Content uploaded by Ricardo Suarez-Bertoa
Author content
All content in this area was uploaded by Ricardo Suarez-Bertoa on Aug 26, 2022
Content may be subject to copyright.
Accident Analysis and Prevention 174 (2022) 106743
Available online 11 June 2022
0001-4575/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Driver models for the denition of safety requirements of automated
vehicles in international regulations. Application to motorway
driving conditions
Konstantinos Mattas , Giovanni Albano , Riccardo Don`
a , Maria Christina Galassi ,
Ricardo Suarez-Bertoa , Sandor Vass , Biagio Ciuffo
*
European Commission Joint Research Centre, Ispra, Ispra, Italy
ARTICLE INFO
Keywords:
Connected and automated vehicles
Driver model
Safety requirements
Fuzzy logic
ABSTRACT
UN Regulation 157, the rst global regulation regarding the type-approval of Automated Driving Systems (ADS),
has been adopted in 2021. In it, safety performance requirements are being dened for vehicles of automation
Level 3, according to the SAE J3016, with a limited Operational Design Domain (ODD). In particular, for three
types of events that are related to motorway driving, two models are provided to distinguish between pre-
ventable trafc scenarios, for which the ADS is expected to avoid an accident, and unpreventable trafc sce-
narios, for which accidents cannot be avoided and the ADS can only mitigate their severity. The models recreate
the short-term behavior of a driver who reacts to an emergency. Two possible actions are predicted: either no
reaction or full braking when danger is identied. In the present paper the two models are analyzed and
compared with two additional models: an industry proposed model, the Responsibility Sensitive Safety frame-
work (RSS), and the Fuzzy Safety Model (FSM) proposed by the authors. As in the case of the two regulation
models, also the RSS, although more sophisticated, assumes that the possible reaction by the driver is binary.
This approach neglects the ability of a human driver to drive defensively and anticipate possible risks. Defensive
drivers, indeed, may use comfortable decelerations in anticipation, to avoid nding themselves in an emergency
situation. The FSM uses fuzzy logic to mimic this behavior.
Results show that anticipation plays a very important role to reduce the number of unpreventable trafc
scenarios. In addition, by validating the classication capabilities of the four models with real trafc data, the
FSM proved to be the most suitable of the investigated models. On the basis of these results, the FSM has been
included in the proposal for amending UN Regulation 157, thus allowing to set higher safety standards for the
rst automated vehicles that will be introduced into the market.
1. Introduction
In 2021, the rst global regulation on the approval of automated
driving systems (ADSs) has been adopted (UNECE, 2021). The UN
Regulation 157 concerns the approval of a system able to drive without
the strict supervision of a human driver in a very limited operational
design domain (ODD), namely on motorways up to a maximum speed of
60 km/h (given lane changing is not possible, the system is also referred
to as the Automated Lane Keeping System, ALKS). Since a driver is al-
ways required to be present to take-over the driving task, this is a Level 3
automated system according to the SAE classication (SAE, 2021). The
limited and very well-dened scope of the Regulation helped to shape a
legislation on the approval of ADSs in a reasonable timeframe and
allowed to understand the issues to be tackled in its extension to wider
and more complex scopes. It can be therefore considered a crucial
milestone in the future legislative development in this eld.
One of the most challenging issues the regulators had to address was
the denition of the Level of safety the new vehicles have to guarantee
and the way to prove it by the ADS developer. In principle, ADSs should
i) respect trafc rules, ii) be able not to cause any accident, and, to the
extent possible, iii) prevent accidents caused by other road users. It is
easy to understand that turning this principles into quantitative re-
quirements that can be proven by ADS developers during the type-
approval process is not a simple task (Kalra and Paddock, 2016).
* Corresponding author.
E-mail address: biagio.ciuffo@ec.europa.eu (B. Ciuffo).
Contents lists available at ScienceDirect
Accident Analysis and Prevention
journal homepage: www.elsevier.com/locate/aap
https://doi.org/10.1016/j.aap.2022.106743
Received 5 February 2022; Received in revised form 11 May 2022; Accepted 4 June 2022
Accident Analysis and Prevention 174 (2022) 106743
2
In general, four different approaches could be followed by the
regulator (own elaborations based on Blumenthal et al., 2020):
- a rst approach would set an overall safety target (in terms of e.g.
accident probability, number of injured people, number of deaths,
etc.) in the ADS’ ODD and leave the ADS developer to provide
convincing evidence that this is achieved;
- a second approach would set a safety target on a xed number of
trafc scenarios and witness that the ADS is able to achieve it on the
basis of test results in all of them (in which the results are weighted
by the probability of occurrence of each scenario);
- a third approach would be not to set any safety target but to set some
performance requirement allowing to dene the trafc scenarios that
an ADS should be able to safely handle without incurring into an
accident; and nally
- a fourth approach setting operational requirements (namely re-
quirements on the way the ADS should operate in specic circum-
stances) can also be considered.
The different approaches have all advantages and disadvantages. The
rst is, for example, the desirable approach for policy makers who can
directly turn the policy objectives on road safety into a requirement for
ADSs, but it is also very difcult to prove its fulllment by an ADS
developer. The third, on the contrary, is the desirable approach for ADS
developers and Approval Authorities. It indeed clearly sets the vehicle
design target for the formers and it can be assessed by the latter in a
fairly straightforward way. However, this approach does not allow to
directly understand the impact of ADS introduction on road safety as the
performance requirements say little on e.g. the number of accidents that
by respecting them an ADS is able to avoid compared to a human driver.
Consequently, if these targets are not properly set or are set in a
completely arbitrary way, the regulation may turn into a “procrustean
bed
1
” with potentially harmful side effects. The risk of falling into a
procrustean bed is even higher with the fourth approach. By arbitrarily
limiting the design space of ADS manufacturers, indeed, there is a high
risk to produce sub-optimal driving logics with possible negative effects
on the safety of ADSs and on other dimensions such as, for example, the
trafc ow. For these cases, it is therefore crucial that the requirements
are based on strong empirical evidence and try to be as less arbitrary as
possible.
UN Regulation 157 mainly adopts the third approach, without an
overall safety target and with a series of performance requirements
(complemented by a limited number of operational requirements on the
need to respect trafc rules and on the minimum time headway to
maintain from the vehicle in front). In the attempt to be as less arbitrary
as possible, these requirements are based on physical considerations,
such as the minimum braking distance of a vehicle, the minimum re-
action time of their sensing systems, a quantitative denition of a
competent and careful human driver, all to be used as a reference for the
performances of future ADSs. This approach was in line with what had
been used in the past for systems with lower levels of automation and
was justied by the limited scope of this rst regulation, which, in the
near future, would therefore not be able to produce harmful effects
neither to the safety nor to the efciency of trafc ow.
As new regulations are currently being developed at UN and national
levels it is however important to have a closer look at the implications of
the aforementioned requirements and, in particular, if other approaches
than those currently used could contribute to make them more robust.
The attractiveness of using performance requirements is indeed
undeniable, as they make the type-approval process transparent and
easily veriable. However, they are not easy to be properly set, espe-
cially when the ADS complexity increases. Therefore, it is important to
understand whether they are a suitable tool or not.
The present paper tries to contribute to answering this question. It
focuses on the type of performance requirements dened in UN Regu-
lation 157. They are analytical models acting as classiers, as they are
used to dene whether a specic trafc scenario can be considered
preventable or not by the ADS. Such requirements are compared to other
frameworks recently proposed in the literature such as the Re-
sponsibility Sensitive Safety (RSS) model, an industry-developed
framework, and the Fuzzy Safety Model (FSM) proposed by the au-
thors. Fuzzy logic has been introduced by Zadeh (Zadeh, 1965), and has
been widely used in multiple disciples. In a previous work, the use of
fuzzy logic to evaluate car-following safety levels has been investigated
(Mattas et al., 2020). The relevant fuzzy safety metrics are exploited in
this work to design the FSM, a driver reaction model. In this way, a
scheme to extract performance requirements out of safety metrics is
presented. All the models are then compared and validated against
human driver data from the openly available highD dataset (Krajewski
et al., 2018), to understand whether they are suitable to fulll their task.
It is shown that, while the current performance requirements in UN
Regulation 157 are a step forward in establishing reasonable and
transparent rules, in some cases they neglect the ability of drivers to
prevent an accident by driving defensively. FSM is shown to be capable
of reproducing this capability by using mild actions earlier and to avoid
the need for harsh and imminent reactions later.
The paper is organized as follows. Section 2 introduces the relevant
literature in the eld. Section 3 includes the details of the four different
models. The methodology used for the assessment is presented in section
4. Finally, results and conclusions are discussed in sections 5 and 6
respectively.
2. Related work
Connected and automated mobility is considered an important tool
to achieve different policy objectives. In Europe, for example, they are
expected to support the transition to a safer, more efcient, and more
sustainable transport system (Marques dos Santos et al., 2022). Ensuring
that automated vehicles are deployed with sufciently ambitious safety
requirements should be therefore a policy assumption. Furthermore,
achieving high safety standards is very important for the public accep-
tance of automated vehicles, even regarding automated public transport
(Kassens-Noor et al., 2020). Researches on users’ acceptance show that
the perceived safety and trust are among the most relevant factors
affecting Automated Vehicles (AVs) acceptance (Xu et al., 2018; Zhang
et al., 2019), and that the experience with AVs can increase trust.
Different large-scale surveys highlighted that people valued very highly
being able to take over control of an AV (European Commission, 2020;
Nordhoff et al., 2018), and that this willingness to take over control of an
AV depends on specic socio-demographic characteristics (Grosso et al.,
2021b). The respondents desire a level of control over the vehicle
behavior, arguably because of the lack of trust. Similarly, passengers
experiencing a ride with an automated shuttle would often use an
emergency button inside the vehicle, just to see how the vehicle would
react (Nordhoff et al., 2020). Therefore, safety and security should be
clearly and transparently ensured, required by regulation (Fagnant and
Kockelman, 2015) and supported by insurance schemes (Baumann et al.,
2019; Grosso et al., 2021a).
