vehicles’ actions less dependent, or even fully inde-

pendent, from the driver decisions.

However, when taking over, the intelligent sys-

tems in the vehicles need to be aware of their envi-

ronment. By depending only on their onboard sen-

https://orcid.org/0000-0002-6059-7027

https://orcid.org/0000-0002-2008-5932

https://orcid.org/0000-0001-5722-1910

https://orcid.org/0000-0001-6501-2377

https://orcid.org/0000-0001-7537-606X

https://orcid.org/0000-0002-6839-9359

sors, the vehicles have limited perception of their sur-

rounding environment. Therefore, enabling commu-

nication to exchange perception data between vehi-

cles and other road users is a pivotal aspect to ensure

reaching the intended level of environmental aware-

ness.

With the emerging Vehicle-to-everything (V2X)

technology, road users can exchange data, which

opens the horizons for a multitude of applications

(Boban et al., 2018), like Cooperative Awareness

(Sjoberg et al., 2017), Cooperative Maneuver (3GPP,

2016), Teleoperated Driving (3GPP, 2016), and Col-

lective Perception (Shan et al., 2021; Fukatsu and

Sakaguchi, 2019; Pilz et al., 2021; Shan et al., 2021;

Barbieri et al., 2021).

In this work, we focus on some of today’s issues

of Collective Perception (CP) and introduce a possi-

ble solution for enhancing the vehicles’ perception in

speciﬁc environments.

Having different perception levels of the sur-

rounding environment, vehicles and other road users

can participate in the collective perception by ex-

changing their detected objects. This helps the road

users to extend their perception range including the

346

Khalil, A., Meuser, T., Alkhalili, Y., Anta, A., Staecker, L. and Steinmetz, R.

Situational Collective Perception: Adaptive and Efﬁcient Collective Perception in Future Vehicular Systems.

DOI: 10.5220/0011065000003191

In Proceedings of the 8th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2022), pages 346-352

ISBN: 978-989-758-573-9; ISSN: 2184-495X

objects outside their Field of view (FOV) (e.g., ob-

jects hidden by a building). Nevertheless, there is

a wide set of challenges within CP that needs to be

considered. As CP is performed in three main steps,

sensing, communicating, and fusing (Pilz et al., 2021),

challenges can arise in each one of these steps. More-

over, having diverse levels of sensing technologies

on the vehicles, the huge amount of transmitted data

(Fukatsu and Sakaguchi, 2019), and the use of stati-

cally trained onboard Deep Neural Networks (DNNs)

used for objects detection, are open challenges that

have to be tackled by researchers.

In this work, we handle the issue of having on-

board DNNs which are statically trained with context-

agnostic data sets. These DNNs are statically trained

and deployed on the vehicles, which could affect their

detection performance in some certain circumstances

(e.g., fog, heavy rain at night). Moreover, DNNs suf-

fer from the high costs of performance improvement

(NEIL et al., 2021), as the amount of data required

for generalized DNNs scale exponentially with the re-

quired accuracy of these models.

Federated Learning (FL) has recently attracted

considerable interest in the ﬁeld of vehicular networks

research, and more literature has been published in

recent years showing its ability to train DNNs with a

less amount of transmitted data (Otoum et al., 2020;

Yu et al., 2020; Yuan et al., 2021; Boualouache and

Engel, 2021; Zhao et al., 2021; Ding et al., 2021).

Nonetheless, previous research has tended to focus on

using FL to dynamically train generalized vehicles’

onboard models and failed to make them context-

aware.

To alleviate such issues, and inspired by FL con-

cepts, the main contribution of this work is to in-

troduce our forward-looking Situational Collective

Perception (SCP), which enables dynamically-trained

and situation-aware onboard DNNs. We tailor the CP

architecture to enable training the situational DNNs

with context-speciﬁc data, and deploy it on the vehi-

cles to enhance their detection capability in speciﬁc

situations.

The rest of the paper is organized as follows: In

Section 2 we provide more insight into CP, its ele-

ments, key features, and its main challenges. Section

3 presents an overview of FL and highlights the recent

works on applying FL concepts in vehicular networks.

