Autonomous Driving Architectures, Perception and Data Fusion: A Review

Abstract

Over the last 10 years, huge advances have been made in the areas of sensor technologies and processing platforms, pushing forward developments in the field of autonomous vehicles, mostly represented by self-driving cars. However, the complexity of these systems has been also increased in terms of the hardware and software within them, especially for the perception stage in which the goal is to create a reliable representation of the vehicle and the world. In order to manage this complexity, several architectural models have been proposed as guidelines to design, develop, operate and deploy self-driving solutions for real applications. In this work, a review on autonomous driving architectures is presented, classifying them into technical or functional architectures depending on the domain of their definition. In addition, the perception stage of self-driving solutions is analysed as a component of the architectures, detailing into the sensing part and how data fusion is used to perform localisation, mapping and object detection. Finally, the paper is concluded with additional thoughts on the actual status and future trends in the field.

Full text

Autonomous Driving Architectures, Perception and
Data Fusion: A Review
Gustavo Velasco-Hernandez∗†, De Jong Yeong∗, John Barry∗and Joseph Walsh∗
∗IMaR Research Centre / Lero, The Irish Software Research Centre
Institute of Technology Tralee
Tralee, Ireland
†gustavo.velascohernandez@staff.ittralee.ie
ORCID 0000-0002-2177-6348
Abstract—Over the last 10 years, huge advances have been
made in the areas of sensor technologies and processing plat-
forms, pushing forward developments in the field of autonomous
vehicles, mostly represented by self-driving cars. However, the
complexity of these systems has been also increased in terms
of the hardware and software within them, especially for the
perception stage in which the goal is to create a reliable
representation of the vehicle and the world. In order to manage
this complexity, several architectural models have been proposed
as guidelines to design, develop, operate and deploy self-driving
solutions for real applications. In this work, a review on au-
tonomous driving architectures is presented, classifying them into
technical or functional architectures depending on the domain
of their definition. In addition, the perception stage of self-
driving solutions is analysed as a component of the architectures,
detailing into the sensing part and how data fusion is used to
perform localisation, mapping and object detection. Finally, the
paper is concluded with additional thoughts on the actual status
and future trends in the field.
Index Terms—Architecture, Autonomous Driving, Autonomous
Vehicles, Data Fusion, Localisation, Mapping, Perception, Self-
Driving Car, Sensor Fusion.
I. INTRODUCTION
Several initiatives for the development of self-driving solu-
tions have been created in industry and academia, but there
is not a one-size-fits-all approach that could accomplish all
the goals or to solve all the issues present in each appli-
cation scenario. However, there are certain components and
processing stages that are shared among different projects
in order to complete the main objective of such a system,
this is, autonomous navigation of a platform in a specified
environment.
On the component side, there are hardware parts, like
positioning and range sensors, networks, embedded and High-
Performance Computing (HPC) platforms, and also software
components, low-level embedded and high-level application
software. Regarding the processing stages, they can be sum-
marised in these four categories: sensing and perception, pro-
cessing and planning, vehicle control, and system supervision.
No matter whether the platform is a car on the road, a robot
in a warehouse, a tractor in a crop field or a lifting vehicle
in a building site, these stages are part of any self-driving
architecture definition. The differences between them will be
978-1-7281-9080-8/20/$31.00 ©2020 IEEE
the conditions where they will operate, encompassed in their
Operational Design Domain (ODD), which defines the scope
and limitations of the environment where a self-driving system
or feature should work, including but not limited to weather,
terrain, time-of-day, location, etc.
An overall description of the common elements of a self-
driving solution can be found in [1] and [2], however, it is
required to organise these elements in a way that enables a suc-
cessful product development life cycle, and avoid reinventing
the wheel for every self-driving development to be done. For
this reason, several authors have proposed architectural models
for the design, development, and deployment of autonomous
driving systems, both from a technical and a functional point
of view.
Some of these initiatives have a wider scope while others
focus on certain specific aspects. One of these aspects is
the perception stage of autonomous vehicles, where multiple
sensors provide information using different modalities with the
objective of creating a reliable representation of the vehicle,
its status and the world that surrounds it. This supposes a
challenge because all the sensors could be seen as separate
subsystems generating streams of data, but this data needs to
be merged in an appropriate way. This process is known as
data fusion and it handles the collection, association of data
and the creation of better representations than those obtained
by using the sensors individually.
The rest of this paper is structured as follows: Section II pro-
vides an overview of autonomous driving architectures found
in the literature. These are classified based on their domain or
abstraction level (functional vs technical) and each one of their
inner elements is described. Additional concepts on sensors,
perception and data fusion are presented in section III, starting
with a description of typical sensors used in a self-driving
environment and then, different developments in the field of
perception are highlighted, especially in two areas: localisation
and mapping, and object detection. Then an overview of data
fusion, its relevance and its challenges is presented, including
recent work in the area. Finally, conclusions are presented in
section IV with additional thoughts on the actual status and
future trends on autonomous vehicles, especially in perception
and data fusion as part of a self-driving architecture.

II. AUTONOMOUS DRIVING ARCHITECTURES
When representing a complex system using an architectural
model, there are different perspectives from which the system
can be viewed, for example in terms of physical components,
development stages, logical functions or process blocks, as
described in [3]. In the present work, autonomous driving ar-
chitectures are considered from two viewpoints: 1) a technical
viewpoint, which is related to hardware and software com-
ponents and their implementations, and 2) from a functional
viewpoint, which includes the description of processing stages
that a self-driving vehicle must have as logical blocks of the
whole system.