Policies on AVs can help accelerate the development, reduce control
uncertainties, and smoothen the migration from conventional to AV
trafc (Li et al., 2019). Some of the rst relevant regulations for AV
targeted on-road testing, with different countries having very distinctive
approaches (Lee and Hess, 2020). Moreover, on-road testing alone may
not be sufcient to prove the increased safety of ADS. For human drivers,
a fatal crash happens with a return period of 3.4 million hours of driving,
1
Procrustes was a mythological bandit, who would put his captives on an
iron bed that had the “ideal” height for a human. He would harm them in the
effort of tting them to that ideal height. While he has been slain by the
mythological hero Theseus, the “procrustean bed” lives on, as an expression to
describe arbitrary standards that bring harmful side effects.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
3
which corresponds to a vehicle driving for 380 years straight (Shladover
and Nowakowski, 2019). Karla and Paddock have evaluated that, for
demonstrating that AVs are safer than human drivers, a eet of 100 AVs
should be tested for 24 h a day, for about 225 years (Kalra and Paddock,
2016). Changes in software or hardware could require the repetition of
the test.
In this light, novel ways to evaluate and assure safety by setting
proper requirements are required, and Trafc Conict Techniques
(TCTs) can be useful. TCTs have been developed to assess safety by
identifying conicts (Laureshyn et al., 2010), that are dangerous situa-
tions with much higher frequency than that of accidents. Moreover, the
frequency of conicts is correlated with the frequency of accidents for
human drivers (Tarko, 2018). However, this correlation may change
with the introduction of AVs, as different kinds of accidents may occur
and the human driver behavior may change (Wang et al., 2021). One of
the most widely used safety metrics of TCT is the Time-To-Collision
(TTC) (Laureshyn et al., 2016; Mahmud et al., 2019; Wang and Stama-
tiadis, 2014). Using TTC, different trafc states can be identied in the
famous NGSIM database (Kovvali et al., 2007), with the states shown to
correspond to different levels of risk (Zhao et al., 2018). The potential
risk for human drivers in dangerous car following situations can be
assessed (Liu et al., 2019). TTC has been shown to act complementary to
the calculation of minimum time headway, as the two values are suitable
for different purposes (Vogel, 2003). Moreover, TTC together with time
headway and nal relative distance under braking, have been used to
create Fuzzy rules, based on the Virginia 100-car database (Xiong et al.,
2019). Different models have been developed to evaluate risks in situ-
ations that include lane-changing maneuvers, based on real data (Park
et al., 2018; Xie et al., 2019).
TTC-based models have been used in safety evaluations for active
safety of Advanced Driver Assistance Systems (ADAS) using experi-
mental data, as in the work of Seiniger et al. (Seiniger et al., 2013).
However, in contrast with ADAS systems which should intervene at the
last moment in an emergency, AVs are responsible to avoid emergencies,
when possible, by driving defensively. Those decisions are made on a
tactical level, hence the term “Tactical Safety” has emerged (Schoener,
2020). This is an important difference in the responsibility and the
performance evaluation of different levels of vehicle automation.
More elaborate metrics can consider estimating the stopping distance
(Brunson et al., 2002; Doi et al., 1994), or even trajectory optimization
techniques (Junietz et al., 2018). Additionally, model predictive safety
metrics have been developed (Weng et al., 2020). The notion of safety
force elds has also been introduced, which can be useful for pre-
collision warning systems (Wang et al., 2016) and path planning
(Rasekhipour et al., 2017; Wolf and Burdick, 2008). Safety eld metrics
can also carry information on the potential accident severity (Mullakkal-
Babu et al., 2017). A safety force eld algorithm has been suggested and
veried through simulation by NVIDIA (NVIDIA, 2019). Moreover, the
concept of a safety eld has been used to model human like driving
behavior in non-emergency conditions (Kolekar et al., 2020), or coupled
with a behavior prediction model for the surrounding trafc, to over-
come some disadvantages of traditional safety metrics (Wang et al.,
2022). Such prediction schemes can be also based on naturalistic data,
and be useful for the design or the test of an ADS (Feng et al., 2021), but
for the design of regulation requirements, it may be desirable to avoid
them. Safety relevant choices can be made not only on the tactical level
but also on the strategic level, by planning a trip that avoids the routes
with locations of high-risk (Ryan et al., 2020).
All the aforementioned papers provide signicant contributions for
accident analysis or controller design but cannot be directly used in the
development of regulation requirements, which need additional infor-
mation. A performance requirement, indeed, provides the set of condi-
tions in which a collision shall be avoided or mitigated by the ADS.
Therefore, the denition of sets of possible events is necessary. Those
could be derived by observations in real data, by extracting correlations
(Thal et al., 2020), or sample high-risk situations (Akagi et al., 2019).
For the upcoming Level 3 ALKS, the riskiest cases will arguably be
receiving a reckless cut-in by a slower vehicle, and a formulation to
classify preventable and unpreventable cases is already in the text of the
regulation [paragraph 5.2.5.2,(UNECE, 2021)]. In the same regulation,
in Appendix 3, there is a “human driver” model, dened for cut-in, cut-
out, and car-following scenarios. The model, which will be described in
detail in the following sections, is proposed to simulate the behavior of a
Competent and Careful (CC) human driver. The rationale is that the
ALKS should guarantee at least the same level of safety of a careful and
competent human driver, and therefore that any accident avoidable by
that type of human driver, shall be also avoided by the ADS. This rst CC
human driver model is a major step towards a viable regulation
framework, establishing a way to form performance requirements using
simulation and assuming an intuitive driving style (VMAD, 2019a).
Moreover, the developers of the CC human driver model provided esti-
mations of the value of its parameters on the basis of extensive empirical
observations (VMAD, 2019b).
Another model that could be used in safety regulation is the Re-
sponsibility Sensitive Safety (RSS) model, developed by Intel/Mobileye
(Shalev-Shwartz et al., 2017). The model provides rules that could be
superimposed to any AV control strategy, to make sure that the AV
would not be responsible to cause an accident. The longitudinal safety
control part of the model is similar to a formally veried control strategy
(Loos et al., 2011). Even though RSS represents an operational
requirement, it may be used to dene the sets of preventable and
unpreventable conditions through simulation, providing a performance
requirement similarly as in the CC human driver model. RSS includes
some parameters and an effort has been carried out to calibrate and
evaluate the model based on two hundred cut-in events from the
Shanghai Naturalistic Driving Study data (Liu et al., 2021). In the same
work, it has been observed that human drivers would predict danger and
adopt preventive strategies, which is not seen in the formulation of the
CC driver model.
This type of performance requirement is suitable for the next de-
velopments in ADS regulation. Until the use cases considered by the
regulation remain sufciently simple and relatively easy to be described,
it is indeed possible to introduce a set of conditions where collisions
should be avoided in the riskiest trafc situations. However, “Tactical
Safety” should be explicitly taken into account when using as perfor-
mance reference the concept of a careful and competent human driver,
though. Furthermore, apart from classifying whether a certain scenario
would lead to a collision or not, the model should also be able to assess
how challenging it is to avoid a collision, in order to offer a support tool
in the choice of the scenario(s) to be tested. Therefore, such a model
should classify preventable and unpreventable cases in a binary way,
while also indicating a non-binary value of risk or severity. In this light,
a novel model is presented here, based on Fuzzy Surrogate Safety Met-
rics (SSMs). The Fuzzy SSMs have been recently developed and veried
using real data (Mattas et al., 2020; 2019). The model can be used to set
performance requirements, similarly to the CC human driver model in
the existing regulation. At the same time, on the basis of the fuzzy
membership value, it also embodies a direct way to dene a level of
difculty in avoiding the collision.
3. Models
Four models are presented and compared in the paper, the two that
are already part of Regulation 157, the RSS model proposed by Intel/
Mobileye, and the Fuzzy Safety Model (FSM), originally proposed in this
work. The model described in paragraph 5.2.5.2. Regulation 157 is
referred to as Reg157 model, while the model in Appendix 3 of the same
regulation will be referred hereinafter as the CC human driver model.
3.1. Reg157 model
The model, described in paragraph 5.2.5.2. refers to collision
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
4
avoidance when another vehicle is cutting-in. If the cutting-in vehicle
has a lower speed than the ALKS vehicle, and the lateral movement of
the cutting-in vehicle has been visible for at least 0.72 sec, a collision
should be avoided when the following equation (1) holds:
TTC >vrel2×6ms2+0.35s(1)
where TTC is the time to collision between the two vehicles and vrel is
their relative velocity. Equation (1) is evaluated when the front wheel of
the cutting-in vehicle is closer than 0.3 m to the outside edge of the
visible lane markings. The assumption is that the maximum deceleration
should be at least 6 m/s
2
. Moreover, the perception time, together with
the time needed to achieve the deceleration of 6 m/s
2
is equal to 0.35 s
for a vehicle with high-level automated functionalities.
A simulation model is developed, based on this behavior so that the
Reg157 for cut-in safety can be directly compared to the rest. In each
simulation step, the ego vehicle is evaluating the lateral distance to the
cutting-in vehicle. If the distance is unsafe, so the cutting-in vehicle’s
edge is 0.3 m inside the ego vehicle’s lane, the TTC is calculated on the
longitudinal dimension. If the TTC is large enough, and equation (1) is
satised, an abrupt deceleration is not needed, as the situation is safe.
On the other hand, if the TTC is small, the ego vehicle decelerates, with
6 m/s
2
, 0.35 s after the danger is identied. During this reaction time,
the ego vehicle speed is assumed to be constant.