We introduce Situational Collective Perception (SCP)

in Section 4, presenting its potential for enhancing the

detection capability of the vehicles. In this section,

we also highlight the possible future research direc-

tions and challenges of adopting SCP in vehicular net-

works. Finally, in Section 5, we conclude this work.

2 COLLECTIVE PERCEPTION

The term collective perception refers to one of the

emerging technologies in Intelligent Transportation

Systems (ITSs). Figure 1 illustrates the main compo-

nents of collective perception. Each vehicle uses its

onboard sensors (e.g., camera, lidar, radar) to gather

data about the surrounding environment. The gener-

ated raw sensor data is preprocessed (e.g., size re-

duction, fusion) and fed to Deep Neural Networks

(DNNs) for object detection and classiﬁcation. The

vehicle uses a list of the detected objects and merges

it with their spatial information to build its local envi-

ronment model. This local environment model is the

basis of decision-making and planning systems.

However, the main weakness of this individual

perception (without data exchange between vehi-

cles) is that it highly depends on the onboard sensor

technology (obstruction, detection range, precision),

which restricts the perception range to each vehicle’s

Field of view (FOV). To tackle this issue, and as il-

lustrated in Figure 1, each vehicle can exchange per-

ception data with other vehicles (or road users) to ex-

tend its awareness beyond its limited individual per-

ception.

In general, there are two ways to exchange data

between the vehicles (Meuser, 2020): (a) with the

help of base stations (infrastructure-based) which

play the role of data forwarders between the vehi-

cles, (b) without any help of external infrastructure

(infrastructure-less) in which vehicles directly ex-

change data with each other. The appropriate com-

munication type depends on the use case. More-

over, and in some use cases, a hybrid type should

be applied, using both aforementioned communica-

tion types for serving a certain use case. Regardless

of the communication type, each vehicle encapsulates

its perception data in Collective Perception Messages

(CPMs) (ETSI, 2019) and sends them to other vehi-

cles. This perception data comes with different de-

tail levels (Pilz et al., 2021). The receiving vehicle

then fuses the received objects with its local environ-

ment model. This includes different processing tasks,

like coordinate transformation (Shan et al., 2021), and

enriches the vehicle’s perception of the surrounding

environment. This enhanced perception can be ex-

ploited in many other safety and efﬁciency applica-

tions in vehicular networks.

Although collective perception sounds promising,

and applying it can be essential for enhancing safety

and efﬁciency, previous works, however, show that it

comes with different kinds of challenges. In the next

section, we introduce some of the main collective per-

ception challenges.

Situational Collective Perception: Adaptive and Efﬁcient Collective Perception in Future Vehicular Systems

347

KOM – Multimedia Communications Lab 4

Decisions making and planning systems

Detection and classification

Collective Perception Architecture

Vehicle 1

Preprocessing

https://mtri.org/automotivebenchmark.html

https://www.eenewsautomotive.com/news/deep-learning-method-improves-environment-perception-self-driving-cars

Data Exchange

Base Station

Sensors’ data

Local environment model

Decisions making and planning systems

Detection and classification

Vehicle 2

Preprocessing

Sensors’ data

Local environment model

(a)

(b)

Figure 1: Illustration of the main components of collective perception. (a) Two vehicles are exchanging perception data with

the help of a base station (infrastructure-based). (b) Two vehicles are directly exchanging perception data (infrastructure-less).

2.1 Collective Perception Challenges

Collective perception is a complex task and comes

with many challenges. While performing collective

perception, issues can occur in its all stages, sens-

ing, communicating, and fusing (Pilz et al., 2021).

This introduces major computational and commu-

nication difﬁculties. In the following, we will fo-

cus on two main challenges of collective perception:

we start with illustrating the issue of having stati-

cally trained DNNs and making the case for context-

speciﬁc DNNs. Then, networking issues that could

occur while exchanging the heavy perception data are

presented.