A. Technical View
Hardware and software are the two main layers of the
technical view of an autonomous driving architecture and each
layer includes components that represent different aspects of
the whole system. Some of these components can be seen as
an isolated group, but there are some components that act like
a backbone within their own layer providing a structure and
guidelines for the interactions between all the components.
This description is depicted in figure 1.
Sensors Processing
units
V2X/Cloud
Comms
Internal Networking Interfaces
(CAN, LIN, GigaEthernet, USB 3.X, etc)
Software Frameworks and Standards
AUTOSAR, ROS, ROS2, RTOS etc.
ML, AI, DL
algorithms
Data
Collection
Real-time and
Critical Control
Software
UI/UX and
infotainment
Mobile
Platform/
Actuators
Hardware Software
Fig. 1. Technical architecture for an autonomous driving system
Autonomous vehicles nowadays are large complex systems
equipped with several sensors, for internal and external mon-
itoring and generating massive amounts of data per day. In
order to handle all that information, communications and
processing units in the vehicles are no longer limited to a num-
ber of Electronic Control Units (ECUs) with low-bandwidth
networks as it used to be. More powerful devices are used
to collect and process the data coming from the sensors,
like heterogeneous computing platforms with multiple cores,
Graphical Processing Units (GPUs) and Field Programmable
Gate Arrays (FPGAs). In addition to the data generated by
the vehicle, external, data is also available from the internet,
other vehicles or infrastructure, what is known as Vehicle-
to-Anything communications (V2X). The hardware part also
includes the vehicle itself, this is the mobile platform and
actuators, which can be of different kinds depending on
the application and terrain where the system will operate.
The internal networking interfaces allow each subsystem to
exchange information with each other, for example high band-
width interfaces like USB 3.x or Gigabit Ethernet for sensor
data transport, or CAN and LIN networks for low bandwith
communication.
The processing capabilities of current self-driving vehicles
is such that they are sometimes they are referred to as
supercomputers on wheels. This statement is not far from
reality because due to the complexity in the hardware side,
the software side of vehicles has been also evolving, from the
embedded software running on top of Real-Time Operating
Systems (RTOS) on ECUs, to high-level software includ-
ing frameworks, libraries and modules that support Machine
Learning (ML), Artificial Intelligence (AI) and Deep Learning
(DL) algorithms required for processing the data. But also,
there are software components dealing with different aspects
of the vehicle operation, like drivers for data collection from
the sensors, user interfacing through the infotainment system,
and real-time and critical software for controlling actuators
and monitoring the status of the vehicle. This complexity
in software creates the need to follow certain patterns and
implement standards that enable a successful development,
management and deployment of such systems, both at a low-
level (hard real-time software/firmware) and at a high-level
(detection, inference or forecasting software). These software
frameworks and standards provide a structured way for the
software to operate in a concurrent and cooperative way.
One example of software guidelines and frameworks is
the AUTomotive Open System ARchitecture, AUTOSAR [4]
[5], widely used in the automotive industry. Its main goal
is to create and establish an open and standardized archi-
tecture for ECUs, using a component-based software design
approach. Another example is the Robot Operating System,
ROS [6], a well-established software framework providing
tools and libraries for robotic applications, including ready-
to-use implementations for perception, navigation and motion
control algorithms. However, as the robotics and self-driving
landscape has changed considerably since its introduction in
2009, a new version, ROS2 [7], was re-designed based on new
considerations in order to make it suitable for its use in a new
range of applications, like deterministic, real-time and safety-
critical systems [8]. In the case of software components as part
of an autonomous driving architecture, Autoware Foundation
offers its projects autoware.ai and autoware.auto [9]. They are
built on top of ROS and ROS2 respectively and offer software
modules for algorithms and tasks commonly used in self-
driving systems.
B. Functional View
From another perspective, autonomous vehicles are com-
posed of logical or functional blocks, which are defined based
on the flow of information and the processing stages performed
from data collection to the control of the vehicle, and including
the internal monitoring of the system. From this, four main
functional blocks can be identified across most of the proposed
architectures and solutions from literature in both academia
and industry: perception, planning and decision, motion and

vehicle control, and system supervision. These blocks are
represented in figure 2.
Perception
Planning
and
Decision
Motion and
Vehicle
Control
System Supervision
Sensors
External interaction:Maps,
rules, user IO, world status, etc.
Actuators
Fig. 2. Functional architecture for an autonomous driving system
The main goal of the perception stage is to receive data from
sensors and other sources (vehicle sensors configuration, maps
databases, etc), and generate a representation of the vehicle
status and a world model. For performing these two tasks,
sensors are categorized into proprioceptive and exteroceptive
sensors. Proprioceptive sensors are those used for sensing
the vehicle state, like Global Navigation Satellite Systems
(GNSS), Inertial Measurement Units (IMUs), Inertial Navi-
gation Systems (INS), and Encoders. These are used to get
position, movement and odometry information of the platform.
Exteroceptive sensors monitor the environment surrounding
the vehicles to obtain data of the terrain, the environment
and external objects. Cameras, lidars (LIght Detection And
Ranging), radar and ultrasonic sensors belong to this category.
