Collision Avoidance: A Literature Review on Threat-Assessment Techniques

Abstract

For the last few decades, a lot of attention has been given to intelligent vehicle systems, and in particular to automated safety and collision avoidance solutions. In this paper, we present a literature review and analysis of threat-assessment methods used for collision avoidance. We will cover algorithms that are based on single-behavior threat metrics, optimization methods, formal methods, probabilistic frameworks and data driven approaches, i.e., machine learning. The different theoretical algorithms are finally discussed in terms of computational complexity, robustness and most suited applications.

Full text

1
Collision Avoidance: a Literature Review on
Threat-Assessment Techniques
John Dahl, Gabriel Rodrigues de Campos, Claes Olsson, and Jonas Fredriksson
Abstract—For the last few decades, a lot of attention has
been given to intelligent vehicle systems, and in particular to
automated safety and collision avoidance solutions. In this paper,
we present a literature review and analysis of threat-assessment
methods used for collision avoidance. We will cover algorithms
that are based on single-behavior threat metrics, optimization
methods, formal methods, probabilistic frameworks and data
driven approaches, i.e., machine learning. The different theoret-
ical algorithms are finally discussed in terms of computational
complexity, robustness and most suited applications.
Index Terms—Threat-assessment algorithms, decision-making
methods, intelligent vehicles, active-safety systems,
I. INTRODUCTION
IN order to help drivers deal with complex traffic situa-
tions, several types of automated safety systems have been
developed during the last decades. These Advanced Driver
Assistance Systems (ADAS) have revolutionized the auto-
motive industry and better solutions are continuously being
introduced in new passenger cars. Leveraging arrays of sensors
that detect nearby objects and obstacles, these features can
keep the vehicle within the lane, maintain a safe distance,
and even avoid collisions in several critical situations. They
are also included in the safety rating procedures of different
national New Car Assessment Programmes (NCAP) in a effort
to encourage automotive manufacturers to build even safer cars
and, ultimately, reduce road fatalities worldwide.
In complex traffic scenarios, situation assessment is
paramount for an effective automotive safety system. An
illustration for human-driving and ADAS process is given in
Fig. 1. One can see that: i) the human-driving process in-
cludes Perception,Driver Intention and Driver Action submod-
ules; ii) the ADAS process comprises Sensing & Estimation,
Threat-assessment/Decision-making and Actuation functional-
ities. Naturally, ADAS processes (also referred to in some
literature as semi-autonomous systems) are design to be an
abstraction of the human-driving process, and great progresses
have been made to enlarge the range and complexity of the
scenarios handled today. Nevertheless, a major theoretical
challenge in the presence of human drivers remains how
to precisely distinguish a safe (though perhaps aggressive)
driving behaviour from an unsafe one.
In this paper, we will focus our attention solely on Threat-
Assessment (TA) and Decision-Making (DM) aspects for
J. Dahl, G. Rodrigues de Campos and C. Olsson are with Zenuity, Gothen-
burg, Sweden. e-mail: john.dahl@zenuity.com, gabriel.campos@zenuity.com,
claes.olsson@zenuity.com.
J. Dahl and J. Fredriksson are with the Department of Electrical Engi-
neering, Chalmers University of Technology, Gothenburg, Sweden. e-mail:
jonas.fredriksson@chalmers.se.
Fig. 1: Driving task: an engineering-designed ADAS process
versus a human-driving process. HMI stands for Human-
Machine Interface, and can take different forms (e.g., sound
feedback, haptic feedback and displays).
semi-autonomous systems. While some literature review pa-
pers on motion prediction algorithms exist, see [1] and [2]
for example, there is not, to the best of our knowledge,
any major survey paper on threat-assessment. We will also
include in our discussion pertinent TA-approaches designed for
fully Automated Driving (AD), that unlike ADAS systems are
aimed at guaranteeing perpetual safety in absence of a skilled
driver and in all type of circumstances. We propose here a
wide scope article on threat-assessment and decision-making
approaches, covering research areas such as safety verifica-
tion of dynamical systems, scheduling-based approaches, set-
based algorithms and machine learning-based methods. More-
over, we will leverage in the discussions from an industrial
perspective and experience on the design of critically safe
systems. Note that vehicle control or path planning-related
aspects are excluded of our discussion, even if they present
fundamental theoretical problems worth of discussion and a
deeper study. The reader is nevertheless referred to [1]–[3] for
a recent overview of motion planning and control techniques
for automated vehicles.
The remaining of the paper is organized as follows. Section
II presents a literature review on TA-methods, while Section III
summarizes the TA-methods from an application and actuation
point of view. Section IV discusses the different methods in
terms of computational complexity, robustness and potential
for future developments.

2
II. TH RE AT-ASSESSMENT METHODS
In the following we provide a literature overview on existing
TA-methods. Methods found in the literature can be divided,
on a higher level, into physical model-based methods and data
driven-based methods. The main difference among the two
classes is that in the first one, threat-assessment and decision-
making are made based on learning from physical insights and
models, while data driven methods rely on data-based learning,
i.e., black-box modeling. Furthermore, physical-based meth-
ods can be further divided into several categories, and in this
paper we divided such methods into the following categories:
single-behavior threat metrics (SBTM), optimization-based
methods, formal methods and probabilistic approaches. While
the classification of the different works mentioned in this paper
is not always straightforward or unambiguous, we have tried
to objectively classify them based on the inherent decision-
making process. If the threat-assessment is made assuming a
single future behavior of the different traffic participants, or if
in some cases a closed-form expression can be found, it will
fall into the category of single-behavior threat metrics. If the
decision-making is a result from an optimization problem it
will fall into the optimization-based methods category. If it is
possible to find formal guarantees regarding the presence of
a threat, these methods will fall into the category of formal
methods (including set-based approaches). If the decision-
making is made on a probabilistic setup, it will fall into
the probabilistic methods category. Finally, if the decision-
making is data-driven, such works will be lumped together
into one separate category. It is worth noting that sometimes
the borderline between different categories is subtle, and there-
fore overlaps between the chosen categories are unavoidable.
Hence, some papers can naturally fit into several categories.
A. Single-Behavior Threat Metrics (SBTMs)
In presence of perfect measurements and a comprehensive
knowledge of the intention of each participant (e.g., its destina-
tion), it can be possible to retrieve an exact deterministic mo-
tion prediction using model-based state propagation. Naturally,
this is rarely the case in reality, and simplifications are usually
introduced in the problem formulation. For example, it is
reasonable to assume that vehicles will continue to follow the
current path during a certain time window, i.e., only a single
future behavior is considered. Another simplification can be
to assume that the measurement data is noise-free. Hence, by
using reasonable simplifications one can reformulate the TA
problem as being driven by threat metrics based on single
future behaviors of the different traffic participants. This type
of Threat Metrics (TM) can be defined in any domain such
as, e.g., time, distance, or acceleration, and can be used to
identify hazardous situations.
1) Time domain TMs: A vastly used TM is Time-to-
Collision (TTC), which represents the time until a collision
between two objects occurs. In the literature, TTC values are
normally used as a threshold in the decision-making process
for enabling a warning or an automatic intervention, see [4]–
[6]. While there exist several approaches to derive TTC, the
most frequently used one computes it as the distance to the
closest object divided by the relative velocity. A more general
definition of TTC can be found in [7], where TTC is calculated
with respect to the relative distance and relative velocity,
given a constant relative acceleration. Using the intuition
behind TTC, [8] used Time-to-Lane Crossing (TLC) for lane
departure warning systems.
Several similar metrics have also been proposed by different
authors. For instance, the Inverse TTC (ITTC). ITTC is
naturally defined as the inverse of TTC and therefore increases
with the risk of collision, see e.g., [9]. For example, the inverse
ITC concept was used in [10] for the derivation of a forward
collision warning system, based on an experimental data set
on maneuvers where drivers execute last-second braking and
steering maneuvers.
As an alternative to TTC, Inter-Vehicle-Time (IVT) was
also introduced in [11]. It is defined as the time it takes for
the host vehicle, given the current host’s velocity, to travel
the distance equal to the relative distance with respect to the
obstacle ahead. Thus, a short relative distance and a high
host velocity yields a low IVT time. An important difference
between the IVT and TTC is that IVT is calculated based
on the host’s velocity while TTC is calculated based on the
relative velocity. Hence, as highlighted in [11], IVT is a
good complement in situations where TTC metric falls short
in identifying the threat. For example, in cases where two
vehicles are traveling at the same velocity but close to each
other, TTC will attain large values even though the situation
is hazardous. Another relevant TM is the Time-to-Manoeuvre
(TTM) (also referred to as Time-to-X), which is the time until
an automated manoeuvre must be initiated to avoid a collision,
e.g., Time-to-Brake (TTB), Time-to-Steer (TTS) and Time-
to-Kickdown (TTK). Moreover, Time-to-React (TTR) defines
the last point in time at which an evasive trajectory still exists.
For instance, [12] have used a set-based approach to determine
TTR, see Section II-C for a more detailed description. Another
approach is presented in [13], where a modified binary search
algorithm computes approximated values for different TTMs
to derive the TTR for multiple object scenarios.
2) Acceleration domain TMs: Some common acceleration-
based metrics include for example the Brake Threat Number
(BTN) and the Steering Threat Number (STN), defined as the
ratio between the longitudinal (lateral) acceleration required
to avoid a collision, and the maximum longitudinal (lateral)
acceleration, respectively. Hence, if the required deceleration
reaches the maximum achievable threshold, then the ratio
is equal to one, a ratio higher than one corresponds to an
unavoidable collision, i.e., collision mitigation solutions are
the only option.
The authors of [7] and [14] computed the required ac-
celeration to avoid a collision based on constant curvature
escape paths while [15] assumed manoeuvres with constant
jerk (third time-derivative of the distance). Using a recursive
search-tree approach, [16] focused on collision avoidance in
a multi-vehicle setup, relying on the assumption that the ego
vehicle’s optimal path (i.e., with the least lateral acceleration)
is either going straight forward or it will always tangent at
least one of the objects. A real-time implementation of this
algorithm was later presented and validated in [17], [18].