2.1.1 Statically Trained DNNs

One pivotal aspect which affects the vehicle’s scene

understanding capability is how the DNNs used for

object detection perform. The main weakness of

the currently used DNNs comes from the fact that

they are statically trained. Moreover, the data sets

used for the training are commonly context-agnostic,

which means that they are not tailored to speciﬁc sit-

uations (context-speciﬁc). These situations can be,

e.g., location-speciﬁc (geographic), time-speciﬁc, or

weather-speciﬁc. Using context-agnostic data sets for

statically training the DNNs reduces their detection

ability in certain situations, which may cause safety

disasters (Dickson, 2020). On the other hand, to

improve the performance of the DNNs, a massive

amount of context-agnostic training data should be

gathered, and a long training time is required. The

training time and the amount of data required for the

generalized DNNs scale exponentially with the re-

quired accuracy of these models (NEIL et al., 2021).

Moreover, these DNNs may not reach, in the end,

the intended performance in speciﬁc situations. This

highlights the interest of evolving to DNNs that are

dynamically trained and more speciﬁc to the differ-

ent kinds of situations. In Sections 3 and 4, we

will describe an approach on training and deploying

situation-speciﬁc DNNs by employing the emerging

Federated Learning (FL) concepts in the collective

perception process.

2.1.2 High-volume Data Exchange

In their work, Fukatsu et al. (Fukatsu and Sakaguchi,

2019) measured the data rate required for vehicles

equipped with 3D LiDAR trying to execute collec-

tive perception for overtaking. Their study showed

that even with vehicles’ velocity of 50 km/h, the re-

quired data rate can easily reach around 6 Gbps if

only raw data was exchanged. Considering that this

data rate is generated only by a single perception ap-

plication, thus, running more vehicular applications

which require cooperation and data exchange between

the vehicles will introduce serious challenges for the

vehicular communication network. This emphasizes

the importance of reducing the amount of transmit-

ted data, either by reducing the number of exchanged

messages or by reducing its volume. In Section 4,

we propose Situational Collective Perception (SCP)

to reduce the data exchanged, making it suitable for

improving vehicles’ perception.

3 FEDERATED LEARNING IN

VEHICULAR NETWORKS

In this section, we provide background on the applica-

tion of Federated Learning (FL) approaches in vehic-

ular networks. We employ FL as a basis for our Sit-

uational Collective Perception (SCP) approach. FL

is a promising machine learning model-training ap-

proach which has been introduced (Brendan McMa-

han, 2017) as a solution for privacy issues of the con-

ventional model training approaches (Li et al., 2020).

With FL, instead of collecting the training data from

clients in the cloud, data is kept locally at each client.

A selected set of clients train their models locally with

their local data and then send the updated models (i.e.,

the gradients) back to the server, which aggregates the

model and sends it back to the clients. With its ability

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

348

KOM – Multimedia Communications Lab 7

Dynamic Model Training Approaches

https://bdtechtalks.com/2020/07/29/self-driving-tesla-car-deep-learning/

https://www.tesladeaths.com/

Initial approach

Collecting

data before

training

Sharing

model after

training

Issues:

 Huge amount of transmitted data.

 Privacy

Federated learning approach

Collecting

gradients

Sharing the

global model

Improvements:

 Less amount of transmitted data (only

gradients and metadata).

 Privacy preserved.

Model

training

Gradients

aggregation

(a) Central training

KOM – Multimedia Communications Lab 7

Dynamic Model Training Approaches

https://bdtechtalks.com/2020/07/29/self-driving-tesla-car-deep-learning/

https://www.tesladeaths.com/

Initial approach

Collecting

data before

training

Sharing

model after

training

Issues:

 Huge amount of transmitted data.

 Privacy

Federated learning approach

Collecting

gradients

Sharing the

global model

Improvements:

 Less amount of transmitted data (only

gradients and metadata).

 Privacy preserved.

Model

training

Gradients

aggregation

(b) FL training

Figure 2: Illustration of two dynamic detection deep neural

networks training approaches, (a) shows the conventional

approach, in which raw data (images, point cloud data) is

transmitted to the central server to be used for training the

DNN. (b) vehicles only transmit their model updates to the

central server to be aggregated.

to work efﬁciently with Non-Independent and Identi-

cally Distributed (Non-IID) data of a huge number of

clients, FL has attracted the interests of many indus-

trial ﬁelds. Recently, FL has received considerable

attention in vehicular networks research (Du et al.,

2020; Elbir et al., 2020). More literature is getting

published in recent years highlighting the privacy-

preserving features of applying FL to vehicular net-

works (Otoum et al., 2020; Yu et al., 2020; Yuan et al.,

2021; Boualouache and Engel, 2021; Zhao et al.,

2021; Ding et al., 2021). In (Vyas et al., 2020), Vyas

et al. proposed an architecture based on FL for de-

veloping a model that predicts the driver stress level,

and then use these predictions as a basis for a driver

recommendation system.