After collecting all the incoming data from the sensors, two
main functions are performed in the perception stage: Locali-
sation and mapping, and object detection. More details on the
perception stage are covered in section III.
Once the vehicle and world status are available for the
planning and decision stage, the system is able to receive
external information like a goal or a travel mission, and then
can start the navigation plan for the vehicle, including a long-
term plan (going from place A to place B in a global map
or journey plan) as well as a short-term plan (execute partial
goals or waypoints considering a dynamic and local map).
Reactive behaviours are also included in this stage, mostly
intended for the safe operation of the vehicle. Autonomous
Emergency Braking (AEB) or collision preventing features
are common examples of these behaviours that will override
high-level commands. As most of the vehicles will interact
with other actors in the road, this stage should also include
external information from external sources to operate safely,
information like traffic rules, maps updates and speed limits
must be included to generate the plans.
The stage of Motion and Vehicle Control is related to
the way in which the trajectory generated in the previous
stage is executed on the platform, taking into account the
configuration, geometry and limitations of it. The movement
commands can be either goal points if the platform abstracts its
configuration and the control of actuators, or movement com-
mands like longitudinal speed, steering and braking. Again,
this stage is highly associated with safety features as it receives
high priority commands from the reactive behaviour modules
to modify or stop the movement of the vehicle.
Another stage within an autonomous driving architecture
is the System supervision which is in charge of monitoring
all the aspects of the vehicle like the hardware (sensors
operation, missing or degraded information, platform stability,
energy management, fault and diagnosis, etc) and the software
(autonomous driving software stack, data values within ranges,
data update rates, data inconsistencies, etc). The importance of
these tasks is that, as a safety-critical system, malfunctioning
of hardware or software in a self-driving vehicle, should not
result in any harm to people, environment or property. The
ISO 26262 [10] is a standard for functional safety in road
vehicles, adapted from the standard IEC 61508 [11], that aims
to address possible hazards arising from faults in electric or
electronic devices in a car, and offers a framework for the full
product development life cycle, from specification and design
to validation and production release. A further discussion on
functional safety of automotive software can be found in [12].
C. Related Work
Different description of architectures can be found in
literature both from a technical or a functional point of
view, but also with different scopes, for example, focused
on methodologies, design and concept, or focused to actual
implementation on a platform. For example, reference [13]
defines three different technical architectures references for
automotive systems focusing on the distribution of processing
and communication hardware components within a vehicle. On
the software side, reference [14] presents an approach for inter-
connecting two environments, ROS2 and Adaptive AUTOSAR
[15], frameworks that are based on a Data Distribution Service
(DDS) middleware and can cooperate in the whole self-driving
system.
Related also to technical architectures, there are works
detailing the actual implementation of self-driving systems,
both from industry and academia, like [16], which presents
a brief description of a software-system architecture from the
industry perspective, and in [17], the design and development
of a research platform for outdoor environments is presented,
which main focus is on the sensors and vehicle platform
components.
On the functional side, a functional reference architecture is
proposed by [18] and [19], describing three main components
(perception, decision and control and platform manipulation)
and two basic layers (cognitive driving intelligence and vehicle
platform) within it. Another architecture is presented in [20]
and its main design goal is to be compatible with safety-
critical applications and standards like ISO-26262. To this
end, a V-shaped, modular architecture with five layers of
abstraction, providing scalability and traceability is proposed.
Also, in [21] a whole framework is proposed for the design of
supervision functions for the components of an autonomous

driving system. The focus is on the methodological aspects
of the engineering process, starting from functional and safety
requirements, and generating a formal behaviour model for the
specified system.
Finally, there are other works with a wider scope, addressing
the technical and functional view of self-driving systems,
presenting full deployments [22] and comparing different
architectures [23], [24]. An example of a deployed system
is found in [22] where a self-driving car architecture solution
based on ROS2 is presented, addressing some limitations of
the previous version of ROS in terms of usability and main-
tainability. On the side of architectures comparison, reference
[23] presents an overview of four architectures of self-driving
solutions, from an industrial case to a resource-constrained
prototype implementation, detailing in the technical implemen-
tation of the hardware and software, and showing similarities
and differences of their functional components and design
decisions. Reference [24] presents a review on three functional
system architectures used in self-driving platforms, comparing
their construction approach and their robustness.
III. SEN SI NG A ND PERCEPTION
Perception in a self-driving architecture is the stage dealing
with information from sensors and turning it into meaningful
data for different high-level tasks like navigation and motion
planning. In this section perception will be covered in three
parts: sensors, localisation and mapping, and object detection.
A. Sensors
Perception has been of interest of the field of intelligent
and autonomous vehicles for more than 20 years [25], [26].
Initially, most of the developments were vision-based and
applied to both infrastructure [27] and vehicles [28]. In recent
years, further developments in sensors devices and processing
platforms, have made possible the inclusion Lidar and Radar
technologies into the suite of available sensors for self-driving
applications, providing the perception stage with more data
and allowing to take advantage of the strengths of each sensor
technology and overcome their weaknesses. However, despite
the advances in sensor technologies, there are still different
challenges that need to be addressed.