3
Working with both BTN and STN metrics, [19] presented an
estimation of worst-case performance for collision avoidance
systems. In particular, the authors derived worst-case deci-
sion timing and identified scenarios that are guaranteed not
to exhibit unnecessary interventions. The authors considered
both steering and braking maneuvers, assuming measurement
errors, non-ideal state prediction, as well as sensor and actuator
delays. More recently, [20] proposed the notion of predictive
steering threat number (PSTN), using a bicycle model with a
time-varying lateral acceleration to represent the dynamics of
the vehicle, what the authors claimed to be a more realistic
modelling when compared to commonly used models in STN-
related works. The PSTN has been used to design an adaptive
Collision Warning Algorithm (CWA) based on an adaptive,
speed-dependent warning threshold.
3) Distance domain TMs: A commonly used distance-
based metric is the Minimal Safe Distance (MSD), defined
as the minimum distance to be kept between the host and the
obstacle, see [11]. This metric is aimed for situations where
spatial margins are important, e.g., when queuing at a traffic
light or in a traffic jam. A review on different distance domain
metrics is given in [21].
4) Multi-domain TMs: In complex scenarios, however, a
single TM may not be enough to fully characterize a situation,
and therefore multiple TMs may be needed. For example, let
a scenario consist of two vehicles traveling next to each other
with approximately the same velocity. If only TTC is used, the
threat level will be low (a high TTC) even if the inter-vehicle
distance decreases to a minimum. Hence, the author of [11]
proposed to use a combination of TMs ( e.g., TTC, IVT and
MSD) to better reflect the real threat level.
B. Optimization methods
Dynamic optimization has become a standard tool for
decision-making in a wide range of areas, able to provide
fuel-efficient rocket controllers or operational-efficient control
strategies for complex chemical production facilities. Indeed,
many practical problems can often be expressed as dynamic
optimization problems, and a popular optimization framework
widely used in the literature is Model Predictive Control
(MPC). It consists of an optimization problem minimizing
an objective cost function subject to constraints such as a
dynamical model and boundaries on the states and/or inputs.
The optimization is performed over a finite-time horizon and
the optimal solution yields the best control law for the given
cost function. While optimization-based approaches may often
require high computation power, it is possible in some cases
to leverage the structure of the problem to retrieve an explicit
control law. This can allow off-line pre-computation of the
explicit feedback policies, reducing the on-line computation
in receding horizon control setup to a function evaluation,
therefore avoiding the on-line solution of complex optimiza-
tion programs. This is of particular interest for safety-critical
and time-critical constrained applications such as automotive
threat-assessment algorithms. A detailed introduction to MPC
can be found in, e.g., [22] and [23].
In [24] and [25], an MPC-based method was introduced
to assess the threat level for forward collision avoidance,
i.e., where another vehicle is ahead of the host vehicle. This
approach relies on the calculation of the minimum front
wheel angle necessary for collision avoidance using an optimal
control framework. The solution is also assumed to be able
to violate some of the constraints, as the problem’s hard
constraints have been relaxed to guarantee the feasibility of the
MPC problem. In order to cope with more complex situations,
the lateral acceleration metric is combined with a different
metric defining how much constraints are violated, such that
an intervention is only triggered if the level of violation of
the constraints or the lateral acceleration exceeds a given
threshold. In [26], the authors combined a look-ahead driver
model and a vehicle model, and used an MPC controller
to compute the minimum steering action necessary to avoid
an obstacle. In order to cope with modelling uncertainties
on the driver’s behaviour, such uncertainties are considered
as probabilistic constraints in this work. By computing an
upperbound on the uncertainties’ deviation, the constraints on
the MPC problem can be refined according, which yields a
robust control scheme.
The authors in [27] proposed a grid occupancy and opti-
mization based method, that can be used to find the optimal
braking point for collision avoidance at low speeds, e.g., for
parking applications. For conflict resolution at traffic inter-
sections, [28] leveraged the structure of the problem in order
to derive a hierarchical decomposition of the original opti-
mization problem where the central coordination problem is
separated from the local optimal control problems on each ve-
hicle, which the authors claim to significantly reduce demands
on computational capabilities and information exchange. More
precisely, assuming that vehicles are following a predefined
path through the intersection, the combinatorial part of the
problem (i.e., defining the vehicle crossing order and collision
free time slots that are feasible under the vehicle dynamics
and physical constraints) is separated from the problem of
finding the appropriate control inputs for a given crossing
order. Partly based on this decomposition, [29] focused later
on the properties of the underlying coordination problem. The
authors formulated a finite-time optimal control problem and
proposed a primal decomposition, showing that the problem
can be efficiently tackled using a standard sequential quadratic
programming such that most computations can be performed
in a distributed manner by the vehicles. More recently, [30]
tried to tackle some of the communication aspects of a dis-
tributed Sequential Quadratic Programming (SQP) framework
for solving the optimization problem. In order to reduce the
communication time and burden, the authors proposed an
asynchronous algorithm where the sensitivity of the optimal
control problem are only updated for a subset of all agents
at each step, and show how one can decide on this subset
of agents so to optimize the contraction properties of the
algorithm.
C. Formal methods
In complex situations, a great challenge in the design
of threat-assessment/decision-making systems is to determine
whether a situation can evolve to a threatful situation, espe-
cially when subject to complicated requirements composed