Furthermore, FL facilitates obtaining

dynamically-trained onboard Deep Neural Net-

works (DNNs) with a reduced volume of exchanged

data. So far, there are two approaches to obtain

dynamically-trained DNNs, to improve its perfor-

mance over time. Figure 2 illustrates those two

approaches. In the conventional DNNs training

approach (Figure 2a), vehicles transmit the raw data

(images, point cloud data, etc.) to a central server.

The central server uses the gathered data to train

the global model and transmits it to the vehicles

after validating it. This approach induces heavy data

volume transmission, which can lead to network

congestion issues. Moreover, the transmitted data can

contain sensitive information which can be leaked

during transmission or while processing it at the

central server (Nasr et al., 2019), which leads to

privacy issues. On the other hand, and as depicted

in Figure 2b, the FL DNN training approach has the

potential to address such networking and privacy

issues. Instead of transmitting heavy volume and

sensitive data to the central server, a selected number

of vehicles can participate in the training process

without transmitting their local data to the server. The

central server collects vehicles’ model updates and

aggregates them using common averaging algorithms

like federated averaging (Kone

y et al., 2016).

After that, the vehicles receive the global model,

which replaces their local models (DNNs).

This approach enables maintaining dynamically-

trained DNNs, taking into consideration the commu-

nication network limitations as well as the privacy

concerns (Barbieri et al., 2021; Nguyen et al., 2021).

Nevertheless, the existing literature using FL in ve-

hicular applications focus on generating DNNs that

are trained with context-agnostic data. Making these

DNNs prone to fail while detecting objects in certain

situations as mentioned in Section 1. A reasonable

approach to tackle this issue could be to introduce

situation-speciﬁc DNNs which can be used by the ve-

hicles in speciﬁc situations.

In Section 4, we tailor the collective perception ar-

chitecture in a FL way to enable dynamically trained,

and situational DNNs which lead to adaptive, and ef-

ﬁcient collective perception.

4 SITUATIONAL COLLECTIVE

PERCEPTION

In this section, we provide a comprehensive expla-

nation of our forward-looking Situational Collective

Perception (SCP) architecture as a potential solu-

tion for the issues of having onboard Deep Neural

Networks (DNNs) which are statically trained with

context-agnostic data.

In the following, we introduce the SCP roles, then

explain the models exchanging workﬂow between

these different roles. Next, we highlight some of

the potential research directions and challenges in the

area of SCP.

4.1 SCP Roles

Our proposed SCP design consists of three compo-

nents (see Figure 3):

• Vehicles or Other Entities Participating in Col-

lective Perception (CP): each vehicle communi-

cates with appropriate edge servers to retrieve the

suitable situational models, after detecting a situa-

tion change. With the increased number of vehic-

ular applications, car manufacturers are continu-

ously improving the cars’ computational power.

Thus, we assume that the vehicles have the re-

quired computational power for training and vali-

dating the received situational models. Moreover,

we assume for simplicity that all vehicles have

the same sensor setup, so that they can work with

copies of the same situational DNNs.

Situational Collective Perception: Adaptive and Efﬁcient Collective Perception in Future Vehicular Systems

349

KOM – Multimedia Communications Lab 12

Detection and classification

Research Challenges/Approaches

Situational models

exchange

Edge server

Cloud server

Decisions making and planning systems

Vehicle

Preprocessing

Sensors’ data

Local environment model

Figure 3: Illustration of three research directions in the situational collective perception.