As stated in the previous section, sensors in self-driving
platforms are categorized into proprioceptive and exterocep-
tive. The former group provides the system with data regarding
the position and movement of the vehicle, in an absolute
reference system, like GNSS, or in a relative reference system
like IMUs and encoders. When two or more of these sensors
are used in conjunction, they can compensate or complement
each other under adverse conditions, like losing GNSS signal,
for example. Exteroceptive sensors enable the system to ”see”
its surrounding environment. Cameras, lidars, radars and ultra-
sonic sensors generate a lot of information in form of images
(2D) and point clouds (3D), from where the conditions and sta-
tus of other vehicles, people and terrain can be obtained. This
allows the vehicle to generate a representation of the external
world, and locate itself in it in order to generate navigation
plans and prevent safety-related issues. A more detailed review
of sensor technology and perception in autonomous vehicles
can be found in [29] and [30].
B. Localisation and Mapping
Localisation refers to the process of obtaining the relation
between the vehicle state (location) and its environment,
represented by a map. Two common approaches are used for
doing so: 1) localisation based on proprioceptive sensors and
2) Localisation using exteroceptive sensors. In the first case,
GNSS is used to provide an absolute location in a global
reference system or a global map, and IMU/encoders provide
relative position information, that can be used to complement
or correct GNSS data when it is degraded due the vehicle been
in an area where the sky is obstructed or signal is weak.
One approach used to provide additional information is
to generate odometry data from range sensors like cameras
and lidars, sometimes referred as visual-odometry [31]–[35]
or lidar-odometry [36]–[38]. In these solutions, features and
landmarks are detected and displacement is calculated from
the difference between frames.
When maps are not available or provided, localisation and
mapping are performed simultaneously. This technique is
known as SLAM (Simultaneous Localisation And Mapping),
where a map is constructed from camera or lidar data and
at the same time the location of the vehicle is estimated.
The main advantage of this method is that a prior map
is not needed and the system could localise itself in an
unknown environment. Reference [39] offers a description of
localisation techniques used in autonomous platforms intended
for warehouse operation and reference [40] presents a deeper
review on localisation techniques for mobile robots. Irrespec-
tive of the application environment or mobile platform, those
techniques and algorithms are shared with self-driving cars or
larger off-road vehicles.
C. Object detection
A self-driving vehicle needs to understand its environment
and be aware of all the objects and actors that could interact
with it. For example, the vehicle must detect the road and
the lanes within it, but it also should detect other vehicles,
pedestrians, obstacles and traffic signs. Or in the case of an
off-road vehicle, it also should detect types of terrain, drivable
areas, livestock, trees, weather conditions, etc. All this is part
of the detection stage. In addition to detection, the system
should be capable of tracking the detected objects in the
space domain in the time domain. The main objective is to
forecast possible incidents based on predicted movements of
other vehicles, people or obstacles. This generated information
can also be integrated into global and local maps to improve
the planning and navigation process.
Another function of detection is scene understanding which
determines if an autonomous vehicle is operating within its
ODD, for example in terms of the environmental conditions.
And this is one of the main challenges for outdoor and off-
road autonomous vehicles, to maintain their reliability across

different weather conditions and handle its impact on sensor
data integrity in a safe way. In [41], the authors propose a
multi-feature method to perform visual localisation across dif-
ferent seasons throughout the year, based on scene-matching.
Rain is another condition that has a big impact on cameras
and lidar sensors, in particular, when droplets remain on the
lenses. Reference [42] proposes a de-raining filtering method
based to overcome this issue and also provides a dataset of
rainy images and a method for creating synthetic water drops
on top of other datasets.
IV. DATA FUSION IN PERCEPTION
Data fusion, also referred as multi-sensor data fusion, infor-
mation fusion or sensor fusion, has received several definitions
from different authors in the literature [43], [44], [45] [46],
[47], but a better understanding of what it is, can be obtained
by answering the following questions:
•What is involved in data fusion?
Combine, merge or integrate homogeneous or heteroge-
neous data.
•What is the aim of data fusion?
Get a better representation of a process or the environ-
ment, infer underlying information, improve the quality
of the data.
•How to apply data fusion?
Data fusion is a multi-level task, depending of the nature
of the sensors, the context and the final application.
Thus, multi-sensor data fusion is a multidisciplinary field,
because the information in a typical process, flows from
sensors to applications, passing through stages of filtering,
data enhancement and information extraction. Because of
this, knowledge in a wide range of fields is required, e.g.
signal processing, machine learning, probability and statistics,
artificial intelligence, etc. [48].
The first step in multi-sensor data fusion is to collect and
associate the data in space and time, done by performing
the calibration and synchronisation of the sensors within the
system. This step is extremely important as the performance
of fusion stages rely on the data from different sources being
consistent and referenced to a common reference frame. Some
of the challenges on aggregating sensors are described by
[49] where a multi-sensor fusion of camera, radar and lidar is
applied to large off-road vehicles and the difficulties of time
synchronisation are highlighted. Also, reference [50] describes
a multi-sensing system for autonomous driving, outlining the
challenges of fusing heterogeneous data from sensors and the
benefits of adhering to a software architecture model based on
defined interfaces and components. Further examples of recent
developments in the calibration and synchronisation of sensors
(cameras, lidar and/or radar) can be found in [51], [52] and
[53].
As mentioned before, two types of localisation approaches
can be done, based on internal sensors, or using range sensors.
In the first case, the fusion of GNSS and IMU data is usually
performed using techniques based on Kalman Filtering [54],
[55]. This is a well-established method, and sometimes this
processing is already done in the sensor, as is the case for
some INSs, which embed a GNSS solution with an IMU into
a single device.