3.2. CC human driver model
In Appendix 3 of Regulation 157, a simulation model is presented,
which assumes to mimic the behavior of a Competent and Careful
human driver for three different critical situations: a) a vehicle cutting-
in; b) a deceleration scenario in which the vehicle preceding the ego
vehicle suddenly decelerates; c) a cut-out scenario in which the vehicle
preceding the ego vehicle suddenly exits the lane, to reveal an obstacle.
For the cut-in scenario, the wandering distance of a vehicle inside its
lane is dened to be 0.375 m, estimated from real data (JAMA, 2020).
When a vehicle exits the wandering zone, it is assumed that it is about to
initiate a lane change. An additional perception distance of 0.72 m is
assumed. Once the vehicle in the adjacent lane is further than 0.72 m
from the wandering zone, the CC human driver perceives the possible
risk. The reaction time of the driver, estimated to be 0.75 s, starts at this
point. During this time, the human driver moves his/her foot from the
acceleration pedal. A small deceleration of 0.4 m/s
2
is assumed for this
period. After that, the maximum deceleration is estimated to be 0.774 g
(7.59 m/s
2
), and it takes 0.6 s to be realized. During those 0.6 s, the
deceleration is increasing linearly, in absolute value, corresponding to a
maximum absolute jerk value of 12.65 m/s
3
. An additional constrain is
that the CC driver would not start the emergency deceleration if the TTC
is more than 2 s, as the situation is not considered to be an emergency,
and a hard deceleration would not be necessary.
The formulation of the cut-out scenario is similar. The CC human
driver perceives the cut-out when the preceding vehicle exceeds the
lateral wandering distance. The deceleration dynamics are the same. For
the simulation of the deceleration scenarios, the perception and reaction
time start when the preceding vehicle starts decelerating. For both those
cases, the ego vehicle maintains a 2 s time headway with the preceding
vehicle, before the preceding vehicle changes its trajectory.
3.3. RSS model
The Responsibility Sensitive Safety model, proposed by Intel/Mobi-
leye is a complete framework that can be used for a wide range of trafc
situations. The main idea behind the framework is that the AV could not
ensure absolute safety, but certain rules could ensure that the AV would
not cause an accident. RSS is a rule-based system, to be combined with
any controller. However, in this work, an RSS abiding driver model is
simulated, to classify preventable and unpreventable scenarios, using
simulation.
Only the paragraphs relevant to multi-lane freeways are considered
in the present work. Two safety distances are introduced, the longitu-
dinal and lateral safety distance, as shown in equations (2)–(3).
dlon =ur
ρ
+1
2amax,accel
ρ
2+ur+
ρ
amax,accel2
2amin,brake
−u2
f
2amax,brake
(2)
where dlon is the minimum safe longitudinal distance, ur and uf are the
speeds of the ego vehicle (rear), and the preceding vehicle (front)
respectively,
ρ
is the reaction time of the ego vehicle, amax,accel is the
maximum acceleration of the ego vehicle, amin,brake is an estimation of the
ego vehicle’s maximum deceleration, and amax,brake and estimation of the
preceding vehicle’s maximum deceleration, with the constrain that
amin,brake is always smaller in absolute value than amax,brake, dlat is the
minimum safe lateral distance,
μ
a lateral safety distance margin, u1 and
u2 the lateral speeds of the ego and preceding vehicle, alateralmax,accel and
alateralmin,brake,correct maximum and minimum absolute values of lateral
acceleration.
If at least one of the safety distances is respected, the situation is safe.
A reaction is required by the ego vehicle only in case both safety dis-
tances are smaller than the relevant safe distance. The proper reaction
depends on which was the latest safety distance that was respected. If it
was the longitudinal distance, the proper reaction is a deceleration with
at least amin,brake after a time equal to
ρ
. On the other hand, if it was the
lateral distance, a possibility for a dangerous cut-in is recognized and the
proper reaction is to decelerate on the lateral direction if the ego vehicle
has any lateral speed towards the other vehicle. If by doing this the ego
vehicle does not exit the unsafe area, even when the lateral speed of the
ego vehicle is 0, then a longitudinal deceleration of at least amin,brake is
required. In the simulations carried out, the deceleration increases lin-
early, exactly as in the CC human driver model.
For the simulation of the cut-out and the deceleration scenarios, the
initial distance between the ego vehicle and the preceding vehicle is
within the control of the AV driving strategy. For the RSS simulation
model, the time headway should not be equal to the minimum safe
longitudinal distance, as this would mean that in the process of
achieving and maintaining the steady-state headway, the ego vehicle
could enter the unsafe distance. The same would happen in case of a
deceleration of the preceding vehicle, even if it is mild. This would be
against the RSS rules. Hence, the equilibrium distance is calculated such
that the ego vehicle will not enter the unsafe distance after a time equal
to
ρ
, if the preceding vehicle decelerates with amax,brake for this time, as
shown in equation (4).
dlat =
μ
+2u1+
ρ
alateral
max,accel
2
ρ
+u1+
ρ
alateral
max,accel2
2alateral
min,brake,correct
−2u2+
ρ
alateral
max,accel
2
ρ
+u2+
ρ
alateral
max,accel2
2alateral
min,brake,correct
(3)
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
5
dtime−headway =ur
ρ
+1
2amax,accel
ρ
2+ur+
ρ
amax,accel2
2amin,brake
−uf−
ρ
amax,brake2
2amax,brake
+1
2amax,brake
ρ
2
(4)
3.4. The fuzzy safety model
The Fuzzy Safety Model (FSM) is based on fuzzy surrogate safety
metrics for rear-end collisions. The simulation model consists of three
steps. First, the lateral safety distance is checked, and if there is no po-
tential risk identied, no reaction is required by the ego vehicle.
Otherwise, the process continues by checking the longitudinal distance.
This rst step is necessary only for the cut-in scenarios. For cut-out and
deceleration scenarios, it is skipped, as the vehicles are in the same lane
and the lateral distance is always potentially dangerous, so the process
starts from checking the longitudinal distance. If a risk is identied ac-
cording to the longitudinal distance check, the proper reaction is
calculated, in the term of a deceleration value to be achieved. The check
is carried out in every simulation step.
The lateral safety check uses three criteria. If all criteria are true, a
potential risk is identied. The three criteria are:
•The cutting-in vehicle is downstream.
•The cutting-in vehicle has lateral speed towards the ego vehicle.
•The following is true:
o If the longitudinal velocity of the ego vehicle is greater than the
longitudinal velocity of the cutting-in vehicle, equation (5):
distlat
ucut−in,lat
<distlon +lengthego +lengthcut−in
uego,lon −ucut−in,lon
+s1(5)
o If the longitudinal velocity of the ego vehicle is lower than the lon-
gitudinal velocity of the cutting-in vehicle, equation (6):
distlat
ucut−in,lat
<xego,front −xcut−in,back
ucut−in,lon −uego,lon
+s1(6)
where distlat is the lateral distance between the two vehicles, distlon is the
longitudinal distance between the two vehicles, lengthego and lengthcut−in
the length of the ego vehicle and the cutting-in vehicle respectively, s1 is
a time safety margin of 0.1 s, ucut−in,lat is the lateral velocity of the
cutting-in vehicle, uego,lon and ucut−in,lon the longitudinal velocity of the
ego vehicle and the cutting-in vehicle respectively. For the second case,
xego,front is the longitudinal position of the center of the front bumper of
the ego vehicle and xcut−in,back is the longitudinal position of the center of
the back bumper of the cutting-in vehicle.
The lateral safety check evaluates the lateral movement of the
cutting-in vehicle. Assuming no reaction, the position of the cutting
vehicle in the trajectory of the ego vehicle is estimated. For small lateral
speeds, it is possible that this position would be behind the ego-vehicle,
and there would be no risk of an accident. Otherwise, there is danger of
the cutting-in vehicle crashing to the side of the ego-vehicle or getting in
front of it in an unsafe distance, such that the ego-vehicle would not be
able to decelerate enough to avoid an accident. There is a safety margin
of 0.1 s that is further explained in the results section.
Two fuzzy surrogate safety metrics are evaluated for the longitudinal
safety check, the Proactive Fuzzy surrogate Safety metric (PFS) and the
Critical Fuzzy surrogate Safety metric (CFS) (Mattas et al., 2020). If both
metrics take the value 0, the situation is safe, and no reaction is required.
If any of the two takes a non-zero value, the condition is not safe, and a
reaction is required.
The value of PFS is calculated according to equation (7):
PFS(distlon) =
1,if 0<distlon −d1<dunsafe
0,if distlon −d1>dsafe
distlon −dsafe −d1
dunsafe −dsafe
,if dunsafe <distlon −d1<dsafe
(7)
where distlon is the longitudinal distance between the two vehicles, d1 is a
safety margin of 2 m, and dsafe, dunsafe are calculated according to
equations 8–9:
dsafe =uego,lon
τ
+u2
ego,lon
2bego,comf
−u2
cut−in,lon
2bcut−in,max
+d1(8)
dunsafe =uego,lon
τ
+u2
ego,lon
2bego,max
−u2
cut−in,lon
2bcut−in,max
(9)
with uego,lon the velocity of the ego vehicle, ucut−in,lon the velocity of the
cutting-in vehicle,
τ
the reaction time of the ego vehicle, bego,comf the
comfortable deceleration of the ego vehicle, bego,max the maximum
deceleration of the ego vehicle, bcut−in,max the maximum deceleration of
the cutting-in vehicle, d1 a safety distance margin of 2 m.