• Edge Server: unlike a simple base station illus-

trated in Figure 1 which only plays the role of data

forwarder between the vehicles, we assume that

the edge server has the required computation and

communication power to maintain its responsibil-

ities. The edge server is an essential part of the

SCP architecture, as it is responsible for manag-

ing the different situational models, and sending

the proper one to the legitimate vehicles. Besides

communicating with vehicles, the edge server has

to communicate with other edge servers and the

cloud server in order to exchange and validate the

different situational models.

• Cloud Server: the main task of the cloud server

is to orchestrate the overall SCP processes. The

cloud server is responsible for initializing the sit-

uational models and transferring them to the edge

servers. Having a wide set of different situa-

tional models spread across the edge servers, the

cloud server communicates with all edge servers

to maximize the beneﬁts by combining the learn-

ing parameters contained in the situational mod-

els.

4.2 SCP Workﬂow

To express the validity of the situational models, we

introduce two different models’ ﬂags: The training

ﬂag indicates that the current version of the model is

non-valid and vehicles can not rely on it for detec-

tion purposes. Thus, vehicles have to train the model

with their local data, to eventually make it valid. On

the other hand, the valid ﬂag indicates that the cur-

rent version of the model is valid and thus, vehicles

can use it for detection purposes and train it simulta-

neously.

The overall process starts when the cloud server

uses a pre-deﬁned set of situations to initiate the sit-

uational models. These models then are trained ini-

tially by the cloud server with the available context-

speciﬁc data sets. If no context-speciﬁc data set is

found for a situational model, the cloud server skips

the initial training process for that model. Later, the

cloud server transmits the situational models to the

edge servers, after marking the models with training

ﬂags, as they are not ready yet to be used for detection

purposes.

Situation

detection

Model training ||

Model training

and using

Situation

concluded

Yes

Other edge servers

Cloud server

“training || valid”

Edge Server

Vehicle

Model merging

with local model

Model marking

with “valid” flag

Merged model

validation

Valid

model

Situational model

checking

Model

available

Model

transmission

Start

Yes

Model

transmission

Model marking

with “training”

flag

Figure 4: Illustration of situational models exchanging

workﬂow.

After that, and as depicted in Figure 4, when

the vehicle detects a new situation, it sends a situa-

tional model request to the edge server. The Edge

server checks the availability of the requested situ-

ational model. If the situational model is available,

the edge server transmits it to the requesting vehi-

cle. Upon receiving the situational model, the vehi-

cle checks the model’s ﬂag. If the model is training

ﬂagged, the vehicle only trains it with its locally gath-

ered data. On the other hand, if the vehicle receives a

valid model, it can use it for detection purposes while

training it simultaneously. As soon as the vehicle de-

tects the situation change, it re-transmits the model

back to the edge server for validation. When receiving

a model from the vehicle, the edge server merges it

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

350

with its local situational model and validates the pro-

duced model. If the validation for produced model

fails (e.g., the new version is less performing than the

previous one and it does not reach a speciﬁc perfor-

mance threshold), the edge server marks it with the

training ﬂag. On the other hand, if a valid model is

obtained, the edge server marks it with the valid ﬂag.

Finally, the edge server replaces the old version of the

model with the newly produced version.

Although this workﬂow sounds feasible, one

question that needs to be asked, however, is how the

vehicles will be able to annotate their local data. In

the next section 4.3, we will provide one possible

technique that enables the vehicles to annotate their

local data before using it for training the situational

models.

4.3 SCP Research Directions and

Challenges

Figure 3 reveals three possible research areas related

to SCP. Each research area is associated with a set of

challenges that needs to be addressed in order to ob-

tain beneﬁts from employing SCP in real-world appli-

cations.

We organize SCP future research directions under

three headings:

1. Utilize Situational Collective Perception to En-

able Situational Models Exchange: since the

idea of using situation-speciﬁc models is not in-

troduced in the past in the vehicular applications

area, it is important to evaluate its applicability in

real-world scenarios. Moreover, during the mod-

els exchanging process (see Figure 4), the vehicles

have to annotate their locally gathered data before

using it for training the situational models. This

is considered to be one of the main challenges

in Federated Learning (FL). A possible way to

handle this challenge is by using Active Learning

(AL) approach (Ahmed et al., 2020). With AL,

vehicles can use pre-trained models for selecting

and annotating the training data with an expected

level of uncertainty. This annotated data can be

used then to train the situational models. More-

over, comprehensive studies are required to eval-

uate if the SCP approach outperforms in speciﬁc

situations the conventional approach of having

statically trained models using context-agnostic

data. In addition, the performance of SCP relies

mainly on the number of participating vehicles.