In the area of visual- and lidar-odometry, there are also
developments using both sensors to improve the performance
compared to a unimodal solution. In [56], a method is pro-
posed where visual-odometry is made at an initial stage, and
then lidar is fused to refine the estimation, making it possible
to use in indoor and outdoor conditions. Another approach
is found in [57] where two robust algorithms are coupled,
VISO2 for visual-odometry and LOAM for lidar- odometry. A
different approach is presented in [58], where they use a multi-
camera system composed of four fisheye cameras and generate
virtual lidar data for fusing it into the odometry process.
A variety of SLAM techniques using different configura-
tions of range sensors is found in the literature. There are
solutions based on single cameras, multiple cameras, stereo
cameras, depth cameras and 2D and 3D lidars. For example,
a multi-camera SLAM system is proposed in [59], where a
panoramic view is created by fusing the data from 5 cameras.
The performance of the system using a single camera approach
and a 3, 4 or 5 camera configuration is presented. Reference
[60] presents a review on multimodal localisation approaches
for mobile robots, evaluating Visual SLAM, Visual Odometry
and Place Recognition in terms of the requirements to adapt
to dynamic conditions and terrains, like unstructured, off-road
and outdoor environments. A further review focused only on
SLAM technologies is presented in [61].
All these range sensors are also integrated for the purposes
of object detection and tracking. Most of the recent work in
this area is based on Deep Learning and Deep networks tech-
niques. Reference [62] proposes a lidar-camera fusion system
based on Convolutional Neural Networks. They evaluate the
individual performance of each sensor in addition to the fused
approach. Also, in [63] is presented a Deep-Learning-based
architecture for the fusion of Radar and Camera for object
detection applications. Another recent work is presented in
[64]: a Camera-Radar fusion for object detection using Region
Proposal Networks (RPN) as a layer in a combination of
networks for fusing 2D and 3D data. On the other hand,
reference [65] presents an approach for fusing INS, camera
and Lidar data to perform 3D object tracking based on Kalman
filtering.
A further comprehensive review on multi-sensor fusion in
automated driving focusing on the fusion of heterogeneous
data from Camera, Lidar, Radar, Ultrasonics, GNSS, IMU and
V2X communications is presented in [66].
V. CONCLUSIONS
Over the last 10 years, several advances have been made
in different aspects of autonomous driving systems, from the
creation of reference architectures, standards, communities,
evolution of hardware and software, etc. However, in order
to achieve levels 4 and 5 of autonomy, as defined by SAE
standard J3016 [67], there are still different challenges that

must be solved, especially in the field of perception. Devel-
opments on new lidars, radars, stereo and depth cameras, and
the decreasing cost and size of these devices will allow the
inclusion of several sensors per kind, opening the possibility
of creating better representation of the world, but this presents
challenges in terms of processing, bandwidth, synchronisation
and data fusion.
In this regard, different data and information fusion tech-
niques are being developed presently with good results, but
further work must be done for taking them from isolated
developments and solutions to a safety-critical self-driving
platform where they should seamlessly integrate within a
defined architecture or framework without affecting other
subsystems. To do so, a proper methodology should be
followed throughout all the product life cycle development
stages: ODD definition, functional and safety requirements,
architecture design, software and hardware development, and
testing, verification, validation and product release.
Driving is a complex task even for humans, and we are good
at dealing with unexpected situations and at making sense of
multitude of multi-modal information while driving, allowing
us to make good decisions and move safely through the road.
However, we have to deal with stress and tiredness, and these
are risk factors in non-autonomous vehicles. Through im-
proved perception and data fusion developments, self-driving
vehicles should surpass our capabilities in the pursuit of better,
safer and greener transport.
ACK NOW LE DG EM EN T
This work was supported, in part, by Science Foundation
Ireland grant 13/RC/2094 and co-funded under the European
Regional Development Fund through the Southern & Eastern
Regional Operational Programme to Lero - the Science Foun-
dation Ireland Research Centre for Software (www.lero.ie)
This project has received funding from the European
Union’s Horizon 2020 research and innovation programme un-
der the Marie Skłodowska-Curie grant agreement No 754489
REFERENCES
[1] J. Levinson, J. Askeland, J. Becker, J. Dolson, D. Held, S. Kammel, J. Z.
Kolter, D. Langer, O. Pink, V. Pratt, M. Sokolsky, G. Stanek, D. Stavens,
A. Teichman, M. Werling, and S. Thrun, “Towards fully autonomous
driving: Systems and algorithms,” IEEE Intelligent Vehicles Symposium,
Proceedings, no. Iv, pp. 163–168, 2011.
[2] S. Kato, E. Takeuchi, Y. Ishiguro, Y. Ninomiya, K. Takeda, and
T. Hamada, “An Open Approach to Autonomous Vehicles,” IEEE Micro,
vol. 35, no. 6, pp. 60–68, nov 2015.
[3] P. Kruchten, “The 4+1 View Model of architecture,” IEEE Software,
vol. 12, no. 6, pp. 42–50, 1995.
[4] AUTOSAR, “Autosar standards,” https://www.autosar.org/standards,
Last accesed on 2020-07.
[5] M. Staron and D. Durisic, “AUTOSAR Standard,” in Automotive Soft-
ware Architectures. Cham: Springer International Publishing, 2017, pp.