4
of safety, task sequentiality and restrictiveness arguments, for
example. Hence, a lot of attention has been given in the
literature to formal methods as a way to handle these aspects.
The idea behind formal methods is that performing appropriate
mathematical analysis can contribute to the reliability and
robustness of a system, and allow the derivation of correct-
by-design control systems. Such methods have been greatly
used in computer science, specifically in software engineering
and hardware engineering. More recently, researchers from
the control theory communities have also been increasingly
interested in combining control-theoretic tools for complex
physical systems with formal methods for accommodating
complex specifications, see [31] and [32]. Hence, the literature
in this field is vast and covers different approaches, from
temporal logics [33] and robust control techniques [34], to
set-based and supervisory control approaches [35] and [36].
Within the scope of this article, several formal approaches
for autonomous and semi-autonomous vehicles exist in the
literature, and usually fall into one of two sub-categories:
logic-based and set-based approaches.
1) Logic-based approaches: Logic-based approaches focus
on formalizing a requirement by translating it into logical
sentences, and are aimed at verifying the design of a sys-
tem instead of specifying complex requirements. In [37], the
authors proposed a distributed control system for collision-
free highway driving, where each vehicle is controlled by
adaptive cruise control. The authors also propose a formal
safety proof for the proposed control approach, by leveraging
hybrid systems models and concepts (e.g, quantified hybrid
programs (QHP) and quantified differential dynamic logic). An
interesting aspect of the proposed safety verification approach
is that the safety proof structure is strictly modular, which
reduces the proof to modular stages that can be verified
without the details in lower levels of abstraction. In [38]
and [39], the authors used Multi-Lane Spatial Logic (MLSL)
for proving traffic safety on multi-lane roadways and double-
sided traffic country roads, respectively. A very interesting idea
behind this set of works is that the authors separate the purely
spatial reasoning from the car dynamics for safety analysis,
such that the control layer is given by spatial properties
formalized in MLSL. Inspired by Interval Temporal Logic,
Duration and Shape Calculus, the MLSL approach presented
in [38] can be seen as a two-dimensional extension of interval
temporal logic. This yields that safety amounts to proving that,
under certain assumptions, the occupancy of certain parts of
the road are always disjoint. The abstract traffic model for
multi-lane motorways given in [38] was later refined by taking
traffic directions into account in order to extend this approach
to country roads with two-way traffic in [39]. A formal design
and verification methodology for cooperative driver assistance
systems have also been proposed in [40], that formally verifies
timed probabilistic requirements on the successful completion
of a safe (i.e., collision-free) driving task. Finally, [41] focused
on early specification phases. The authors used Higher Order
Logic (HOL), a purely logical environment, to formalize traffic
rules, and showed that it is possible to formally check the
compliance of traffic rules by autonomous vehicles.
2) Set-based approaches: In set-based approaches, on the
other hand, requirements are often formalized by specifying
a set of acceptable/unacceptable behaviours or system con-
figurations. For example, [42] proposed a hybrid architecture
for collision avoidance at intersections that guarantees safety.
This approach consists of an interaction between a central-
ized component, i.e., a scheduler that assigns a time slot to
vehicles, and distributed vehicles that need to determine if
they can go through the intersection in the assigned time slot
without violating perpetual safety of the system. By leveraging
the structure and the ordering properties of the intersection
coordination problem, the authors build their safety concept
around an infinite sequence of safety inputs that each vehicle
possesses, which plays the role of an infinite horizon contin-
gency plan.
In [43] the safety of automated vehicles is formally verified
by predicting the set of all possible occupancies of the
automated vehicle as well as other traffic participants. The
authors apply reachability analysis using mathematical models
subject to uncertain inputs (e.g., sensor noise and disturbances)
and uncertain initial states, and verified their results on real
automated vehicles. Using identical concepts, [44] presented
a method for formal collision avoidance detection on arbitrary
road networks, where the existence of a collision assessed
by checking if the host’s occupancy set intersects with the
occupancy set of any other object. State constraints are used
to limit the expansion of the occupancy set and, should an
object violate a specific constraint, such constraint is removed
and the object’s occupancy set recomputed accordingly. The
same research group also described in [45] a new tool that uses
reachability analysis to predict the occupancy of surrounding
traffic participants, named SPOT for Set-Based Prediction Of
Traffic Participants. More precisely, the authors focused on
the problem of how to ensure safe motion plans and consider
all possible manoeuvres (e.g., full throttle or full braking),
physical constraints, as well as traffics rules that the traffic
participants are assumed to obey. Given these assumptions,
the obtained results are provably an over-approximation with
respect to the exact set of possible occupancies, which yields
that the SPOT toolbox can be used for instance to verify
intended trajectories, as the underlying approach is safe.
Another interesting aspect of this work is that the authors
also describe how one can react to observed violations of
assumptions such as, for example, when an illegal lane change
is performed. Recently, the same authors have proposed in [12]
a novel approach using reachability analysis to derive Time-
to-React (TTR). The authors show how over-approximations
of the TTR can be efficiently computed for different urban
and rural traffic scenarios, and compare their approach to an
optimization-based estimation technique.
Using identical concepts, [46] also dealt with safety veri-
fication of motion plans, and focused in particular on cases
where (previously safe) motion plans becomes unsafe due to
the violation of assumptions, i.e., a misbehavior of dynamic
obstacles. To do so, the authors derived invariably safe sets in
order to determine the Point of No Return (PNR) and the Point
of Guaranteed Arrival (PGA). These concepts allowed the
authors to divide motion plans into safe sections and safety-

5
critical passageways, where in the latter the system is exposed
to potential collisions if obstacles violate assumptions used for
verification. Most interestingly, the authors developed in this
paper the concept of safety costs that allows the minimization
of safety-critical passageways by assigning costs to paths
before their execution. By integrating the cost function into
the optimization algorithm, the planner directly determines
the safest trajectory possible and ultimately trajectories with
a passageway of size zero, i.e., trajectories that guarantee
safety for an infinite time horizon. For validation purposes,
the authors used the previously mentioned SPOT toolbox to
demonstrate their approach for overtaking manoeuvres on a
double-sided traffic road.
In [8] and [47], a group of researchers from Chalmers Uni-
versity of Technology used reachability theory to determine
the likelihood of a collision. Inputs and states are assumed to
be bounded, e.g., constraints on steering angle, steering angle
rate, lateral distance to lane markings, slip angle, etc, in order
to reflect a regular driving behaviour. The goal is to determine
if vehicles can reach a pre-defined safe set without violating
either input or state constraints, using reachability tools to
identify the safe, invariant set. Naturally, if the current state
belongs to the invariant set, the situation is considered to be
safe and no intervention is to be triggered, while in the other
cases a suitable assisting control policy is requested. Within
the same department, other researchers described in [48] a con-
flict resolution technique for traffic intersections that combines
optimal control with sequential decision-making. In particular,
the authors used reachability theory to derive model-based
heuristics, characterizing for instance each vehicles degree of
freedom for avoiding collisions, and present a coordination
scheme that scales linearly with the number of participants.
The authors exploited the notion of a decision order, firstly
introduced in [49], based on which vehicles sequentially solve
local optimal control problems. This yields that each vehicle
avoids collisions by adapting to the already computed plans
by vehicles preceding it in the order. A receding horizon
implementation of the above mentioned strategy was also
studied in [50], where the authors divided the decentralized
solution of the local optimization problems in two parts: a
finite-time problem where collision avoidance is enforced as
terminal constraints, and an infinite horizon problem defining
the cost-to-go that can be calculated offline.
A lot of attention has recently been given to formal veri-
fication and supervisory control for collision avoidance, and
in particular for traffic intersections, see e.g. [36], [51]–[58].
In particular, a set-based approach for the design of a least
restrictive supervisor has extensively been used by MIT’s and
Politecnico di Milano groups, see [36], [54]. A least restrictive
supervisor is defined as a control algorithm that ensures the
safety of the requested input (either the command of a driver or
potentially from an autonomous vehicle’s driving logic), while
intervening as little as possible (least restrictiveness) [36].
The system takes as input the user-requested command and
returns either (i) the same command if it is considered to be
safe or (ii) an overriding, safe-by-design control command.
In practice, identifying the set of safe manoeuvres equates
with determining the largest set of states for which there
MCIS
Forbidden
setCapture
set
Safe
trajectory
Inevitable collision
trajectory
Fig. 2: Illustration of a formal verification approach for
threat-assessment. Here, the Forbidden Set corresponds to a
configuration corresponding to a collision, the Capture Set to
all configurations from which the system will inevitably reach
the Forbidden Set (i.e., set of configurations corresponding
to a collision), and the MCIS denotes the Maximum control
invariant Set (i.e., the set for which it exist at least one
control input that allows collision avoidance for all future time
instants).
exists a control input (e.g., throttle or steering input) that
can avoid collisions. This set is commonly referred to as the
Maximal Controlled Invariant Set (MCIS), the reader can refer
to [59] and [60] for further details. The approaches proposed
in [54] and [36] are based on the solution of two separate
problems: (i) the Verification Problem, that determines if there
exists an input signal that leads all vehicles safely through the
intersection; and (ii) the Supervisor Problem, returning a safe
overriding control signal when safety conditions are violated.
From a control perspective, the solution to the Verification
Problem can be solved by determining the membership to the
MCIS, see Fig. 2 for an illustration of the MCIS concept. As
illustrated in Fig. 2, the notion of MCIS is closely related
to the concepts of Forbidden Set, which corresponds to all
configurations representing a collision, as well as the Capture
Set (sometimes also called Attraction Set in some literature),
i.e., the set that incorporates all configurations from which the
system will inevitably reach the Forbidden Set or, in other
words, will result in a collision. It is important to notice
that determining the MCIS is often a computationally difficult
problem, and it has been proved to be NP-hard in the case
of some collision avoidance problems, see [54] for further
details. Hence, even if standard general algorithms exist in the
literature, they are limited by numerical complexity to handle
problems involving just a few road users (typically two).
A set of efficient solutions for the intersection collision
avoidance problem was proposed in [54], leveraging the equiv-
alence between the Verification Problem and a Scheduling
Problem, which consists in assigning jobs to a resource while
satisfying given requirements [61]. The Verification Problem
can be written as a scheduling problem where the intersection
represents the resource, the vehicles represent the jobs to be
assigned to the resource, and the time spent by each vehicle
in the intersection is the length of the job to be executed.
Hence, determining the safety of the system can be posed
as a control problem determining the existence of a crossing
schedule, wherein the objective corresponds to minimizing