Thus, enabling efﬁcient SCP in rural areas could

be a difﬁcult task.

2. Situational Models’ Validation: it is important

to bear in mind that vehicular applications are

very safety and security-sensitive. Intensive val-

idation mechanisms need to be developed to en-

sure the validity of the different situational mod-

els. Validation is not only required from the edge

server-side but also the vehicles must validate the

received models. Vehicles may require to perform

a validity check on the valid models received be-

fore relying on them for detection. On the other

hand, edge servers are required to validate the

situational models continuously, with the help of

other edge servers and the cloud server. It is cru-

cial though to develop validation criteria, deﬁning

the most important parameters which make the

model either marked with valid or training ﬂag.

Another possible area of future research would be

to investigate on how to ensure that SCP architec-

ture is resistant against the anomalous nodes (ve-

hicles, or edge servers).

3. Situational Collective Perception at Large

Scale: perhaps one of the most powerful fea-

tures of SCP comes from having a huge collec-

tion of multiple situational models spread across

a wide area. However, in order to maximize

the beneﬁts, future research has to develop inten-

sive mechanisms to deal with the heterogeneity at

a large scale (different models for diverse situa-

tions). Another possible area of future research

would be to investigate on how to manage the sit-

uational models’ exchange between the different

entities, to ensure robustness.

5 CONCLUSION

To detect their surroundings, vehicles employ stati-

cally trained onboard DNNs. These DNNs are prone

to detection failures in some situations, as they are

commonly trained with context-agnostic data. In

this work, we introduced the forward-looking SCP to

tackle these issues. Inspired by the FL approach, we

tailor the collective perception architecture to enable

dynamically trained and situation-speciﬁc DNNs. We

aim to facilitate enabling adaptive, and efﬁcient col-

lective perception in future vehicular networks. We

highlighted three possible SCP research areas and em-

phasized some of its expected challenges.

ACKNOWLEDGEMENTS

This work has been funded by the German Research

Foundation (DFG) within the Collaborative Research

Center (CRC) 1053 MAKI.

Situational Collective Perception: Adaptive and Efﬁcient Collective Perception in Future Vehicular Systems

351

REFERENCES

3GPP (2016). Study on enhancement of 3gpp support for

5g v2x services (release 15). Tech. Report TR, 22.

Ahmed, L., Ahmad, K., Said, N., Qolomany, B., Qadir, J.,

and Al-Fuqaha, A. (2020). Active learning based fed-

erated learning for waste and natural disaster image

classiﬁcation. IEEE Access, 8:208518–208531.

Barbieri, L., Savazzi, S., and Nicoli, M. (2021). Decen-

tralized federated learning for road user classiﬁca-

tion in enhanced v2x networks. In 2021 IEEE Inter-

national Conference on Communications Workshops

(ICC Workshops), pages 1–6. IEEE.

Boban, M., Kousaridas, A., Manolakis, K., Eichinger, J.,

and Xu, W. (2018). Connected roads of the future:

Use cases, requirements, and design considerations

for vehicle-to-everything communications. IEEE ve-

hicular technology magazine, 13(3):110–123.

Boualouache, A. and Engel, T. (2021). Federated learning-

based scheme for detecting passive mobile attackers

in 5g vehicular edge computing. Annals of Telecom-

munications, pages 1–20.

Brendan McMahan, Daniel Ramage, R. S. (2017). Feder-

ated learning: Collaborative machine learning without

centralized training data.

Dickson, B. (2020). Why deep learning won’t give us level

5 self-driving cars.

Ding, A. Y., Peltonen, E., Meuser, T., Aral, A., Becker,

C., Dustdar, S., Hiessl, T., Kranzlmuller, D., Liyan-

age, M., Magshudi, S., et al. (2021). Roadmap

for edge ai: A dagstuhl perspective. arXiv preprint

arXiv:2112.00616.