81–116.
[6] M. Quigley, B. Gerkey, K. Conley, J. Faust, T. Foote, J. Leibs, E. Berger,
R. Wheeler, and A. Ng, “Ros: an open-source robot operating system,”
in Proc. of the IEEE Intl. Conf. on Robotics and Automation (ICRA)
Workshop on Open Source Robotics, Kobe, Japan, May 2009.
[7] D. Thomas, W. Woodall, and E. Fernandez, “Next-generation ROS:
Building on DDS,” in ROSCon Chicago 2014. Mountain View, CA:
Open Robotics, sep 2014.
[8] B. Gerkey, “Why ros 2?” https://design.ros2.org/articles/why ros2.html,
Last accesed on 2020-07.
[9] The Autoware Foundation, “The autoware foundation,” https://www.
autoware.org/, Last accesed on 2020-07.
[10] ISO, “ISO 26262 - Road vehicles – Functional safety,” 2011.
[11] IEC, “IEC 61508 - Functional safety of electri-
cal/electronic/programmable electronic safety-related systems,” 2005.
[12] M. Staron and P. Johannessen, “Functional Safety of Automotive
Software,” in Automotive Software Architectures. Cham: Springer
International Publishing, 2017, pp. 201–222.
[13] A. Bucaioni and P. Pelliccione, “Technical architectures for automotive
systems,” Proceedings - IEEE 17th International Conference on Soft-
ware Architecture, ICSA 2020, pp. 46–57, 2020.
[14] N. Parmar, V. Ranga, and B. Simhachalam Naidu, “Syntactic Interoper-
ability in Real-Time Systems, ROS 2, and Adaptive AUTOSAR Using
Data Distribution Services: An Approach,” 2020, pp. 257–274.
[15] AUTOSAR, “Autosar adaptive platform,” https://www.autosar.org/
standards/adaptive-platform/, Last accesed on 2020-07.
[16] S. Furst, “System/ Software Architecture for Autonomous Driving Sys-
tems,” in 2019 IEEE International Conference on Software Architecture
Companion (ICSA-C). IEEE, mar 2019, pp. 31–32.
[17] S. Kyberd, J. Attias, P. Get, P. Murcutt, C. Prahacs, M. Towlson, S. Venn,
A. Vasconcelos, M. Gadd, D. De Martini, and P. Newman, “The Hulk:
Design and Development of a Weather-proof Vehicle for Long-term
Autonomy in Outdoor Environments,” Tokyo, Japan, 2019, pp. 1–14.
[18] S. Behere and M. T¨
orngren, “A Functional Architecture for Autonomous
Driving,” in Proceedings of the First International Workshop on Auto-
motive Software Architecture - WASA ’15. New York, New York, USA:
ACM Press, 2015, pp. 3–10.
[19] ——, “A functional reference architecture for autonomous driving,”
Information and Software Technology, vol. 73, pp. 136–150, 2016.
[20] S. Akkaya, Y. Gurbuz, M. G. Zile, E. Baglayici, H. A. Seker, and
A. Erdogan, “A Modular Five-Layered V-Shaped Architecture for Au-
tonomous Vehicles,” ELECO 2019 - 11th International Conference on
Electrical and Electronics Engineering, pp. 850–854, 2019.
[21] R. Cuer, L. Pi´
etrac, E. Niel, S. Diallo, N. Minoiu-Enache, and C. Dang-
Van-Nhan, “A formal framework for the safe design of the Autonomous
Driving supervision,” Reliability Engineering and System Safety, vol.
174, no. February, pp. 29–40, 2018.
[22] M. Reke, D. Peter, J. Schulte-Tigges, S. Schiffer, A. Ferrein,
T. Walter, and D. Matheis, “A self-driving car architecture in
ROS2,” in 2020 International SAUPEC/RobMech/PRASA Conference,
SAUPEC/RobMech/PRASA 2020. IEEE, jan 2020, pp. 1–6.
[23] C. Berger and M. Dukaczewski, “Comparison of architectural design
decisions for resource-constrained self-driving cars-A multiple case-
study,” in Lecture Notes in Informatics (LNI), Proceedings - Series of
the Gesellschaft fur Informatik (GI), vol. P-232, 2014, pp. 2157–2168.
[24] O. S. Tas, F. Kuhnt, J. M. Zollner, and C. Stiller, “Functional system
architectures towards fully automated driving,” in 2016 IEEE Intelligent
Vehicles Symposium (IV), vol. 2016-Augus, no. Iv. IEEE, jun 2016,
pp. 304–309.
[25] E. Dickmanns, “The development of machine vision for road vehicles in
the last decade,” in Intelligent Vehicle Symposium, 2002. IEEE, vol. 1.
IEEE, 2002, pp. 268–281.
[26] U. Nunes, C. Laugier, and M. M. Trivedi, “Guest Editorial Introducing
Perception, Planning, and Navigation for Intelligent Vehicles,” IEEE
Transactions on Intelligent Transportation Systems, vol. 10, no. 3, pp.
375–379, sep 2009.
[27] A. Broggi, K. Ikeuchi, and C. E. Thorpe, “Special issue on vision ap-
plications and technology for intelligent vehicles: part I-infrastructure,”
IEEE Transactions on Intelligent Transportation Systems, vol. 1, no. 2,
pp. 69–71, jun 2000.