The value of CFS is calculated only when the ego vehicle velocity is
higher than the preceding vehicle, according to equation (10):
CFS(distlon) =
1,if 0<distlon <dunsafe
0,if distlon >dsafe
distlon −dsafe
dunsafe −dsafe
,if dunsafe <distlon <dsafe
(10)
where distlon is the longitudinal distance between the two vehicles, and
dsafe, dunsafe are calculated according to equations 11–15:
dsafe =
uego,lon −ucut−in,lon2
−2a′
ego
,if uego,lon,NEXT ≤ucut−in,lon
dnew +uego,lon,NEXT −ucut−in,lon 2
2bego,comf
,if uego,lon,NEXT >ucut−in,lon
(11)
dunsafe =
uego,lon −ucut−in,lon2
−2a′
ego
,if uego,lon,NEXT ≤ucut−in,lon
dnew +uego,lon,NEXT −ucut−in,lon 2
2bego,max
,if uego,lon,NEXT >ucut−in,lon
(12)
a′
ego =max(aego,−bego,comf )(13)
uego,lon,NEXT =uego,lon +a′
ego
τ
(14)
dnew =(uego,lon +uego,lon,NEXT)
2−ucut−in,lon
τ
(15)
with uego,lon the velocity of the ego vehicle, uego,lon,NEXT the expected ve-
locity of the ego vehicle after the reaction time assuming constant ac-
celeration, aego the current acceleration of the ego vehicle, a′
ego a
modied acceleration of the ego vehicle that cannot be harder than
bego,comf , dnew is the expected change in the distance between the two
vehicles after the response time, ucut−in,lon the velocity of the cutting-in
vehicle,
τ
the reaction time of the ego vehicle, bego,comf the comfortable
deceleration of the ego, bego,max the maximum deceleration of the ego
vehicle.
The deceleration intensity is relative to the values of PFS and CFS.
The reaction is that of a simplied rule-based Fuzzy Inference System, as
in equation (16).
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
6
breaction =CFSbego,max −bego,comf +bego,comf ,if CFS >0
PFSbego,comf ,if CFS =0(16)
where CFS and PFS are the values calculated in equations (10) and (7)
respectively. The reaction starts after a time equal to
τ
, and the decel-
eration value increases linearly, exactly as in the CC human driver
model. A signicant difference between this model to the other three is
that the simulated driver may react using a calm deceleration, in
anticipation of an emergency maneuver. For all the other models, the
decelerations required by the control strategy are either 0 or the
maximum.
4. Assessment methodology
4.1. Simulation framework
A simulation framework is developed in Python, to test the four
different models. The three types of scenarios, as previously described,
are a) cut-in, b) cut-out, and c) deceleration of the preceding vehicle.
The trajectories of the other vehicles are pre-determined to t the sce-
nario and vary depending on different initial conditions. The simulated
vehicle keeps constant longitudinal velocity and 0 lateral velocity unless
required to decelerate by the model used each time. All trajectories are
stored to investigate the level of challenge in each situation. The same
rules apply to the simulation of the real-world trajectories for the vali-
dation. The parameter values used for the Reg157 model, and the CC
human driver model are as presented in the Regulation 157. For the RSS
and FSM simulations, the parameters are presented in Table 1 and
Table 2, respectively. They are corresponding to the CC human driver
model where is tting. However, the parameter values can be subject to
discussion and change. In the present paper, the attributes and quali-
tative results of each model are compared.
4.2. Replication of regulation example cases
For all three types of scenarios, the examples presented in Appendix
3 of the regulation (UNECE, 2021) are recreated, with additional sim-
ulations for higher speed, up to 130 km/h to widen the scope of the
assessment to the whole motorway driving. The simulation framework
regarding the regulatory application is openly available
2
. In the cut-in
scenarios, the simulations are extended backward in time, so the
cutting-in vehicle starts with 0 lateral speed and accelerates with a
lateral acceleration of 1.5 m/s
2
. The distances are calculated so when the
lateral distance is equal to the scenario setting (1.6 m for all tests as in
Appendix 3), the lateral velocity and longitudinal distance of the vehi-
cles has just reached the indicated value. This is crucial for the RSS and
FSM, which could react in an anticipatory fashion. This would probably
t the way the experiments would be carried out on a test track. For all
models except the Reg157, the jerk is bounded, so the deceleration of the
ego vehicle is decreasing linearly when a risk is identied.
4.3. Validation using real data
Throughout the paper, real data are used for comparison purposes,
for xing parameter values, and, most importantly, for the comparison
of the models’ efciency. The data are from the highD database (Kra-
jewski et al., 2018). The human driver trajectories in the database are
collected on German highways, at six different locations near Cologne,
using unmanned aerial vehicles. The locations vary by the speed limits
and the number of lanes. The typical position error of the detected ve-
hicles is 10 cm.
The main limitation of this work is the small sample size of data.
Fatal trafc accidents are rare (Kalra and Paddock, 2016) and a vastly
larger dataset should be used. In fact, there are no accidents reported in
the highD dataset. Moreover, as any human driver drives differently
under different conditions, a perfect deterministic model of a competent
and careful human driver cannot exist. Due to those limitations, a
parameter calibration is not carried out. The validation part aims to
investigate the different mechanics of each model and to show why and
how the proposed model is better suitable to describe the reaction of a
competent and careful human driver.
The different models are not used to accurately replicate the
behavior of a human driver, which is a challenging task. Instead, they
are used to classify if a certain scenario should be considered prevent-
able (i.e. not leading to a collision) or unpreventable (i.e. leading to a
collision). Therefore, for the validation of the models, the way the
human driver reacted to a situation does not necessarily need to be
exactly replicated. On the contrary, what is important is that the sce-
narios that are preventable for a human driver in real life are also
classied as preventable by the models. Furthermore, the models are
also checked on cases in which there was no cut-in but two vehicles were
driving closely in adjacent lanes, to investigate whether the models are
too conservative and therefore also not suitable to be used as classiers
(as they would generate too many false-positive cases).
To examine the capacity of the models to correctly classify pre-
ventable cases, the most hazardous cut-in cases have been extracted
from the highD dataset. The models should classify the cases as pre-
ventable, as the human drivers were able to avoid an accident and as
there is no accident recorded in the highD dataset. Some assumptions
are required to obtain the worst cases:
•The vehicle receiving the lane change needs to be classied as a car.
Vehicles classied by the developers of the dataset as “truck” would
not be used as the ego vehicle.
•The ego vehicle should not have changed lanes during the recording.
Since every vehicle is recorded for a few seconds, a lane change by
the vehicle receiving the cut-in might mean that this was not a “pure”
cut-in maneuverer. Maybe the ego vehicle changed lanes to avoid a
dangerous situation, and the models are not capable of replicating
this movement. Therefore, any vehicle eligible to be considered as a
vehicle receiving a cut-in must have been completely inside the lane
markings for the whole recording.
Table 1
Value of the parameters used in the simulations involving
the RSS model.
Parameter RSS
ρ
0.75 s
amax,accel 3 m/s
2
amin,brake 6 m/s
2
amax,brake 6 m/s
2
μ
0.3 m
alateralmax,accel 1 m/s
2
alateralmin,brake,correct 1 m/s
2
Max abs deceleration 0.774 g
Max abs jerk 12.65 m/s
3
Table 2
Value of the parameters used in the simulations involving
the FSM.
Parameter FSM
τ
0.75 s
bego,comf 3 m/s
2
bego,max 6 m/s
2
bcut−in,max 7 m/s
2
Max abs deceleration 6 m/s
2
Max abs jerk 12.65 m/s
3
2
github.com/ec-jrc/JRC-FSM.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
7
•The cutting-in vehicle should start from another lane and end with at
least one wheel in the ego vehicle’s lane. Moreover, it should not
move to another lane during the recording. Those cases may corre-
spond to a double lane change and not to a cut-in.
•The trajectories are obtained only for the period that there are re-
cordings for both vehicles. Moreover, the time instance in which the
cutting-in vehicle rst touched the lane markings is identied. The
parts of the trajectories referring to a timestamp that is more than 5 s
earlier than this time instance are discarded as the behavior during
this is irrelevant to the maneuver.
•The TTC is calculated for each instance of the 25 Hz recordings. Only
the cases in which the TTC achieved a value of 5 s or less are obtained
for the validation. This large TTC threshold was chosen to include
cases where the driver’s proactive behavior prevented a more severe
conict.
Regarding the cases where no reaction would be necessary as the
“other” vehicle did not perform a lane change, pairs of vehicles traveling
in adjacent lanes are identied. Again, several assumptions are used:
•Both vehicles have to be always inside the markings of their
respective lanes. If not, there are cases where one of the vehicles
drifts in the other lane, without completing the lane change during
the observed period, and the vehicle upstream would need to
decelerate or move laterally to avoid a collision.
•The time for which there are data for both vehicles should be larger
than 3 s, otherwise the maneuverer is too short to be used as an
example.
•One of the two vehicles should always be in front of the other one,
looking at their rightmost position from the highD data. Overtaking
maneuvers are not relevant to the current models’ focus.
5. Results
In this section, the results of the simulated scenarios are presented.
The 4 different reaction models are designed to estimate through
simulation if a scenario is preventable or unpreventable. The Reg157
model is used only for the cut-in scenarios. Finally, the efciency of the
models regarding the cut-in scenarios is investigated using real trajec-
tory data.
5.1. Cut-in scenarios
The simulations reconstruct the results presented in the regulation
Appendix 3. The lateral distance between the two vehicles is 1.6 m. For
the Reg157 and the CC human driver model, the position of the lane
markings affects the time of reaction of the ego vehicle. For the simu-
lation experiments presented, the lane markings are assumed to be
exactly in the middle of the distance between the two vehicles. The rest
of the parameters are presented in Table 3 for the low-speed scenarios
and Table 4 for the high-speed scenarios.
The most challenging combination of velocities for the low-speed
scenarios is the cutting-in vehicle running with 10 km/h longitudinal
velocity and the ego vehicle 60 km/h. The results regarding the pre-
ventability of each scenario for this combination, are presented in
Fig. 1a, and an example for the high-speed scenarios with the cutting-in
vehicle running with 40 km/h longitudinal velocity and the ego vehicle
130 km/h in Fig. 1b. The green dots represent the cases that have been
judged to be unpreventable by the RSS model, the magenta “Y” markers
regard the Reg157 unpreventable cases, the red “X” markers the CC
human driver, and the blue transparent large dots represent the pro-
posed FSM, for all combinations of the lateral velocity of the cutting-in
vehicle and the initial longitudinal distance.