Du, Z., Wu, C., Yoshinaga, T., Yau, K.-L. A., Ji, Y., and Li,

J. (2020). Federated learning for vehicular internet of

things: Recent advances and open issues. IEEE Open

Journal of the Computer Society, 1:45–61.

Elbir, A. M., Soner, B., and Coleri, S. (2020). Feder-

ated learning in vehicular networks. arXiv preprint

arXiv:2006.01412.

ETSI (2019). Etsi tr 103 562.

Fukatsu, R. and Sakaguchi, K. (2019). Millimeter-wave v2v

communications with cooperative perception for auto-

mated driving. In 2019 IEEE 89th Vehicular Technol-

ogy Conference (VTC2019-Spring), pages 1–5. IEEE.

Kone

y, J., McMahan, H. B., Yu, F. X., Richt

arik, P.,

Suresh, A. T., and Bacon, D. (2016). Federated learn-

ing: Strategies for improving communication efﬁ-

ciency. arXiv preprint arXiv:1610.05492.

ofﬂer, R. (2021). How many cars are in the world?

Li, T., Sahu, A. K., Talwalkar, A., and Smith, V. (2020).

Federated learning: Challenges, methods, and fu-

ture directions. IEEE Signal Processing Magazine,

37(3):50–60.

Meuser, T. (2020). Data management in vehicular

networks-relevance-aware networking for advanced

driver assistance systems.

Nasr, M., Shokri, R., and Houmansadr, A. (2019). Compre-

hensive privacy analysis of deep learning: Passive and

active white-box inference attacks against centralized

and federated learning. In 2019 IEEE symposium on

security and privacy (SP), pages 739–753. IEEE.

NEIL, C., THOMPSON, K., GREENEWALD, K., LEE, G.,

and MANSO, F. (2021). Deep learning’s diminishing

returns.

Nguyen, A., Do, T., Tran, M., Nguyen, B. X., Duong, C.,

Phan, T., Tjiputra, E., and Tran, Q. D. (2021). Deep

federated learning for autonomous driving. arXiv

preprint arXiv:2110.05754.

Otoum, S., Al Ridhawi, I., and Mouftah, H. T. (2020).

Blockchain-supported federated learning for trustwor-

thy vehicular networks. In GLOBECOM 2020-2020

IEEE Global Communications Conference, pages 1–

6. IEEE.

Pilz, C., Ulbel, A., and Steinbauer-Wagner, G. (2021). The

components of cooperative perception-a proposal for

future works. In 2021 IEEE International Intelligent

Transportation Systems Conference (ITSC), pages 7–

14. IEEE.

Shan, M., Narula, K., Wong, Y. F., Worrall, S., Khan, M.,

Alexander, P., and Nebot, E. (2021). Demonstrations

of cooperative perception: safety and robustness in

connected and automated vehicle operations. Sensors,

21(1):200.

Sjoberg, K., Andres, P., Buburuzan, T., and Brakemeier, A.

(2017). Cooperative intelligent transport systems in

europe: Current deployment status and outlook. IEEE

Vehicular Technology Magazine, 12(2):89–97.

Vyas, J., Das, D., and Das, S. K. (2020). Vehicular edge

computing based driver recommendation system us-

ing federated learning. In 2020 IEEE 17th Interna-

tional Conference on Mobile Ad Hoc and Sensor Sys-

tems (MASS), pages 675–683. IEEE.

Yu, Z., Hu, J., Min, G., Xu, H., and Mills, J. (2020). Proac-

tive content caching for internet-of-vehicles based on

peer-to-peer federated learning. In 2020 IEEE 26th

International Conference on Parallel and Distributed

Systems (ICPADS), pages 601–608. IEEE.

Yuan, X., Chen, J., Zhang, N., Fang, X., and Liu, D. (2021).

A federated bidirectional connection broad learning

scheme for secure data sharing in internet of vehicles.

China Communications, 18(7):117–133.

Zhao, L., Ran, Y., Wang, H., Wang, J., and Luo, J. (2021).

Towards cooperative caching for vehicular networks

with multi-level federated reinforcement learning. In

ICC 2021-IEEE International Conference on Commu-

nications, pages 1–6. IEEE.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

352