[28] ——, “Special issue on vision applications and technology for intelligent
vehicles: Part II - vehicles [Editorial],” IEEE Transactions on Intelligent
Transportation Systems, vol. 1, no. 3, pp. 133–134, sep 2000.
[29] S. Campbell, N. O’Mahony, L. Krpalcova, D. Riordan, J. Walsh,
A. Murphy, and C. Ryan, “Sensor Technology in Autonomous Vehicles
: A review,” in 29th Irish Signals and Systems Conference, ISSC 2018.
IEEE, jun 2018, pp. 1–4.
[30] J. Zhao, B. Liang, and Q. Chen, “The key technology toward the self-
driving car,” International Journal of Intelligent Unmanned Systems,
vol. 6, no. 1, pp. 2–20, 2018.
[31] B. Zhao, T. Hu, and L. Shen, “Visual odometry - A review of ap-
proaches,” 2015 IEEE International Conference on Information and

Automation, ICIA 2015 - In conjunction with 2015 IEEE International
Conference on Automation and Logistics, no. August, pp. 2569–2573,
2015.
[32] Q. Lin, X. Liu, and Z. Zhang, “Mobile Robot Self-LocalizationUsing
Visual Odometry Based on Ceiling Vision,” 2019 IEEE Symposium
Series on Computational Intelligence, SSCI 2019, pp. 1435–1439, 2019.
[33] K. S. Krishnan and F. Sahin, “ORBDeepOdometry - A feature-based
deep learning approach to monocular visual odometry,” 2019 14th
Annual Conference System of Systems Engineering, SoSE 2019, pp. 296–
301, 2019.
[34] M. Aladem and S. A. Rawashdeh, “A Combined Vision-Based Multiple
Object Tracking and Visual Odometry System,” IEEE Sensors Journal,
vol. 19, no. 23, pp. 11 714–11 720, 2019.
[35] H. Ragab, M. Elhabiby, S. Givigi, and A. Noureldin, “The Utilization of
DNN-based Semantic Segmentation for Improving Low-Cost Integrated
Stereo Visual Odometry in Challenging Urban Environments,” 2020
IEEE/ION Position, Location and Navigation Symposium, PLANS 2020,
pp. 960–966, 2020.
[36] B. Zhou, Z. Tang, K. Qian, F. Fang, and X. Ma, “A LiDAR Odometry
for Outdoor Mobile Robots Using NDT Based Scan Matching in GPS-
denied environments,” 2017 IEEE 7th Annual International Conference
on CYBER Technology in Automation, Control, and Intelligent Systems,
CYBER 2017, pp. 1230–1235, 2018.
[37] I. Hamieh, R. Myers, and T. Rahman, “Construction of Autonomous
Driving Maps employing LiDAR Odometry,” 2019 IEEE Canadian
Conference of Electrical and Computer Engineering, CCECE 2019, pp.
15–18, 2019.
[38] L. Qingqing, F. Yuhong, J. Pena Queralta, T. N. Gia, H. Tenhunen,
Z. Zou, and T. Westerlund, “Edge Computing for Mobile Robots: Multi-
Robot Feature-Based Lidar Odometry with FPGAs,” 2019 12th Inter-
national Conference on Mobile Computing and Ubiquitous Network,
ICMU 2019, pp. 54–55, 2019.
[39] C. Cronin, A. Conway, and J. Walsh, “State-of-the-art review of au-
tonomous intelligent vehicles (AIV) technologies for the automotive and
manufacturing industry,” 30th Irish Signals and Systems Conference,
ISSC 2019, pp. 1–6, 2019.
[40] S. Campbell, N. O’Mahony, A. Carvalho, L. Krpalkova, D. Riordan,
and J. Walsh, “Where am I? Localization techniques for Mobile Robots
A Review,” 2020 6th International Conference on Mechatronics and
Robotics Engineering, ICMRE 2020, pp. 43–47, 2020.
[41] Y. Qiao, C. Cappelle, and Y. Ruichek, “Visual Localization across Sea-
sons Using Sequence Matching Based on Multi-Feature Combination,”
Sensors, vol. 17, no. 11, p. 2442, oct 2017.
[42] H. Porav, T. Bruls, and P. Newman, “I Can See Clearly Now: Image
Restoration via De-Raining,” in 2019 International Conference on
Robotics and Automation (ICRA), vol. 2019-May. IEEE, may 2019,
pp. 7087–7093.
[43] F. E. White, “Data Fusion Lexicon,” The Data Fusion Subpanel of the
Joint Directors of Laboratories, Technical Panel for C3, vol. 15, no.
0704, p. 15, 1991.
[44] R. Luo, “Multisensor fusion and integration: approaches, applications,
and future research directions,” IEEE Sensors Journal, vol. 2, no. 2, pp.
107–119, apr 2002.
[45] R. C. Luo, C. C. Chang, and C. C. Lai, “Multisensor Fusion and
Integration: Theories, Applications, and its Perspectives,” IEEE Sensors
Journal, vol. 11, no. 12, pp. 3122–3138, dec 2011.
[46] W. Elmenreich, “A Review on System Architectures for Sensor Fusion
Applications,” in Software Technologies for Embedded and Ubiquitous
Systems, R. Obermaisser, Y. Nah, P. Puschner, and F. J. Rammig, Eds.