A rst observation is that the unpreventable cases dened by the
Reg157 and CC human driver models are similar. There is some differ-
ence, because of the different assumptions of the reaction time,
maximum deceleration, and the different assumptions on when a vehicle
initiates the reaction. The formulation of both models is based on
calculating the TTC. Thus, the simulated driver reacts to emergencies.
On the other hand, both the RSS model and the FSM model assume
anticipative behavior. The velocity and stopping distance of both vehi-
cles is used to foresee upcoming emergencies and react proactively.
Hence, both the RSS and FSM avoid accidents when the initial distance is
large and the cutting-in lateral velocity is small. While the model pa-
rameters are in essence assumptions, their values would not change this
observation qualitatively. The current formulation of the CC human
driver model and the Reg157 model is not capable of anticipating
emergencies.
The main argument for the use of a model that is capable of recre-
ating anticipatory behavior is that human drivers are also capable of
such behavior. This is conrmed by investigating lane changes in the
highD data set. Three examples are presented in Fig. 2, with the red line
denoting the longitudinal velocity of the cutting-in vehicle and the blue
line the velocity of the vehicle receiving the cut-in. The time in which
the closer side of the cutting-in vehicle reached the lane marking is
denoted by the black, vertical line. In all three cases, the vehicle that
received the cut-in was able to anticipate the maneuver and started
decelerating a few seconds earlier. Since the data refer to human drivers,
it can be assumed that their reaction time is non-negligible, and the
situation has been predicted even earlier than when the deceleration
was realized. Thus, it may be argued that AVs would be expected to have
the same capabilities and avoid accidents by anticipation if a human
driver would avoid it as well.
Two aspects separate RSS from the FSM. The right boundary of the
unpreventable area for the RSS model is on the left of that of the FSM
model. Hence, some cases that would be classied as preventable using
the RSS, are unpreventable according to the FSM. The FSM uses calm
deceleration when an emergency is predicted, while the RSS model can
only react with a full deceleration. Hence, the RSS decelerates harder
and can avoid an accident with slightly less space. This is directly
affected by the assumed parameters of the models.
The second difference between the two modeled results is an area for
small initial distance and low lateral velocity of the cutting-in vehicle,
which is identied as unpreventable only by the RSS of all the models.
From a visual inspection of the simulations, it becomes clear that for
such cases, the longitudinal relative speed is large, and the approaching
rate of the cutting-in vehicle is small. Hence, if the ego vehicle does not
decelerate, it would be in front of the cutting-in vehicle when their
trajectories intersect. However, the conservative behavior of the RSS
Table 3
Parameter values for low-speed simulation experiments of cut-in.
Minimum
value
Maximum
value
Step
Longitudinal distance 1 m 60 m 1 m
Cutting-in vehicle lateral
velocity
0.1 m/s 1.8 m/s 0.1 m/s
Ego vehicle velocity 10 km/h 60 km/h 10 km/h
Cutting-in vehicle velocity 10 km/h 50 km/h 10 km/h
Table 4
Parameter values for high-speed simulation experiments of cut-in.
Minimum
value
Maximum
value
Step
Longitudinal distance 1 m 120 m 2 m
Cutting-in vehicle lateral
velocity
0.1 m/s 1.8 m/s 0.1 m/s
Ego vehicle velocity 70 km/h 130 km/h 20 km/h
Cutting-in vehicle velocity 10 km/h 100 km/h 30 km/h
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
8
requires the simulated vehicle to decelerate in anticipation, and a
collision occurs.
These accidents are avoided with the FSM, because of the constrain
in equation (5). When the inequality is not satised, the approaching
rate of the cutting-in vehicle is too small compared to the relative speed,
so the ego vehicle can avoid the accident by keeping a constant speed. In
this area, small changes in the parameters can drastically change the
result of the simulation. A small difference in the relative velocity of the
vehicles could be the deciding factor between a very critical unpre-
ventable situation and a situation that would be prevented without any
corrective maneuvers. This discontinuity may lead to sensor and model
uncertainties becoming important. Hence, the safety margin of 0.1 s is
used in the equation (5), so these cases that are almost accidents are not
considered preventable. Overall, the FSM is close to the intersection of
the other three models showing a good compromise between the
different approaches.
The results for other scenario parameters are qualitatively the same.
The results for other cases are presented in this paper’s annex.
An advantage of using a simulation framework is that a complete
trajectory is generated. Using the trajectory and relevant safety metrics,
different scenarios can be classied for the level of challenge, and not
just as being preventable or unpreventable. This can be useful for the
regulation of ADS, as it should cover the testing procedures that a
technical service should run to check for compliance. An idea discussed
in the amendment of Regulation 157 is that the testing authority should
be allowed to test any scenario and not only specic ones, to avoid
Fig. 1. Unpreventable cases for all 4 models for the (a) ego vehicle running with 60 km/h and the cutting-in vehicle running with 10 km/h longitudinal speed, (b)
ego vehicle running with 130 km/h, and the cutting-in vehicle running with 40 km/h longitudinal speed.
Fig. 2. Three cases observed in the highD dataset, in which the human driver reacts in anticipation of a cut-in, using a soft deceleration.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
9
designs tted on specic test cases. Therefore, an evaluation of the
challenge level for all preventable scenarios is important, to ensure that
at least a minimum number of challenging scenarios will be tested.
Both the Reg157 model and CC human driver model use the TTC to
identify emergencies and request a reaction. Therefore, the minimum
TTC value observed would be very small, even for scenarios that are not
very challenging. On the other hand, the RSS framework requests a very
strong and early deceleration. An example of the results for the same
speeds is presented in Fig. 3. The lower TTC values are denoted by the
red coloring, and the higher, and thus safer, TTC values with green, for
the Reg157 (a) and RSS model (b). For the Reg157 the TTC should be
low to trigger a reaction so all cases of the vehicle using a deceleration to
avoid an accident are colored red. On the other hand, the hard reaction
of RSS shows cases that are close to the unpreventable front, to be much
Fig. 3. The minimum TTC value observed in the cut-in simulations for a) Reg157 model, b) RSS model.
Fig. 4. Classication of easy (green), medium (yellow), and difcult (red) scenarios according to the FSM.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
10
less challenging.
The FSM is based on the PFS and CFS safety metrics. Thus, it natu-
rally lends itself to classication in three different classes. When the CFS
has achieved values close to 1, the scenario can be assumed to be very
challenging. In the present example, the threshold of CFS to be higher
than 0.9 is used. Otherwise, if large values of PFS have not been ach-
ieved, the scenario was rather easy to handle. The threshold used in this
example is for PFS to achieve values lower than 0.85. Scenarios with
values in-between can be classied as scenarios of medium challenge.
Examples of the following classication are shown in Fig. 4. The com-
binations of velocities of the ego vehicle and the cut-in vehicle are
60–10, 60–30, 30–20, and 130–100 km/h in Fig. 4a,Fig. 4b, Fig. 4c, and
Fig. 4d respectively. A distinct difference of the areas is presented,
conditional to the velocity of the two vehicles. The results for the rest of
the cases are presented in this paper’s annex.
5.2. Preceding vehicle sharp deceleration scenarios
Simulations of deceleration scenarios have been run for the CC
human driver model, the RSS model, and the FSM. The scenarios have
different initial conditions, as the steady-state distance required by each
model is different. For the CC human driver, it is assumed to be 2 s time
headway for any speed. The initial distance of the RSS model is ac-
cording to equation 4. Finally, the initial distance of the FSM is
conservatively assumed to be when PFS is 0.
Both the CC human driver model and the FSM did not nd any
unpreventable cases. On the other side, the RSS model identied a few
when the speed is very high, and the deceleration of the preceding
vehicle is much higher than the parameter assumed in this work. The
RSS simulated driver with more conservative parameter values, espe-
cially regarding the maximum deceleration of the preceding vehicle,
would also be able to avoid all accidents. The gures are presented in the
annex of the current paper, along with the classication of different
scenarios according to the level of challenge, when using the FSM. The
results are intuitive as harder decelerations are classied as the most
difcult cases.
5.3. Cut-out scenarios
The cut-out scenarios presented in the current regulation text
consider only cases of the ego vehicle speed not exceeding 60 km/h. The
variables are the lateral speed of the preceding vehicle and the distance
in front of the preceding vehicle where a static obstacle is located. The
ego vehicle and preceding vehicle have the same speed and the headway
is assumed to be the steady-state one. The preceding vehicle starts a
lateral movement to change lane, revealing a static vehicle (obstacle) in
a certain distance in front of it. The ego vehicle may start reacting when
the preceding vehicle has deviated more than 0.375 m, which is the
wandering zone. The reaction will not be realized before a time equal to
the reaction time of each model.
For such scenarios, no unpreventable cases are predicted in the
regulation text. Moreover, the CC human driver is showing all cases to
be avoidable. This is conrmed also by the FSM and RSS model, as
shown in Fig. 5a, b, and c for the FSM, CC human driver, and RSS
models, respectively. Green dots represent cases where the accident is
prevented, and yellow dots represent cases of the preceding vehicle
Fig. 5. Results of cut-out scenarios with the ego vehicle speed equal to 60 km/h using a) FSM, b) CC human driver, c) RSS.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
11
crashing into the static object. Crashes would be denoted by red ‘x’
markers, but there have been non in this case.
The regulation text does not suggest a model for the cut-out sce-
narios, assuming that all cases should be avoidable. For higher speeds
though, the situation is more critical. In Fig. 6, the relevant results are
presented for 110 km/h. For all three models, there are a few cases that
would be regarded as unavoidable. For this case, in contrast with the
cut-in case, tactical decision-making would not lead to less unprevent-
able scenarios. The risk is not known before the preceding vehicle exits
the wandering zone, so there cannot be an anticipatory deceleration.