6
the total time needed to clear the intersection. Based on the
results of [36], several papers have leveraged such equivalence
for the design of supervisors for more complex scenarios
handling, for instance, disturbances and measurement noise
[56] or uncontrollable vehicles [57]. Optimality arguments
are also introduced in the design of the supervisor in [62].
The algorithm presented in [62] is able to identify the op-
timal corrections to a human-decided input, which allows
more driver-friendly and less aggressive manoeuvres. More
recently, the authors extended their results to a road-network
of interconnected intersections, see [63]. The road-network
verification problem is decomposed in more treatable sub-
problems, and the authors show how the above mentioned
results for a single intersection can be applied to handle
such clearly more complex scenarios. Using similar ideas,
the authors of [64] also employed formal control theoretical
methods for hybrid-systems in order to guarantee collision-free
intersections. In particular, the authors leverage the properties
of systems that evolve on a partial order and whose dynamics
preserve the ordering. For instance, vehicles approaching an
intersection are guaranteed to maintain a certain order within
the same lane, as vehicles cannot go over each other. The
authors showed how such class of systems admits an explicit
solution to the safety problem, and proposed linear complex-
ity discrete-time algorithms with guaranteed termination. An
experimental implementation of such systems has also been
described in [65]. A set of works addressed the problem of
multi-vehicle collisions based on abstraction techniques. For
instance, [66] provided a recent overview over several works
from a research group at the University of Michigan, where
the authors constructed a discrete-event system abstraction,
and formulate the problem in the framework of supervisory
control for discrete-event systems with uncontrollable events,
which allows the mitigation of the computational limitations
related to the presence of continuous dynamics and infinite
state spaces.
Finally, some other works have also combined formal
methods with optimization approaches (discussed in the next
section). For instance, [67] provided a new approach com-
bining set- and optimization-based methods for formal path-
planning for autonomous vehicles. More precisely, the authors
model vehicles as differential inclusions composed of simple
dynamics and a set-based uncertainty formulation, and obtain
reachable sets through an interesting combination of optimiza-
tion techniques and reachability analysis for control synthesis
purposes. From a systems and control perspective, [68] fo-
cused on the robust model predictive control for constrained
systems, proposing an approach combining optimization-based
techniques and set-based concepts. In particular, the authors
focused on constrained linear discrete-time systems that are
subject to state and measurement disturbances, and the main
idea behind this work is to consider the state estimation error
as an additional unknown uncertainty, that is bounded by a
simple, precomputed invariant set.
D. Probabilistic methods
The underlying idea behind probabilistic TA-approaches is
to leverage available information of the system’s uncertainties
t1t2t3
t2t1
t3
Fig. 3: Illustration of a probabilistic threat-assessment. The
figure shows an oncoming scenario, with the host vehicle in
silver and the oncoming vehicle in red. Vehicles are approxi-
mated as point-of-mass (solid orange circles). Illustrated by the
blue and yellow ellipses is the uncertainty associated with the
predicted future occupancies, that grows with the prediction
time tk.
to make decisions with a certain level of confidence. For
the sake of clearness of the discussion, the remaining of this
section is organized in the following parts: subsection II-D.1
focus on probabilistic TA-methods, while subsection II-D.2
discusses probabilistic decision-making.
1) Probabilistic analysis: Generally speaking, a probabilis-
tic threat-assessment method assigns probabilities to different
events, e.g., how likely it is to collide with another object
in a near future given some assumptions on uncertainties.
In automotive applications, some of the major sources of
uncertainty include dynamical modelling errors, measurement
noise and the misinterpretation of the drivers’ intention. The
interested reader can see Fig. 3 for an illustration of the
working principles of a basic probabilistic TA.
While there is an infinite number of actions that a driver
can take (and that are in general hardly predictable), each
action can be linked to a small set of high-level manoeuvres
such as performing a lane change, overtake, etc. This principle
has been used in [69], for instance. By posing the problem
as determining the probability of high-level manoeuvres, one
can then reduce both the search space and the computational
complexity. Here, each object is assigned a goal function based
on different objectives, which is used to define a probability
distribution over the set of manoeuvres for each object. The
probability of a collision is computed via Monte Carlo sim-
ulations, where each object is simulated using samples from
respective maneuver distribution. This method has later been
improved in [70], where the authors considered the driver’s
awareness of other objects’ location, e.g., if the object is
situated behind, ahead, or at either sides of the host vehicle.
It is also assumed that, for example, a driver is paying more
attention to an object ahead rather than behind the ego vehicle.
Based on this work, [71] improved the dynamic vehicle model
in order to reduce modelling mismatch, as the model used
in [70] suffered from a biased mean value for the lateral
and longitudinal acceleration. More recently, [72] introduced
a road-aligned curved coordinate system, aimed at simplifying
the modelling when the road bends. Another work utilizing a
curved coordinate system, see [73], assessed the risk in terms

7
of a probabilistic TTC level, based on uncertainties regarding
occupancy of the adjacent lanes.
Focusing on predicting the occupancy for traffic partici-
pants, [74] proposed a probabilistic approach using a Markov
chain abstraction, considering measurements uncertainties as
well as other traffic participants’ unknown intentions. Addi-
tionally, the road geometry as well as the interaction among
the traffic participants are also taken into consideration, and the
probability of a crash is computed given a defined trajectory
for the host vehicle. This work was later extended in [75]
where Markov chain abstraction and Monte-Carlo simulations
are compared in terms of computational performance and pre-
cision. According to the analysis, the Markov chain abstraction
is superior for probabilistic occupancy prediction while the
Monte-Carlo simulations is the best choice to derive the colli-
sion risk. A different probabilistic motion prediction algorithm,
based on an Unscented Kalman Filter (UKF) combined with
a reachability based decision-making protocol, has also been
given in [76] to decide on an emergency intervention for
intersection scenarios. A safety intervention is based on two
different but naturally correlated thresholds: one defining the
degree of confidence on the model based predictions; the other
representing the probability of an unavoidable collision.
Other works have used Bayesian approaches. For example,
[77] and [78] used a Dynamical Bayesian Network (DBN)
to compute the risk that the driver is missing to stop when
approaching an intersection. In scenarios with multiple objects,
it is typically the case that an action of one object will cause a
reaction of the other participants. Hence, each driver’s action
is conditioned by the behavior of the other participants, which
increases the complexity of the prediction task. The authors of
[79] used a DBN to model the physical relationship together
with the conditional drivers behavior of the objects in the
scenario. To account for changes on the drivers’ actions, the
DBN is embedded in a Partially Observable Markov Decision
Process (POMDP), where each new action can be taken with a
certain probability. The DBN model parameters are learned by
utilizing an expected maximization approach using unlabeled
observations. In [80] the threat level is computed by combining
network level collision predictions with vehicle level collision
predictions using a DBN. On the network level, the threat is
assessed based on high-level traffic information, e.g., average
speed and traffic flow on a road segment, while on vehicle level
the threat is assessed based on individual motion predictions of
the surrounding vehicles. These results show that the vehicle
level assessment is improved in cases where the situation is
classified as hazardous by the network level assessor. Using
the principles of Bayesian Occupancy Filtering (BOF), [81]
and [82] estimated the future occupancy of different partici-
pants. The environment is represented as a 2-dimensional grid,
where each cell contains both static and dynamic information
about occupancy and velocity. In [81], the grid indicates the
occupancy by other objects and is updated via inference based
on sensor measurements. However, prediction performance
is poor in presence of temporarily occluded objects. This
is addressed in [82] by extending the method with prior
map knowledge so to increase the prediction performance for
occluded objects.
Some works have focused on computational efficiency as-
pects, which are crucial for real-time applications. On this
topic, [83] mixed set theoretical methods with a probabilistic
approach, and divided the threat-assessment problem into
a preliminary and a specialized part. The preliminary part
searches for potential collisions with other vehicles using a
geometric occupancy analysis. The occupancy of each object
is efficiently computed via segmented cones representing the
reachable set of states for predefined scenarios. The set of
objects that can cause a collision are thoroughly analyzed
in the specialized part, where the probability of collision is
computed using statistical inference.
2) Probabilistic decision-making: In general, it is not obvi-
ous how to trigger an automatized intervention in a stochastic
setup. However, using hypothesis testing, one can optimize
the triggering threshold and therefore minimize the risk of an
erroneous decision.
The authors of [7] and [84] derived a decision rule for
automated braking based on hypothesis testing for the required
lateral acceleration needed to avoid a collision, where the
estimates of the longitudinal dynamics are uncertain. The
Bayesian framework for the decision-making is generalized in
[85] to deal with any stochastic threat-assessment algorithm.
Furthermore, they show that the risk of a collision can be
computed online with numerical Monte-Carlo integration for
an automated braking system.
A similar approach has been used in [86] and [87], where
the threat-assessment is divided into two levels: one for a
physical system level threat-assessment, i.e., how difficult it is
for a vehicle to avoid a collision; and one for the assessment
on how the driver would perceive a potential intervention.
The authors argue that it might be possible to trig earlier
interventions based on less detailed data, if it can be be
ensured that the driver will most likely not be annoyed by
the intervention.
Finally, [88] has recently focused in behavior generation
aspects for automated vehicles, using motion planning tech-
niques that consider uncertain predictions of other road users.
The main contribution of this work relies on an online Partially
Observable Markov Decision Process (POMDP) algorithm,
which is able to incorporate uncertainties and provide op-
timized solutions for the on intersections with an arbitrary
layout and a variable number of traffic participants.
E. Data driven methods
Machine Learning (ML) methods recently became very
popular for several tasks within ADAS/AD applications, in
particular due to the rapid advances in computational hardware
and algorithms that are strongly supported by large companies
such as NVIDIA and Intel. The underlying idea behind ML
is to learn an environment/behavior/property purely based on
a data set, often denoted as the training set. More precisely,
ML algorithms approximate the relationship from one element
to another within the system, which are usually complex and
not formally described. If the training set is rich and large
enough, in the sense that all configurations of the system are
sufficiently excited/represented, a ML-approximated function