Santorini Islands, Greece: Springer, 2007, pp. 547–559.
[47] H. Bostr¨
om, S. Andler, and M. Brohede, “On the definition of infor-
mation fusion as a field of research,” University of Sk ¨
ovde, Tech. Rep.,
2007.
[48] G. Velasco-Hernandez, “Multisensor Architecture for an Intersection
Management System,” Universidad del Valle, Tech. Rep., 2019.
[49] D. J. Yeong, J. Barry, and J. Walsh, “A Review of Multi-Sensor Fusion
System for Large Heavy Vehicles Off Road in Industrial Environments,”
in ISSC, 2020.
[50] J. P. Giacalone, L. Bourgeois, and A. Ancora, “Challenges in aggregation
of heterogeneous sensors for Autonomous Driving Systems,” SAS 2019
- 2019 IEEE Sensors Applications Symposium, Conference Proceedings,
pp. 3–7, 2019.
[51] H. Hu, J. Wu, and Z. Xiong, “A soft time synchronization framework for
multi-sensors in autonomous localization and navigation,” IEEE/ASME
International Conference on Advanced Intelligent Mechatronics, AIM,
vol. 2018-July, pp. 694–699, 2018.
[52] J. Domhof, K. F. Julian, and K. M. Dariu, “An extrinsic calibration
tool for radar, camera and lidar,” Proceedings - IEEE International
Conference on Robotics and Automation, vol. 2019-May, pp. 8107–8113,
2019.
[53] L. Yang and R. Wang, “HydraView : A Synchronized 360 -View of
Multiple Sensors for Autonomous Vehicles,” pp. 53–61, 2020.
[54] S. Panzieri, F. Pascucci, and G. Ulivi, “An outdoor navigation system
using GPS and inertial platform,” IEEE/ASME Transactions on Mecha-
tronics, vol. 7, no. 2, pp. 134–142, 2002.
[55] Wahyudi, M. S. Listiyana, Sudjadi, and Ngatelan, “Tracking Object
based on GPS and IMU Sensor,” in 2018 5th International Confer-
ence on Information Technology, Computer, and Electrical Engineering
(ICITACEE). IEEE, sep 2018, pp. 214–218.
[56] J. Zhang and S. Singh, “Visual-lidar odometry and mapping: Low-
drift, robust, and fast,” Proceedings - IEEE International Conference
on Robotics and Automation, vol. 2015-June, no. June, pp. 2174–2181,
2015.
[57] M. Yan, J. Wang, J. Li, and C. Zhang, “Loose coupling visual-
lidar odometry by combining VISO2 and LOAM,” Chinese Control
Conference, CCC, pp. 6841–6846, 2017.
[58] Z. Xiang, J. Yu, J. Li, and J. Su, “ViLiVO: Virtual LiDAR-Visual
Odometry for an Autonomous Vehicle with a Multi-Camera System,”
IEEE International Conference on Intelligent Robots and Systems, pp.
2486–2492, 2019.
[59] Y. Yang, D. Tang, D. Wang, W. Song, J. Wang, and M. Fu, “Multi-
camera visual SLAM for off-road navigation,” Robotics and Autonomous
Systems, vol. 128, p. 103505, 2020.
[60] N. O’Mahony, S. Campbell, A. Carvalho, S. Harapanahalli, G. A.
Velasco-Hernandez, D. Riordan, and J. Walsh, “Adaptive multimodal
localisation techniques for mobile robots in unstructured environments
:A review,” in IEEE 5th World Forum on Internet of Things, WF-IoT
2019 - Conference Proceedings. IEEE, apr 2019, pp. 799–804.
[61] A. Singandhupe and H. La, “A Review of SLAM Techniques and
Security in Autonomous Driving,” Proceedings - 3rd IEEE International
Conference on Robotic Computing, IRC 2019, no. 19, pp. 602–607,
2019.
[62] G. Melotti, C. Premebida, and N. Goncalves, “Multimodal deep-learning
for object recognition combining camera and LIDAR data,” 2020 IEEE
International Conference on Autonomous Robot Systems and Competi-
tions, ICARSC 2020, no. April, pp. 177–182, 2020.
[63] F. Nobis, M. Geisslinger, M. Weber, J. Betz, and M. Lienkamp, “A
Deep Learning-based Radar and Camera Sensor Fusion Architecture for
Object Detection,” 2019 Symposium on Sensor Data Fusion: Trends,
Solutions, Applications, SDF 2019, 2019.
[64] Z. T. Li, M. Yan, W. Jiang, and P. Xu, “Vehicle object detection based on
rgb-camera and radar sensor fusion,” Proceedings - International Joint
Conference on Information, Media, and Engineering, IJCIME 2019, pp.
164–169, 2019.
[65] A. Asvadi, P. Gir˜
ao, P. Peixoto, and U. Nunes, “3D object tracking using
RGB and LIDAR data,” IEEE Conference on Intelligent Transportation
Systems, Proceedings, ITSC, pp. 1255–1260, 2016.
[66] Z. Wang, Y. Wu, and Q. Niu, “Multi-Sensor Fusion in Automated
Driving: A Survey,” IEEE Access, vol. 8, pp. 2847–2868, 2020.
[67] SAE, “J3016 - Taxonomy and Definitions for Terms Related to Driving
Automation Systems for On-Road Motor Vehicles,” 2018.