Moreover, for all the unpreventable cases, the emergency deceleration
was not enough, or the reaction time was not short enough. Therefore,
all model formulations are equally capable in the classication of
different cases for cut-out, and the reaction time and maximum decel-
eration parameters are crucial in the result. The results of the rest of the
cases are presented in this paper’s annex.
Again, FSM has the advantage of classifying the different situations
according to the level of challenge. In this specic case, a disadvantage
of the simplistic formulation can affect the results of the classication,
but not of the distinction between preventable and unpreventable. With
a static obstacle upstream, the values of PFS and CFS are very similar. As
a result, the CFS values increase, before the vehicle starts taking action
to avoid the collision. Hence, the CFS value will be large for almost all of
the simulated scenarios. This drawback could be eliminated by using a
more complex control strategy, based on the Fuzzy SSMs. However, it
may be counter-productive to use different model formulations for
different cases or adopt more complicated models containing more pa-
rameters whose values need to be established. Hence, the situations will
not be classied by the maximum value of CFS and PFS, but by the value
of CFS and PFS at the time that the ego vehicle driver identies the
danger when the preceding vehicle is for the rst instance out of the
wandering zone. The thresholds for the classication of the different
values are thus different. Cases of PFS and CFS equal to 0 are considered
easy. Scenarios of medium challenge are those who have a CFS value
that is less than 0.5. The cases of the CFS value being larger than 0.5 are
assumed to be the most challenging situations.
An example is shown in Fig. 7. The green dots represent the easiest
cases and yellow are the medium difculty cases. The most challenging
preventable cases are shown by red dots, while the unpreventable cases
when the simulated ego vehicle crashed into the object are denoted by
red “x” markers. The ego vehicle speed and the distance to the static
object are the most signicant factors. The results for the rest of the cases
are presented in this paper’s annex.
5.4. Validation: Cut-in cases
Overall, 99 cases of severe cut-in maneuvers have been identied in
the highD dataset. As shown in the histograms of Fig. 8a, the median
velocity of the ego vehicle covers a wide range, including low speeds and
high speeds that are over the speed limit. Similarly, the relative velocity
is presented in Fig. 8b, which is shown to be lower than 10 m/s for most
cases.
Out of the 99 cases isolated, the RSS and the FSM are the only ones
that classied all cases correctly as preventable for a human driver.
Finding no false negative cases, i.e., cases incorrectly classied as
unpreventable, is important for such a model in this specic use. The CC
Fig. 6. Results of cut-out scenarios with the ego vehicle speed equal to 130 km/h using a) FSM, b) CC human driver, c) RSS.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
12
human driver model and the Req157 produced 9 and 14 false negatives
respectively. The results are presented in the bar plot of Fig. 9a.
Models allowing collisions for situations in which a human driver
would avoid a collision, could allow for automated driving systems that
are not as safe as human drivers. Especially since the cases are based on
real data so they occur in real trafc. The introduction of such systems to
the market could increase the frequency of such accidents. Moreover, for
all the cases that the CC human driver model and the Reg157 found to be
unavoidable, the human drivers tracked in the dataset managed to avoid
accidents without exploiting the full vehicle dynamics. The minimum
acceleration value achieved for each case is presented in Fig. 9b and c for
the CC human driver model and the Reg157 model respectively. For all
Fig. 7. Classication of easy (green), medium (yellow), and difcult (red) scenarios according to the FSM level and crashes (“x”) for cut-out cases, for ego vehicle
speed of (a) 40 km/h, (b) 80 km/h, (c) 130 km/h.
Fig. 8. Histogram of (a) the median speed of the ego vehicle and (b) the relative median speed of the two vehicles, for the 99 most severe cut-in cases identied.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
13
the cases in which those models identied an unavoidable collision, the
human driver never exceeded a deceleration of 3 m/s
2
. Moreover, dur-
ing the simulations of the FSM for the same cases, the accident was
avoided without the use of decelerations harder than 4 m/s
2
. This
further shows the importance of the “tactical” safety level, with the
driver anticipating emergencies before they become imminent.
5.5. Validation: Vehicles traveling in adjacent lanes cases
The number of vehicle pairs identied has been 243,897. For each
case, the upstream vehicle, which is considered to be the ego vehicle,
may have used some deceleration, either because of the movement of
the other vehicle or as a result of the surrounding trafc. Regardless of
the behavior of the tracked human driver, any deceleration requested by
any of the models is considered to be a false positive, as the vehicle in the
adjacent lane never crossed the lane markings.
During the simulations, the Reg157 model was never activated, as
the other vehicle never reached the lane markings. The CC human driver
model was activated 3 times in total, showing a very small number of
false positives. As presented in Fig. 10, the number of cases in which a
deceleration was requested was 15,782 for the FSM model and 786 for
the RSS-based model. This rst nding indicates the FSM model is prone
to false-positive classication, for the current parameter values. How-
ever, the actual vehicle had at least one instant of deceleration for
190,063 of the cases. For the human driver, we cannot assume that this
deceleration is due to the other vehicle under investigation, as it could
be caused by other reasons such as the surrounding trafc.
A more informative illustration of the results is presented in Fig. 11a
and b. In Fig. 11a the number of vehicles achieving a minimum decel-
eration value is shown for a range of deceleration values, on a loga-
rithmic scale for better readability. The capability of the FSM to judge
Fig. 9. (a). Number of false negatives for each model. Minimum acceleration used for the human driver in false-positive cases of the (b) CC human driver model (c)
Reg157 model.
Fig. 10. Number of cases where a deceleration was observed for each model
and for the real drivers.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
14
the safety level based on fuzzy metrics and react according to the level of
unsafety of the situation creates a large gap between the results of FSM
and RSS. While the RSS model decelerated in much fewer cases, the
deceleration has been much harder. Those kinds of maneuvers can affect
both trafc ow and trafc safety. It seems it is very different from a
human driver’s behavior, so that would probably be a surprising
behavior for the surrounding trafc. Setting different parameter values
would affect the magnitude of the deceleration for the RSS, but the shape
of the distribution is not expected to change. This is because the reaction
of the controller is binary (active or inactive), which comes from using
crisp sets for describing the safety conditions.
On the contrary, the FSM decelerated for more cases than any other
model. However, the cumulative distribution plot of the minimum
deceleration shows how this is not dissimilar to the behavior of the
human driver. The larger difference is shown in the range from −4 m/s
2
to −2 m/s
2
that are the boundaries of a comfortable deceleration that
have been set. The parameter values have not been optimized to t the
results. It is shown how the FSM distribution is the only one that has the
capacity to track, with some error, the real distribution. Moreover, for
the proposed parameter values, the cases of hard deceleration with the
FSM are even less than the ones recorded for the human drivers. This
suggests that the decelerations assumed would not be outside what is
common or frequent in the real world and they would not surprise the
surrounding trafc.
Apart from the maximum deceleration value achieved, which rep-
resents the maximum intensity of the deceleration, the duration plays an
important role. To show this, the maximum speed drop from the initial
speed is presented in Fig. 11b for the observed data, the FSM, and the
RSS model. The empirical cumulative distribution of the actual human
driver is higher than both models. The only exception is an outlier of the
RSS model with a speed reduction of 25 m/s. Those results indicate that
FSM would not request very hard decelerations with a high duration.
Moreover, it would not deteriorate the ow by requesting large speed
drops. It has to be claried that any manufacturer could use any control
strategy they would like, as far as it is shown to be as safe as the refer-
ence model. Therefore, a manufacturer, having more accurate knowl-
edge of the response time and the vehicle dynamics, could produce
automated driving systems that are as safe, with an even better impact
on the trafc ow. Especially considering that the response time of an
automated driving system could potentially be shorter than the 0.75 s
assumed for the FSM.
6. Conclusions
Automated driving systems bring important promises to the future of
road transportation. Increased safety and efciency, decreased envi-
ronmental impacts, and time in congestion are some of the possible
signicant benets. Meanwhile, more equitable access to transportation
and overall better use of the infrastructure could, in the long run, bring
large societal benets. For all those to be realized, the ADS should be
accepted by the public. Safety is a crucial precondition for the public’s
trust, and for the systems to become dominant in the market. However,
ensuring safety is not a trivial task. Regulations can be important on this
path, and an example is the rst global regulation on the approval of a
Level 3 ADS adopted in 2021.
In the present paper, we present and explore the models used in the
existing regulation to set safety performance requirements and addi-
tional two that were proposed in the regulation amendment, namely, the
RSS, an industry proposed safety framework, and the FSM proposed by
the authors.
The analysis shows that the models included in the existing regula-
tion only focus on the reaction of the ADS to emergencies (similarly to
what would be appropriate for ADAS systems). On the other hand, some
of the eld experts suggest that ADS should be able to drive defensively,
predicting and avoiding emergencies in advance, as this is an important
skill of an expert driver. An obvious trade-off exists, as the strictest
regulations can have impacts similar to that of operation requirements,
while less restrictive ones may not properly embody the ambitions for
safe automated trafc.
Regarding the RSS model, while this is not the main scope of the
model developers, interesting comparisons can be drawn between the
performance requirements produced by it and the existing regulation.
RSS can shrink the set of unpreventable cases. On the other hand, some
cases that would not lead to an accident if no reaction was required,
resulted in accidents when using the RSS model. Finally, the model
proposed by the authors, based on fuzzy surrogate safety metrics, is able
to simulate a defensive driving behavior and includes the ability of a
calm, proactive reaction. The unpreventable cases identied are quali-
tatively close to an intersection between the other three models.
It is shown that in case anticipative driving is requested, the pro-
posed model, FSM, can be used to derive the sets of preventable or
unpreventable cases. This would not restrict the ADS operation, as it
represents a minimum level, and the way of reacting is not dictated.