8
is expected to efficiently generalize with new data, e.g.,
verification data or real-time sampled data. In the context
of ADAS applications, ML approaches are usually used for
predictions where the input is the current state of the system
and the output a prediction of the future state. A popular
method within the ML family is the Artificial Neural Network
(ANN). For further details on learning methods, the reader can
refer to [89].
There are many different implementations of machine-
learning concepts. For example, the authors of [90] and [91]
used a technique based on Gaussian processes in order to
learn the drivers’ behavior at intersections and lane-change
scenarios, respectively. Furthermore, ANN, Recurrent Neural
Networks (RNN) and Support Vector Machines (SVM) have
been considered to predict the drivers’ behavior specifically
for lane-change scenarios, see [92] where it is shown that
SVM offers superior results with respect to RNN and ANN.
In a similar work, naive Bayes, SVM, logic regression, nearest
neighborhood, decision trees, extra trees and Random Forest
(RF) classifiers are compared in [93] for lane-change scenar-
ios, where it is concluded that the tree based classifiers are su-
perior with respect to the remaining approaches. Furthermore,
a RF algorithm is used in [94] to learn automated parking
manoeuvres, where the solution of the RF is approximated by
a General Radial Basis Function (GRBS) for enhanced real-
time capabilities.
Another important application for ANN is the automated
annotation of scenarios in logged data, see [95]. For example,
a given scenario is automatically classified by its properties,
i.e. ”the scenario is an intersection with a vehicle approaching
from the left side”.
For AD applications, supervised ANN has already been used
for End-to-End learning [96]–[98], where the objective is to
derive an actuation signal directly from the raw image input
obtained from a frontal camera. In other words, the goal for the
ANN is to mimic the behaviour of the driver given the training
data set. While End-to-End learning considers no assumptions
on the underlying problem, other works proposed instead to
use state estimates as inputs to an ANN. This approach is used
in [99], where an ANN is learned to detect unintended lane
departures.
III. APP LI CATI ON ,ACT UATI ON A ND VALIDATI ON
Section II presented a description of the different threat-
assessment methods and the underlying main ideas and prin-
ciples. In order to relate the reviewed works to different
automotive-related applications, a summary table is given in
Table I. The table is organized with respect to three arguments:
i) what type of application the works are aimed for; ii)
the actuation principle; and iii) how the results have been
validated. In particular, we identified five different collision-
prone scenarios: at intersections, frontal and rear-end collisions
(a vehicle is traveling in same direction as the host), lane
departures (LKA), and collisions with Vulnerable Road Users
(VRU). We also classified works with respect to the actuation
principle, i.e., relying on braking and/or steering actions, and
the validation approach separated in two categories, simulation
only and experimental (or simulations based on real-data).
Note that broad scope publications such as dissertations and
review papers, as well as literature related to automotive appli-
cations that do not fit with the above mentioned structure are
gathered at the end of the table. Furthermore, non automotive-
oriented literature focusing on general aspects of sensing,
control theory, or computer science, for instance, are excluded.
From Table I, one can see that a great part of the papers in
this literature review used probabilistic and formal methods.
Moreover, one can also observe that, among the cited litera-
ture, a lot of attention has been given to conflict resolution
at traffic intersections, as such scenarios incorporate the same
technical challenges such as round-abouts or on-ramp merging,
for instance. A significant part of the considered publications
provided experimental results, even if most of the works rely
on simulation-only results. It is important mentioning that,
naturally, the distribution of Table I and the discussion (see
next section) is not necessarily representative of all the existing
literature, but rather an overview of the papers analyzed in this
review.
IV. DISCUSSION
In the rest of the paper, we will discuss the different ap-
proaches in terms of system performance, robustness, real-time
properties and also challenges and implications of increased
autonomy.
A. System performance
While all methods can benefit from an efficient data-filtering
and sensor-fusion algorithms, the different approaches differ
on how environment data is used within the algorithms. For
instance, SBTMs usually only consider the most likely state
estimates, discarding in many cases the associated distribution
of the states. Instead, robustness with respect to uncertainties
are gained by putting margins in the decision-making stage,
that are typically tuned by empirically testing to maximize the
system performance.
Probabilistic methods, on the other hand, tend to make
use of the entire uncertainty model in order to estimate the
probability of a collision. Nevertheless, it is generally not
obvious how to derive the uncertainty model, especially in
cases of time-varying uncertainties. This is, e.g., the case for
uncertainties related to driver’s intentions. From this literature
review, it follows that probabilistic methods are preferred in
scenarios where the uncertainties can be modeled by a few
random variables, or in cases where the prediction horizon is
sufficiently short.
Regarding formal methods, the underlying objective is usu-
ally to verify and formally guarantee whether a situation can
evolve to an unavoidable dangerous situation. These methods
are best suited for applications where formal guarantees of
performance are required, or when the system is subject to
complex combinations of safety, comfort, or logic require-
ments, for example.
In data-driven approaches, the only underlying assumption
is that there exists a relationship between the inputs and
the outputs, i.e., the current system’s state can be mapped

9
TABLE I: An overview of the reviewed articles. Here, CA stands for Collision Avoidance, LKA for Lane Keep Assistance,
and VRU for Vulnerable Road Users.
Single-behavior TMs Optimization methods Formal methods Probabilistic methods Data driven methods
Application
CA with VRU [4] [13] [17] [69] [72] [86] [87]
[83]
CA at intersec-
tions
[5] [13] [27] [28] [29] [30] [12] [35] [36] [42] [45]
[48] [49] [50] [53] [54]
[55] [56] [57] [58] [62]
[64] [65] [66]
[74] [75] [76] [77],
[78] [79] [82] [83]
[88]
[79] [90]
Frontal CA [4] [5] [14] [15] [20]
[21]
[24] [25] [26] [12] [35] [39] [43] [44]
[45] [46] [52]
[69] [70] [71] [72]
[73] [74] [75] [86]
[87]
[91] [92] [93]
Rear-end CA [5] [6] [7] [9] [10]
[11] [16] [17] [18]
[19] [21]
[24] [25] [26] [12] [35] [36] [37] [38]
[39] [40] [43] [44] [45]
[46] [63] [67]
[7] [69] [70] [71]
[72] [73] [74] [75]
[84] [85] [86] [87]
[91] [92] [93]
LKA [8] [8] [35] [47] [96] [97] [98] [99]
Actuation
Steering control [9] [10] [11] [13]
[16] [17] [18] [19]
[20]
[24] [25] [26] [8] [12] [38] [39] [40]
[43] [44] [47] [52] [67]
[72] [79] [91] [94] [96] [97] [98]
Throttle/Braking
control
[6] [7] [9] [10] [11]
[13] [14] [15] [19]
[21]
[27] [28] [29] [30] [12] [36] [37] [38] [39]
[40] [42] [43] [44] [48]
[49] [50] [53] [54] [55]
[56] [57] [58] [36] [62]
[63] [64] [65] [66] [67]
[7] [76] [77] [79]
[85] [86] [87] [88]
[90] [91] [94]
Validation approach
Simulation only [14] [16] [17] [18]
[19] [20]
[26] [28] [29] [30] [12] [36] [37] [38] [39]
[40] [42] [45] [46] [48]
[49] [52] [54] [56] [57]
[62] [63] [64] [66]
[69] [70] [71] [79]
[83] [84] [88]
[79] [92] [99]
Experimental/
with real-data
[4] [5] [6] [7] [8]
[9] [10] [11] [13] [15]
[16] [21]
[24] [25] [27] [8] [35] [43] [44] [47]
[53] [55] [58] [64] [65]
[67]
[7] [70] [72] [73]
[74] [75] [76] [77]
[78] [79] [82] [85]
[86] [87]
[90] [91] [93] [94] [96]
[97] [98]
Related literature
Correlated topics
and broad-scope
publications
[1] [2] [3] [4] [7] [35] [41] [79] [80] [81] [95] [100] [101] [102] [103]
to a threat level. In the case of supervised learning (e.g.,
deep learning), a practical challenge is that the data used for
training the algorithm needs to be classified a-priori, either
by humans or by some automated logic. Furthermore, training
data-set must be persistently exciting (i.e., exciting all possible
modes of the system), which may not always be possible in
collision avoidance (i.e., there is often a discrepancy between
the amount of data on regular driving behaviors and collision-
prone behaviors). On the other hand, one of the strong benefits
of using machine-learning based methods is their ability to
capture complex nonlinear behaviors, which makes it, at least
in theory, suitable to find relationships that may be hard to
model using classical engineering principles.
B. Real-time properties
From an industrial perspective, the systems performance
is naturally not the only factor under consideration. Indeed,
any safety system needs to be cost-efficient, both during the
development phase but also upon deployment (e.g., real-time
computational requirements). There is therefore an emphasis
on adopting efficient but not unnecessarily complex methods.
Generally speaking, SBTMs are computationally cheap by
design. For set based approaches, recent works have proposed
a toolbox that is able to compute set based predictions in under
20 ms, see [45]. Other solutions, though, aim at leveraging
the problem structure to reduce the online computational
load or set and reachability computation, by solving sub-
parts of the problem offline during the development phase and
later deployed in vehicles with lower computational demands.
Regarding optimization based approaches, their computational
complexity usually scales poorly with the size of the problem,
which mean that only small problems can be efficiently solved
in real-time. Unfortunately, constrained MPC-problems tend to
be non-convex, which together with long prediction horizons
usually yields prohibitive complex problems.
Probabilistic methods may also lead to computational-
efficient solutions in some special cases where the noise
is strictly Gaussian and the dynamics linear, for example.
However, that is rarely the case in practice and the solution
often needs to be numerically approximated by Monte-Carlo
simulations or particle filtering, which are known to be accu-
rate but computationally expensive.
For data-driven algorithms, it is crucial that the training
data is sufficiently rich and large in size. However, large
networks with many layers and neurons, combined with com-
plicated activation functions, offer great challenges in terms
of computational power and memory allocation for real-time
deployment. Moreover, the training process itself is normally

