27th Telecommunications forum TELFOR 2019 Serbia, Belgrade, November 26-27, 2019.
978-1-7281-4790-1/19/$31.00 ©2019 IEEE
Abstract — Advancing to higher levels of driving
automation brings unpredicted challenges and with them,
many situations that cannot be foreseen. In order to overcome
these problems, set of functionalities in modern vehicle is
growing in terms of algorithmic complexity and required
hardware. Risk of testing implemented solutions in real world
is high, expensive and time consuming. That is why virtual
simulation tools for automotive testing are heavily acclaimed.
Original Equipment Manufacturers (OEMs) use these tools to
create closed sense, compute, act loop. Production software is
tested against simulated sensing data and simulated action
consequences are generated according to the given software
commands. This gives OEMs ability to optimize design of the
vehicles before any physical prototypes are produced. Early
optimization brings reduced costs and less time delays. This
paper presents development of simple C++ perception
applications using ROS as a prototyping platform that are
validated and tested with “Software-In-the-Loop” (SIL)
methods. Simulations are created using CARLA simulator to
provide data and transform commands given by the
autonomous platform into simulated actions. Validation is
done by connecting Autoware autonomous platform with
CARLA simulator in order to test against various scenes in
which applications are applicable.
Keywords — Autonomous driving, perception, ROS,
CARLA, AUTOWARE, SIL, ADAS, C++, Python.
I. I
NTRODUCTION
evelopment of autonomous vehicles is a major trend
in automotive industry. Pushing towards SAE levels
[1] four and five and fully automated vehicle as an ultimate
product, engineers are facing issues that have never been
addressed. As they progress with development of some
functionality new problems arise because of uncertainty of
physical world, as there are many unpredicted situations
that could cause accidents. Due to this, development of
virtual simulators to test vehicle’s cognitive computing [2]
becomes crucial part of the development. With this
Stevan Stević is with the RT-RK Institute for Computer Based
Systems, Novi Sad, Serbia (e-mail: stevan.stevic@rt-rk.com).
Momčilo Krunić is with the Faculty of Technical Sciences,
University of Novi Sad, Serbia, Novi Sad, Serbia (e-mail:
momcilo.krunic@rt-rk.com).
Marko Dragojević is with the RT-RK Institute for Computer Based
Systems, Novi Sad, Serbia (e-mail: marko.dragojevic@rt-rk.com).
Nives Kaprocki is with the RT-RK Institute for Computer Based
Systems, Novi Sad and Faculty of Technical Sciences, University of
Novi Sad, Serbia (e-mail: @rt-rk.com).
approach the perception module [3] receives input from
computer-generated scenes and mathematically modelled
movement patterns for pedestrians, bicycles, and other
entities. Acting module [3] on the other side outputs
commands to simulators that implement these as actions.
Using this, billions of kilometres that are required [4] to
demonstrate reliability of autonomous vehicles in terms of
fatalities and injuries, have been already simulated by
OEMs [5]. By using simulator’s abstract visualizations,
engineers can focus on development of core capabilities
for autonomous driving, such as: driving models and
systems, remote assistance, mapping, localization,
perception etc.
This paper emphasizes and explains the importance of
using simulators in the modern automotive development
and gives practical example by showing how simple
advanced driver-assistance system (ADAS) applications
can be tested within the context. The rest of the material is
organized as follows. First part explains current state of the
industry, problems and practices used. Also, provides
some academical and industry related as a background.
Second section explains platforms and tools that are for
development and simulations. Section IV describes the
purpose of test applications and presents existing setup for
connecting Autoware [6] with CARLA and describes
sensor and start up file configuration. Section V presents
validation for these use-cases and final section concludes
the paper with the review on work done and some future
steps.
II. S
IMULATORS
Simulators use different models of environments that
can be built from high resolution LiDARs, cameras or even
virtual maps that provide annotations with tools like
OpenDRIVE [7]. Furthermore, some types of simulators
can augment existing data like point cloud to create
obstacles [8]. Based on this, modern day simulators like
rFpro [9], LGSVL [10], AVS [11] by Uber, CARLA [12],
DRIVE CONSTELLATION Simulator [13] by NVIDIA
and others are able to create a very realist scenes and
complex layouts like road paint or road separation that can
be difficult to discern even for humans. On top of that,
these ecosystems are scalable and can implement different
feature requests, unlike early development tools of this
type that stretch their capabilities to satisfy different use-
cases.
Simulators usually offer simulation of wide variety of
Stevan Stević, Momčilo Krunić, Marko Dragojević and Nives Kaprocki, Members, IEEE
Development and validation of ADAS
perception application
in ROS environment
integrated with CARLA simulator
Authorized licensed use limited to: ROBERT BOSCH. Downloaded on June 18,2020 at 11:02:25 UTC from IEEE Xplore. Restrictions apply.