Fig. 11. Cumulative distribution of (a) maximum deceleration value (b) maximum speed drop.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
15
Thus, such a model would not delay the potential trafc ow and ef-
ciency benets. Moreover, the proposed model can be used for the
classication of preventable cases, according to the level of challenge.
This can be very useful for the testing requirements, which can also be a
part of such regulation.
Finally, the models have been validated for the cut-in case, using
trajectory data from real drivers. It is shown that careful human drivers
are able to proactively identify potential emergencies and use calm re-
actions to avoid them. Neglecting this crucial part, both the CC human
driver model and the Reg157 model produced an unnecessarily large
number of false-negative cases. This accounts for a non-optimal ambi-
tion in the actual efciency of automated driving systems and could
potentially lead to systems that are not as safe as a human driver
reaching the public streets.
On the other hand, the investigation of false-positive cases has shown
that FSM would request a deceleration for more cases than the RSS
model. However, the requested deceleration would be milder, and the
overall speed-drop smaller for the FSM. The RSS model is still based on a
binary denition of what is considered to be safe and unsafe, so all re-
actions are requesting a very hard deceleration. FSM was shown to be
better equipped to track the actual cumulative distribution of de-
celerations for human drivers. In the future the authors plan to validate
on additional datasets, possibly coming from different regions, to further
validate the model and the suitable parameter values.
For all the above, the FSM was successfully proposed by the authors
to represent a new way to dene the performance requirements of Level
3 motorway ADS in the amendment of UN Regulation 157.
CRediT authorship contribution statement
K. Mattas: Conceptualization, Data curation, Methodology, Visual-
ization, Formal analysis, Software, Writing – original draft, Writing –
review & editing. G. Albano: Data curation, Formal analysis, Software.
R. Don`
a: Formal analysis, Software, Writing – original draft, Writing –
review & editing. M.C. Galassi: Funding acquisition, Project adminis-
tration, Writing – original draft, Writing – review & editing. R. Suarez-
Bertoa: Writing – original draft, Writing – review & editing. S. Vass:
Writing – original draft, Writing – review & editing. B. Ciuffo:
Conceptualization, Methodology, Visualization, Funding acquisition,
Project administration, Writing – original draft, Writing – review &
editing.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
This research has been funded by the Joint Research Centre (JRC) of
the European Commission. The opinions expressed in this manuscript
are those of the authors and should not be considered to represent an
ofcial opinion of the European Commission. The authors are grateful to
Louison Duboz for helping with suggestions and recommendations
concerning parts of the relevant literature. The authors are specially
grateful to Antony Lagrange from the European Commission DG GROW
for the continuous suggestions and for the support to our proposal, and
to the experts of the Special Interest Group on Regulation 157 operating
under the UNECE-GRVA working party, for all the technical discussions
during the preparation of the regulation amendment, which have
certainly helped improving the approach proposed by the authors. The
work by G. Albano and R. Don`
a has been entirely carried out for the
JRCJ.
References
Akagi, Y., Kato, R., Kitajima, S., Antona-Makoshi, J., Uchida, N., 2019. A Risk-index
based Sampling Method to Generate Scenarios for the Evaluation of Automated
Driving Vehicle Safety*. In: in: 2019 IEEE Intelligent Transportation Systems
Conference (ITSC). Presented at the 2019 IEEE Intelligent Transportation Systems
Conference (ITSC), pp. 667–672. https://doi.org/10.1109/ITSC.2019.8917311.
Baumann, M.F., Br¨
andle, C., Coenen, C., Zimmer-Merkle, S., 2019. Taking responsibility:
a responsible research and innovation (RRI) perspective on insurance issues of semi-
autonomous driving. Transp. Res. Part A: Policy Pract. 124, 557–572. https://doi.
org/10.1016/j.tra.2018.05.004.
Blumenthal, M.S., Fraade-Blanar, L., Best, R., Irwin, J.L., 2020. Safe Enough: Approaches
to Assessing Acceptable Safety for Automated Vehicles. RAND Corporation, Santa
Monica, CA. 10.7249/RRA569-1.
Brunson, S.J., Kyle, E.M., Phamdo, N.C., Preziotti, G.R., 2002. ALERT ALGORITHM
DEVELOPMENT PROGRAM: NHTSA REAR-END COLLISION ALERT ALGORITHM.
Doi, A., Butsuen, T., Niibe, T., Takagi, T., Yamamoto, Y., Seni, H., 1994. Development of
a rear-end collision avoidance system with automatic brake control. JSAE Rev. 15,
335–340.
European Commission, 2020. Special Eurobarometer 496: Expectations and Concerns
from a Connected and Automated Mobility - Data Europa EU [WWW Document].
URL https://data.europa.eu/data/datasets/s2231_92_1_496_eng?locale=en
(accessed 1.4.22).
Fagnant, D.J., Kockelman, K., 2015. Preparing a nation for autonomous vehicles:
opportunities, barriers and policy recommendations. Transp. Res. Part A: Policy
Pract. 77, 167–181. https://doi.org/10.1016/j.tra.2015.04.003.
Feng, S., Yan, X., Sun, H., Feng, Y., Liu, H.X., 2021. Intelligent driving intelligence test
for autonomous vehicles with naturalistic and adversarial environment. Nat.
Commun. 12, 748. https://doi.org/10.1038/s41467-021-21007-8.
Grosso, M., Cristinel Raileanu, I., Krause, J., Alonso Raposo, M., Duboz, A., Garus, A.,
Mourtzouchou, A., Ciuffo, B., 2021a. How will vehicle automation and
electrication affect the automotive maintenance, repair sector? Transp. Res.
Interdiscip. Perspect. 12, 100495 https://doi.org/10.1016/j.trip.2021.100495.
Grosso, M., Duboz, A., Alonso Raposo, M., Krause, J., Mourtzouchou, A., Marques dos
Santos, F.L., Ciuffo, B., 2021b. Connected and Automated Driving: Expectations and
Concerns of European Citizens. Presented at the Transportation Research Board
100th Annual MeetingTransportation Research BoardTransportation Research
Board.
JAMA, 2020. Automated Driving Safety Evaluation Framework Ver. 1.0 -Guidelines for
Safety Evaluation of Automated Driving Technology.
Junietz, P., Bonakdar, F., Klamann, B., Winner, H., 2018. Criticality Metric for the Safety
Validation of Automated Driving using Model Predictive Trajectory Optimization. In:
in: 2018 21st International Conference on Intelligent Transportation Systems (ITSC).
Presented at the 2018 21st International Conference on Intelligent Transportation
Systems (ITSC), pp. 60–65. https://doi.org/10.1109/ITSC.2018.8569326.
Kalra, N., Paddock, S.M., 2016. Driving to safety: how many miles of driving would it
take to demonstrate autonomous vehicle reliability? Transp. Res. Part A: Policy
Pract. 94, 182–193. https://doi.org/10.1016/j.tra.2016.09.010.
Kassens-Noor, E., Kotval-Karamchandani, Z., Cai, M., 2020. Willingness to ride and
perceptions of autonomous public transit. Transp. Res. Part A: Policy Pract. 138,
92–104. https://doi.org/10.1016/j.tra.2020.05.010.
Kolekar, S., de Winter, J., Abbink, D., 2020. Human-like driving behaviour emerges from
a risk-based driver model. Nat. Commun. 11, 4850. https://doi.org/10.1038/
s41467-020-18353-4.
Kovvali, V.G., Alexiadis, V., Zhang, P.E., L.,, 2007. Video-Based Vehicle Trajectory Data
Collection. Presented at the Transportation Research Board 86th Annual
MeetingTransportation Research Board.
Krajewski, R., Bock, J., Kloeker, L., Eckstein, L., 2018. The highD Dataset: A Drone
Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of
Highly Automated Driving Systems. In: in: 2018 21st International Conference on
Intelligent Transportation Systems (ITSC). Presented at the 2018 21st International
Conference on Intelligent Transportation Systems (ITSC), pp. 2118–2125. https://
doi.org/10.1109/ITSC.2018.8569552.
Laureshyn, A., Johnsson, C., Ceunynck, T.D., Svensson, Å., Goede, M. de, Saunier, N.,
Włodarek, P., Horst, R. van der, Daniels, S., 2016. Review of current study methods
for VRU safety. Appendix 6 – Scoping review: surrogate measures of safety in site-
based road trafc observations: Deliverable 2.1 – part 4. [WWW Document].
undened. URL /paper/Review-of-current-study-methods-for-VRU-safety.-6-%E2%
80%93-Laureshyn-Johnsson/bf97ad6cb2233c33ad3f1779dcacc2d414ce8d73
(accessed 6.3.20).
Laureshyn, A., Svensson, Å., Hyd´
en, C., 2010. Evaluation of trafc safety, based on
micro-level behavioural data: theoretical framework and rst implementation.
Accid. Anal. Prev. 42, 1637–1646. https://doi.org/10.1016/j.aap.2010.03.021.
Lee, D., Hess, D.J., 2020. Regulations for on-road testing of connected and automated
vehicles: assessing the potential for global safety harmonization. Transp. Res. Part A:
Policy Pract. 136, 85–98. https://doi.org/10.1016/j.tra.2020.03.026.
Li, S., Sui, P.-C., Xiao, J., Chahine, R., 2019. Policy formulation for highly automated
vehicles: emerging importance, research frontiers and insights. Transp. Res. Part A:
Policy Pract. 124, 573–586. https://doi.org/10.1016/j.tra.2018.05.010.
Liu, R., He, J., Zhu, X., 2019. Potential Risk Assessment Algorithm in Car Following (SAE
Technical Paper No. 2019- 01–1024). SAE International, Warrendale, PA. 10.4271/
2019-01-1024.
Liu, S., Wang, X., Hassanin, O., Xu, X., Yang, M., Hurwitz, D., Wu, X., 2021. Calibration
and evaluation of responsibility-sensitive safety (RSS) in automated vehicle
performance during cut-in scenarios. Transp. Res. Part C: Emerg. Technol. 125,
103037 https://doi.org/10.1016/j.trc.2021.103037.