10
computational expensive but can usually be done offline.
Leveraging the large database that automakers gathered over
the years from field tests and test track verification tests, deep-
learning has been attracting an increasing amount of interest
for ADAS and autonomous driving systems, in particular for
image segmentation1as illustrated in Fig. 4. But while the first
applications have shown promising results for lane-markings
and road sign detection, for example, their applicability to core
safety features is not straightforward and will require several
years of development.
C. Robustness
Another important aspect for any safety system is ro-
bustness, here defined as the measure of the system’s abil-
ity to handle uncertainties due to e.g., measurement noise
or deviations from other assumptions such as the modeled
driver behaviour. Formal methods have an intrinsic ability
to guarantee robustness, where the level of robustness can
be adjusted by over- or underestimating the uncertainties in
the set definitions. SBTM- and optimization-based methods,
on the other hand, fall short in terms of robustness, as a
result of over-simplifications made on the system modelling
and problem statement, see [104]. Imagine a solution to the
optimization problem that tangents the constraints for some
variables. In this case, the solution is feasible with respect
to the objective function, constraints and vehicle model. But
what if the model is uncertain and the state measurements are
noisy? To handle such issues, it is common practice to measure
the state after some time period and to solve the dynamic
optimization problem again starting from the new measured
state. By using feedback on the measurement information,
this provides the whole procedure with a robustness that is
typical for closed-loop systems, and usually referred to as
Model Predictive Control (MPC) that we discussed earlier
in the paper. To enhance robustness properties, extensions to
the MPC framework have been made in [105] and [106].
The underlying idea of Robust MPC is that the problem
contains one or several sources of uncertainties and several
techniques have been proposed to solve the problem such
as stochastic MPC [107] or scenario-tree MPC [108] and
[109]. In the first set of works the problem is defined as
a typical MPC problem, but where the constraints are only
fulfilled with a certain probability. In the second approach, for
problems with a finite number of realizations of uncertainties,
one can formulate a problem for each uncertainty realization
and structure them in a tree formation. Hence, the overall goal
with scenario-tree MPC is to find a control sequence that can
handle all realizations of the uncertainties and potentially using
a distributed computational scheme.
D. Challenges, implications and opportunities
Despite the substantial technical progress on sensing, com-
putation and communication capabilities in the last few years,
1Image segmentation consists of a process of partitioning a digital image
into multiple segments that share certain characteristics. Image segmentation
is particular interesting for automotive systems in order to simplify and/or
change the representation of an camera image into something that is more
meaningful and easier to analyze.
Fig. 4: Image segmentation with deep-learning algorithms,
which are able to identify areas of interest in front of the
vehicle: the drivable-road area is highlighted in blue, lane
markers in yellow, road edges in green and vehicles’ delimiters
in red.
the development and deployment of autonomous vehicles will
raise great technical and societal challenges in the decades
to come. A discussion on the such challenges but also future
opportunities is given in, e.g., [100] or [101]. In [100], the
authors discuss different necessary technologies, covering a
wide spectrum of aspects including in-car computing and data
management, road side infrastructures and cloud solutions.
Most interestingly, the authors particularly discuss the need
of high-definition maps, identified as core technology for
autonomous vehicles. The opportunities and future implica-
tions of autonomous vehicles for transportation policies are
also tackled in [101], where the authors provide a review on
relevant literature for aspects going from safety to machine
ethics. Finally, the interdisciplinary nature of the challenges
raised by increasingly autonomous safety systems is also
discussed in [102]. Here, the authors highlight the need for
a multi-disciplinary approach across software and hardware
engineering, testing or social and legal fields, and particularly
point out the challenges to validate machine-learning based
systems with respect to the ultra-dependable levels required
for autonomous vehicle fleets. The same group of researchers
also discussed in [103] the challenges in testing and validation
of autonomous vehicles, a corner stone for a large-scale
deployment of such technology.
V. CONCLUSIONS AND FUTURE PERSPECTIVES
In this paper we provide a literature review and a technical
discussion on the different methods used for threat-assessment
for automotive active-safety systems. We have covered a
wide range of publications focusing on intelligent and auto-
mated vehicles, and classified them in comprehensive cate-
gories based on the underlying threat-assessment and decision-
making process. It is worth highlighting that the discussions
driven in this paper are articulated around a decision dilemma,
according to which there exists a clear trade-off between
high performance (i.e., the ability to avoid a collision) and
precision (i.e., performing an avoidance maneuver only when
needed). While long prediction horizons and high intensity
interventions are important to properly handle very critical