K. Mattas et al.
Accident Analysis and Prevention 174 (2022) 106743
16
Loos, S.M., Platzer, A., Nistor, L., 2011. Adaptive Cruise Control: Hybrid, Distributed,
and Now Formally Veried, in: Butler, M., Schulte, W. (Eds.), FM 2011: Formal
Methods, Lecture Notes in Computer Science. Springer, Berlin, Heidelberg, pp.
42–56. 10.1007/978-3-642-21437-0_6.
Mahmud, S.M.S., Ferreira, L., Hoque, M.S., Tavassoli, A., 2019. Micro-simulation
modelling for trafc safety: a review and potential application to heterogeneous
trafc environment. IATSS Res. 43, 27–36. https://doi.org/10.1016/j.
iatssr.2018.07.002.
dos Santos, F.L.M., Duboz, A., Grosso, M., Raposo, M.A., Krause, J., Mourtzouchou, A.,
Balahur, A., Ciuffo, B., 2022. An acceptance divergence? Media, citizens and policy
perspectives on autonomous cars in the European Union. Transp. Res. Part A: Policy
Pract. 158, 224–238.
Mattas, K., Botzoris, G., Papadopoulos, B., 2019. Framework for fuzzy surrogate metrics
for modeling road safety. Presented at the CENTRAL EUROPEAN SYMPOSIUM ON
THERMOPHYSICS 2019 (CEST), Banska Bystrica, Slovakia, p. 440006. 10.1063/
1.5114465.
Mattas, K., Makridis, M., Botzoris, G., Kriston, A., Minarini, F., Papadopoulos, B., Re, F.,
Rognelund, G., Ciuffo, B., 2020. Fuzzy Surrogate Safety Metrics for real-time
assessment of rear-end collision risk. A study based on empirical observations. Accid.
Anal. Prev. 148, 105794 https://doi.org/10.1016/j.aap.2020.105794.
Mullakkal-Babu, F.A., Wang, M., Farah, H., van Arem, B., Happee, R., 2017. Comparative
assessment of safety indicators for vehicle trajectories on highways. Transp. Res.
Rec.: J. Transp. Res. Board 2659, 127–136. https://doi.org/10.3141/2659-14.
Nordhoff, S., de Winter, J., Kyriakidis, M., van Arem, B., Happee, R., 2018. Acceptance of
driverless vehicles: results from a large cross-national questionnaire study. J. Adv.
Transp. 2018, 1–22. https://doi.org/10.1155/2018/5382192.
Nordhoff, S., Stapel, J., van Arem, B., Happee, R., 2020. Passenger opinions of the
perceived safety and interaction with automated shuttles: a test ride study with
‘hidden’ safety steward. Transp. Res. Part A: Policy Pract. 138, 508–524. https://doi.
org/10.1016/j.tra.2020.05.009.
NVIDIA, 2019. Safety Force Field for Autonomous Vehicles | NVIDIA Drive [WWW
Document]. URL https://www.nvidia.com/en-us/self-driving-cars/safety-force-
eld/ (accessed 3.16.22).
Park, H., Oh, C., Moon, J., Kim, S., 2018. Development of a lane change risk index using
vehicle trajectory data. Accid. Anal. Prev. 110, 1–8. https://doi.org/10.1016/j.
aap.2017.10.015.
Rasekhipour, Y., Khajepour, A., Chen, S.-K., Litkouhi, B., 2017. A potential eld-based
model predictive path-planning controller for autonomous road vehicles. IEEE Trans.
Intell. Transp. Syst. 18, 1255–1267. https://doi.org/10.1109/TITS.2016.2604240.
Ryan, C., Murphy, F., Mullins, M., 2020. Spatial risk modelling of behavioural hotspots:
risk-aware path planning for autonomous vehicles. Transp. Res. Part A: Policy Pract.
134, 152–163. https://doi.org/10.1016/j.tra.2020.01.024.
SAE, 2021. J3016C: Taxonomy and Denitions for Terms Related to Driving Automation
Systems for On-Road Motor Vehicles - SAE International [WWW Document]. URL htt
ps://www.sae.org/standards/content/j3016_202104/ (accessed 1.4.22).
Schoener, H.-P., 2020. Challenging Highway Scenarios Beyond Collision Avoidance for
Autonomous Vehicle Certication. 10.13140/RG.2.2.29355.05926.
Seiniger, P., Bartels, O., Pastor, C., Wisch, M., 2013. An open simulation approach to
identify chances and limitations for vulnerable road user (VRU) active safety. Trafc
Inj. Prev. 14, S2–S12. https://doi.org/10.1080/15389588.2013.797574.
Shalev-Shwartz, S., Shammah, S., Shashua, A., 2017. On a Formal Model of Safe and
Scalable Self-driving Cars. arXiv:1708.06374 [cs, stat].
Shladover, S.E., Nowakowski, C., 2019. Regulatory challenges for road vehicle
automation: lessons from the California experience. Transp. Res. Part A: Policy Pract.
122, 125–133. https://doi.org/10.1016/j.tra.2017.10.006.
Tarko, A.P., 2018. Estimating the expected number of crashes with trafc conicts and
the Lomax Distribution – A theoretical and numerical exploration. Accid. Anal. Prev.
113, 63–73. https://doi.org/10.1016/j.aap.2018.01.008.
Thal, S., Znamiec, H., Henze, R., Nakamura, H., Imanaga, H., Antona-Makoshi, J.,
Uchida, N., Taniguchi, S., 2020. Incorporating safety relevance and realistic
parameter combinations in test-case generation for automated driving safety
assessment. In: in: 2020 IEEE 23rd International Conference on Intelligent
Transportation Systems (ITSC). Presented at the 2020 IEEE 23rd International
Conference on Intelligent Transportation Systems (ITSC), pp. 1–6. https://doi.org/
10.1109/ITSC45102.2020.9294724.
UNECE, 2021. UN Regulation No. 157. Uniform provisions concerning the approval of
vehicles with regard to Automated Lane Keeping Systems. Available online at
https://unece.org/sites/default/les/2021-03/R157e.pdf. Last access on 7
December 2021.
VMAD, 2019a. Safety Criteria Study on Innovative Safety Validation Methods of
Automated Driving System.
VMAD, 2019b. Safety Criteria study for New Assessment/Test Methods of Automated
Driving System (3rd Report).
Vogel, K., 2003. A comparison of headway and time to collision as safety indicators.
Accid. Anal. Prev. 35, 427–433. https://doi.org/10.1016/S0001-4575(02)00022-2.
Wang, C., Stamatiadis, N., 2014. Evaluation of a simulation-based surrogate safety
metric. Accid. Anal. Prev. 71, 82–92. https://doi.org/10.1016/j.aap.2014.05.004.
Wang, C., Xie, Y., Huang, H., Liu, P., 2021. A review of surrogate safety measures and
their applications in connected and automated vehicles safety modeling. Accid. Anal.
Prev. 157, 106157 https://doi.org/10.1016/j.aap.2021.106157.
Wang, J., Wu, J., Zheng, X., Ni, D., Li, K., 2016. Driving safety eld theory modeling and
its application in pre-collision warning system. Transp. Res. Part C: Emerg. Technol.
72, 306–324. https://doi.org/10.1016/j.trc.2016.10.003.
Wang, X., Alonso-Mora, J., Wang, M., 2022. Probabilistic risk metric for highway driving
leveraging multi-modal trajectory predictions. IEEE Trans. Intell. Transp. Syst. 1–14
https://doi.org/10.1109/TITS.2022.3164469.
Weng, B., Rao, S.J., Deosthale, E., Schnelle, S., Barickman, F., 2020. Model Predictive
Instantaneous Safety Metric for Evaluation of Automated Driving Systems. arXiv:
2005.09999 [cs].
Wolf, M.T., Burdick, J.W., 2008. Articial potential functions for highway driving with
collision avoidance. In: in: 2008 IEEE International Conference on Robotics and
Automation. Presented at the 2008 IEEE International Conference on Robotics and
Automation, pp. 3731–3736. https://doi.org/10.1109/ROBOT.2008.4543783.
Xie, G., Qin, H., Hu, M., Ni, D., Wang, J., 2019. Modeling discretionary cut-in risks using
naturalistic driving data. Transp. Res. Part F: Trafc Psychol. Behav. 65, 685–698.
https://doi.org/10.1016/j.trf.2017.11.022.
Xiong, X., Wang, M., Cai, Y., Chen, L., Farah, H., Hagenzieker, M., 2019. A forward
collision avoidance algorithm based on driver braking behavior. Accid. Anal. Prev.
129, 30–43. https://doi.org/10.1016/j.aap.2019.05.004.
Xu, Z., Zhang, K., Min, H., Wang, Z., Zhao, X., Liu, P., 2018. What drives people to accept
automated vehicles? Findings from a eld experiment. Transp. Res. Part C: Emerg.
Technol. 95, 320–334. https://doi.org/10.1016/j.trc.2018.07.024.
Zadeh, L.A., 1965. Fuzzy sets. Inf. Control 8, 338–353. https://doi.org/10.1016/S0019-
9958(65)90241-X.
Zhang, T., Tao, D., Qu, X., Zhang, X., Lin, R., Zhang, W., 2019. The roles of initial trust
and perceived risk in public’s acceptance of automated vehicles. Transp. Res. Part C:
Emerg. Technol. 98, 207–220. https://doi.org/10.1016/j.trc.2018.11.018.
Zhao, X., Li, Q., Xie, D., Bi, J., Lu, R., Li, C., 2018. Risk perception and the warning
strategy based on microscopic driving state. Accid. Anal. Prev. 118, 154–165.
https://doi.org/10.1016/j.aap.2018.02.012.
K. Mattas et al.