11
situations, they may, when combined with uncertain estimates,
increase the number of unwanted/false interventions. Short
prediction horizon, on the other hand, may provide high
precision perception and situation awareness, but the driver’s
time to react is reduced and the effectiveness of an ADAS may
be limited. Consequently, making use of methods in such a
way that they best handle this intrinsic conflict, is likely one of
the main challenges for future generations of ADAS systems.
Despite the many comprehensive works reviewed here, there
is clearly still room for substantial improvements for future
research. For example, one of the take-home messages is
that the methods covered here have different strengths and
weaknesses that are dependent on the context. Hence, it is
tempting to think that it might be possible to combine different
methods to maximize the over all performance on a broader
set of situations, even though it is not obvious how this can
be achieved. Moreover, the drivers’ intention is also a crucial
aspect of all types of methods, since it determines a substantial
part of the future outcome. Techniques such as driver state
estimation using in-vehicle mounted cameras and vehicle-to-
vehicle communication, likely to enter the market soon, will
allow for new possibilities in terms of threat-assessment and
decision-making strategies.
ACK NOW LE DG ME NT S
This research was supported by Zenuity and the Swedish
Governmental Agency for Innovation Systems/FFI through
Contract 2014-05621. The authors are also grateful to the
reviewers and the editors for their efforts in reviewing and
improving this paper.
REFERENCES
[1] S. Lef`
evre, D. Vasquez, and C. Laugier, “A survey on motion predic-
tion and risk assessment for intelligent vehicles,” Robomech Journal,
Springer, no. 3, 2014.
[2] D. Gonz´
alez, J. P´
erez, V. Milan ´
es, and F. Nashashibi, “A review of mo-
tion planning techniques for automated vehicles,” IEEE Transactions
on Intelligent Transportation Systems, vol. 17, no. 4, pp. 1135–1145,
2016.
[3] B. Paden, M. Cap, S. Z. Yong, D. Yershov, and E. Frazzoli, “A survey of
motion planning and control techniques for self-driving urban vehicles,”
IEEE Transactions on Intelligent Vehicles, vol. 1, no. 1, pp. 33–55,
2016.
[4] A. Van der Horst, A time based analysis of road user behaviour in
normal and critical encounters, 1990, no. HS-041 255.
[5] J. C. Hayward, “Near-miss determination through use of a scale of
danger,” Pennsylvania Transportation and Traffic safety Center, 1971.
[6] R. van der Horst and J. Hogema, Time-to-collision and collision
avoidance systems, 1994.
[7] J. Jansson, Collision Avoidance Theory with Application to Automotive
Collision Mitigation, 2005, no. 950.
[8] P. Falcone, M. Ali, and J. Sj¨
oberg, “Predictive threat assessment via
reachability analysis and set invariance theory,” IEEE Transactions on
Intelligent Transportation Systems, vol. 12, no. 4, pp. 1352–1361, 2011.
[9] R. Kiefer, M. Cassar, C. Flannagan, D. LeBlanc, M. Palmer, R. Deer-
ing, and M. Shulman, “Refining the camp crash alert timing approach
by examining” last second” braking and lane change maneuvers under
various kinematic conditions,” Accident Investigation Quarterly, no. 39,
2004.
[10] R. Kiefer, D. LeBlanc, and C. Flannagan, “Developing an inverse
time-to-collision crash alert timing approach based on drivers last-
second braking and steering judgments,” Accident Analysis Prevention,
vol. 37, no. 2, pp. 295 – 303, 2005.
[11] S. Noh and W. Y. Han, “Collision avoidance in on-road environment for
autonomous driving,” International Conference on Control, Automation
and Systems, 2014.
[12] S. Sontges, M. Koschi, and M. Althoff, “Worst-case analysis of
the time-to-react using reachable sets,” in IEEE Intelligent Vehicles
Symposium, 2018.
[13] A. Tamke, T. Dang, and G. Breuel, “A flexible method for criticality
assessment in driver assistance systems,” IEEE Intelligent Vehicle
Symposium, no. Iv, pp. 697–702, 2011.
[14] M. Br¨
annstr¨
om, E. Coelingh, and J. Sj¨
oberg, “Threat assessment for
avoiding collisions with turning vehicles,” IEEE Intelligent Vehicles
Symposium, pp. 663–668, 2009.
[15] ——, “Model-based threat assessment for avoiding arbitrary vehicle
collisions,” IEEE Transactions on Intelligent Transportation Systems,
vol. 11, no. 3, pp. 658–669, 2010.
[16] A. Eidehall, “Multi-target threat assessment for automotive applica-
tions,” IEEE International Conference on Intelligent Transportation
Systems, 2011.
[17] D. Mad˚
as, M. Nosratinia, M. Keshavarz, P. Sundstrom, R. Philippsen,
A. Eidehall, and K.-M. Dahlen, “On path planning methods for
automotive collision avoidance,” IEEE Intelligent Vehicles Symposium,
2013.
[18] A. Eidehall and D. Mad˚
as, “Real time path planning for threat
assessment and collision avoidance by steering,” IEEE International
Conference on Intelligent Transportation Systems, 2013.
[19] J. Nilsson, A. ¨
Odblom, and J. Fredriksson, “Worst-case analysis of au-
tomotive collision avoidance systems,” IEEE Transactions on Vehicular
Technology, vol. 65, no. 4, pp. 1899–1911, April 2016.
[20] S. Hosseini, N. Murgovski, G. Rodrigues de Campos, and J. Sj¨
oberg,
“Adaptive forward collision warning algorithm for automotive applica-
tions,” American Control Conference, pp. 5982–5987, 2016.
[21] Y. Zhang, E. Antonsson, and K. Grote, “A new threat assessment mea-
sure for collision avoidance systems,” IEEE Intelligent Transportation
Systems Conference, pp. 968–975, 2006.
[22] J. B. Rawlings and D. Q. Mayne, Model Predictive Control : Theory
and Design, 2012.
[23] F. Borrelli, A. Bemporad, and M. Morari, Predictive control for linear
and hybrid systems, 2017.
[24] S. Anderson, S. C. Peters, T. E. Pilutti, and K. Iagnemma, “Design
and development of an optimal-control-based framework for trajectory
planning, threat assessment, and semi-autonomous control of passenger
vehicles in hazard avoidance scenarios,” Springer Tracts in Advanced
Robotics, vol. 70, pp. 39–54, 2011.
[25] S. Anderson, A unified framework for trajectory planning, threat
assessment, and semi-autonomous control of passenger vehicles, 2009.
[26] A. Gray, Y. Gao, T. Lin, J. K. Hedrick, and F. Borrelli, “Stochastic
predictive control for semi-autonomous vehicles with an uncertain
driver model,” IEEE Conference on Intelligent Transportation Systems,
pp. 2329–2334, 2013.
[27] B. Gutjahr and M. Werling, “Automatic collision avoidance during
parking and maneuvering — An optimal control approach,” IEEE
Intelligent Vehicles Symposium, pp. 636–641, 2014.
[28] R. Hult, G. Rodrigues de Campos, P. Falcone, and H. Wymeersch,
“An approximate solution to the optimal coordination problem for
autonomous vehicles at intersections,” American Control Conference,
2015.
[29] R. Hult, M. Zanon, S. Gros, and P. Falcone, “Primal decomposition
of the optimal coordination of vehicles at traffic intersections,” IEEE
Conference on Decision and Control, pp. 2567–2573, 2016.
[30] M. Zanon, S. Gros, P. Falcone, and H. Wymeersch, “Asynchronous
algorithm for optimal vehicle coordination at traffic intersections,”
World Congress of the International Federation of Automatic Control,
2017.
[31] P. Tabuada, Verification and Control of Hybrid Systems. Springer,
2008.
[32] C. Belta, Formal Methods for Discrete-Time Dynamical Systems.
Springer, 2017.
[33] C. Baier and J.-P. Katoen, Principles of model checking. Springer,
2008.
[34] A. Aswani, H. Gonzalez, S. Sastry, and C. Tomlin, “Provably safe and
robust learning-based model predictive control,” Automatica, vol. 49,
no. 5, pp. 1216–1226, 2013.
[35] M. Althoff, “Reachability analysis and its application to the safety
assessment of autonomous cars,” Ph.D. dissertation, Technische Uni-
versitat Munchen, 2010.
[36] A. Colombo and D. Del Vecchio, “Least restrictive supervisors for
intersection collision avoidance: A scheduling approach,” IEEE Trans-
actions on Automatic Control, vol. 60, no. 6, pp. 1515–1527, 2015.

12
[37] S. Loos, A. Platzer, and L. Nistor, “Adaptive cruise control: Hybrid,
distributed, and now formally verified,” Lecture Notes in Computer
Science, Springer, vol. 6664, pp. 42–56, 2011.
[38] M. Hilscher, S. Linker, R. Olderog, and A. Ravn, “An abstract model
for proving safety of multi-lane traffic manoeuvres,” Theories of Pro-
gramming and Formal Methods. Lecture Notes in Computer Science,
Springer, vol. 6991, pp. 404–419, 2011.
[39] M. Hilscher, S. Linker, and E. Olderog, “Proving safety of traffic
manoeuvres on country roads,” Theories of Programming and Formal
Methods. Lecture Notes in Computer Science, Springer, vol. 8051, pp.
196–212, 2013.
[40] W. Damm, H.-J. Peter, J. Rakow, and B. Westphal, “Can we build it:
formal synthesis of control strategies for cooperative driver assistance
systems,” Mathematical Structures in Computer Science, vol. 23, no. 8,
pp. 676–725, 2013.
[41] A. Rizaldi and M. Althoff, “Formalising traffic rules for accountability
of autonomous vehicles,” in IEEE Conference on Intelligent Trans-
portation Systems, 2015.
[42] H. Kowshik, D. Caveney, and P. R. Kumar, “Provable systemwide
safety in intelligent intersections,” IEEE Transactions on Vehicular
Technology, vol. 60, no. 3, pp. 804–818, March 2011.
[43] M. Althoff and J. M. Dolan, “Online verification of automated road
vehicles using reachability analysis,” IEEE Transactions on Robotics,
vol. 30, no. 4, pp. 903–918, 2014.
[44] M. Althoff and S. Magdici, “Set-Based Prediction of Traffic Partici-
pants on Arbitrary Road Networks,” IEEE Transactions on Intelligent
Vehicles, vol. 1, no. 2, pp. 187–202, 2016.
[45] M. Koschi and M. Althoff, “Spot: A tool for set-based prediction of
traffic participants,” in IEEE Intelligent Vehicles Symposium, 2017, pp.
1686–1693.
[46] C. Pek, M. Koschi, M. Werling, and M. Althoff, “Enhancing motion
safety by identifying safety-critical passageways,” pp. 320–326, 2017.
[47] M. Ali, P. Falcone, and J. Sj¨
oberg, “Model-based threat assessment for
lane guidance systems,” American Control Conference, pp. 4586–4591,
2011.
[48] G. Rodrigues de Campos, P. Falcone, R. Hult, H. Wymeersch, and
J. Sj¨
oberg, “Traffic coordination at road intersections: Autonomous
decision-making algorithms using model-based heuristics,” IEEE In-
telligent Transportation Systems Magazine, vol. 9, no. 1, pp. 8–21,
2017.
[49] G. Rodrigues de Campos, P. Falcone, and J. Sj¨
oberg, “Autonomous co-
operative driving: a velocity-based negotiation approach for intersection
crossing,” IEEE Conference on Intelligent Transportation Systems, pp.
1456–1461, 2013.
[50] G. Rodrigues de Campos, P. Falcone, H. Wymeersch, R. Hult, and
J. Sj¨
oberg, “Cooperative receding horizon conflict resolution at traffic
intersections,” IEEE Conference on Decision and Control, pp. 2932–
2937, 2014.
[51] C. Tomlin, G. Pappas, and S. Sastry, “Conflict resolution for air
traffic management: A study in multi-agent hybrid systems,” IEEE
Transactions on Automatic Control, vol. 43, pp. 509–521, 1998.
[52] F. Fadaie and M. E. Broucke, “On the least restrictive control for
collision avoidance of two unicycles,” International Journal of Robust
Nonlinear Control, vol. 16, pp. 553–574, 2006.
[53] R. Verma and D. Del Vecchio, “Semi-autonomous multi-vehicle safety:
A hybrid control approach,” IEEE Robotics & Automation Magazine,
vol. 18, pp. 44–54, 2011.
[54] A. Colombo and D. Del Vecchio, “Efficient algorithms for collision
avoidance at intersections,” ACM Conference on Hybrid Systems:
Computation and Control, 2012.
[55] M. Hafner, D. Cunningham, L. Caminiti, and D. Del Vecchio, “Cooper-
ative collision avoidance at intersections: Algorithms and experiments,”
IEEE Transactions on Intelligent Transportation Systems, 2013.
[56] L. Bruni, A. Colombo, and D. Del Vecchio, “Robust multi-agent
collision avoidance through scheduling,” IEEE Conference on Decision
and Control, 2013.
[57] H. Ahn, A. Colombo, and D. Del Vecchio, “Supervisory control
for intersection collision avoidance in the presence of uncontrolled
vehicles,” American Control Conference, 2014.
[58] H. Ahn, A. Rizzi, A. Colombo, and D. Del Vecchio, “Experimental
testing of a semi-autonomous multi-vehicle collision avoidance algo-
rithm at an intersection testbed,” IEEE/RSJ International Conference
on Intelligent Robots and Systems, 2015.
[59] J. Lygeros, C. Tomlin, and S. Sastry, “Controllers for reachability
specifications for hybrid systems,” Automatica, vol. 35, pp. 349–370,
1999.
[60] E. C. Kerrigan, “Robust Constraint Satisfaction : Invariant Sets and
Predictive Control,” Ph.D. dissertation, University of Cambridge, 2000.
[61] M. Pinedo, Scheduling: Theory, Algorithms, and Systems. Springer,
2008.
[62] G. Rodrigues de Campos, F. Della Rossa, and A. Colombo, “Optimal
and least restrictive supervisory control: Safety verification methods
for human-driven vehicles at traffic intersections,” IEEE Conference
on Decision and Control, 2015.
[63] A. Colombo, G. Rodrigues De Campos, and F. Della Rossa, “Control
of a city road network: Distributed exact verification of traffic safety,”
IEEE Transactions on Automatic Control, vol. 62, no. 10, pp. 4933 –
4948, 2017.
[64] M. R. Hafner and D. Del Vecchio, “Computational tools for the safety
control of a class of piecewise continuous systems with imperfect infor-
mation on a partial order,” SIAM Journal on Control and Optimization,
vol. 49, no. 6, pp. 2463–2493, 2011.
[65] M. Hafner, D. Cunningham, L. Caminiti, and D. Del Vecchio, “Cooper-
ative collision avoidance at intersections: Algorithms and experiments,”
IEEE Transactions on Intelligent Transportation Systems, 2013.
[66] E. Dallal, A. Colombo, D. Del Vecchio, and S. Lafortune, “Supervisory
control for collision avoidance in vehicular networks using discrete
event abstractions,” Discrete Event Dynamic Systems, vol. 27, pp. 1–
44, 2017.
[67] B. Sch¨
urmann, D. Hess, J. Eilbrecht, O. Stursberg, F. K¨
oster, and
M. Althoff, “Ensuring drivability of planned motions using formal
methods,” in IEEE International Conference on Intelligent Transporta-
tion Systems, 2017.
[68] D. Mayne, S. Rakovic, R. Findeisen, and F. Allg¨
ower, “Robust output
feedback model predictive control of constrained linear systems,”
Automatica, vol. 42, no. 7, pp. 1217–1222, 2006.
[69] A. Broadhurst, S. Baker, and T. Kanade, “Monte Carlo road safety
reasoning,” IEEE Intelligent Vehicles Symposium, pp. 319–324, 2005.
[70] A. Eidehall and L. Petersson, “Threat assessment for general road
scenes using monte carlo sampling,” Intelligent Transportation Systems,
pp. 1173–1178, 2006.
[71] S. Danielsson, L. Petersson, and A. Eidehall, “Monte Carlo based
Threat Assessment: Analysis and Improvements,” 2007 IEEE Intel-
ligent Vehicles Symposium, pp. 233–238, 2007.
[72] A. Eidehall and L. Petersson, “Statistical Threat Assessment for Gen-
eral Road Scenes Using Monte Carlo Sampling,” IEEE Transactions
on Intelligent Transportation Systems, vol. 9, no. 1, pp. 137–147, 2008.
[73] J. Kim, K. Jo, W. Lim, M. Lee, and M. Sunwoo, “Curvilinear-
coordinate-based object and situation assessment for highly automated
vehicles,” IEEE Transactions on Intelligent Transportation Systems,
vol. 16, no. 3, pp. 1559–1575, 2015.
[74] M. Althoff, O. Stursberg, and M. Buss, “Model-Based Probabilistic
Collision Detection in Autonomous Driving,” IEEE Transactions on
Intelligent Transportation Systems, vol. 10, no. 2, pp. 299–310, 2009.
[75] M. Althoff and A. Mergel, “Comparison of Markov chain abstraction
and Monte Carlo simulation for the safety assessment of autonomous
cars,” IEEE Transactions on Intelligent Transportation Systems, vol. 12,
no. 4, pp. 1237–1247, 2011.
[76] G. Rodrigues de Campos, A. H. Runarsson, F. Granum, P. Falcone,
and K. Alenljung, “Collision avoidance at intersections: A probabilistic
threat-assessment and decision-making system for safety interventions,”
IEEE International Conference on Intelligent Transportation Systems,
2014.
[77] S. Lef`
evre, “Risk Estimation at Road Intersections for Connected Ve-
hicle Safety Applications,” Ph.D. dissertation, Universite de Grenoble,
2012.
[78] A. Armand, D. Filliat, and J. Iba˜
nez-Guzman, “A Bayesian Framework
for Preventive Assistance at Road Intersections,” IEEE Intelligent
Vehicles Symposium, pp. 1128–1134, 2016.
[79] T. Gindele, S. Brechtel, and R. Dillmann, “Learning driver behavior
models from traffic observations for decision making and planning,”
IEEE Intell. Transp. Syst. Mag., vol. 7, no. 1, pp. 69–79, 2015.
[80] C. Katrakazas, “Developing an advanced collision risk model for
autonomous vehicles,” Ph.D. dissertation, Loughborough University,
2017.
[81] C. Chen, C. Tay, C. Laugier, and K. Mekhnacha, “Dynamic environ-
ment modeling with gridmap: A multiple-object tracking application,”
9th Int. Conf. Control. Autom. Robot. Vision, 2006, ICARCV ’06, 2006.
[82] T. Gindele, S. Brechtel, J. Schr¨
oder, and R. Dillmann, “Bayesian
occupancy grid filter for dynamic environments using prior map
knowledge,” IEEE Intell. Veh. Symp. Proc., pp. 669–676, 2009.

13
[83] D. Greene, J. Liu, J. Reich, Y. Hirokawa, A. Shinagawa, H. Ito, and
T. Mikami, “An efficient computational architecture for a collision
early-warning system for vehicles, pedestrians, and bicyclists,” IEEE
Transactions on Intelligent Transportation Systems, vol. 12, no. 4, pp.
942–953, 2011.
[84] R. Karlsson, J. Jansson, and F. Gustafsson, “Model-based statistical
tracking and decision making for collision avoidance application,”
American Control Conference, pp. 3435–3440, 2004.
[85] J. Jansson and F. Gustafsson, “A framework and automotive application
of collision avoidance decision making,” Automatica, vol. 44, no. 9, pp.
2347–2351, 2008.
[86] F. Sandblom and M. Br¨
annstr¨
om, “Probabilistic threat assessment
and driver modeling in collision avoidance systems,” IEEE Intelligent
Vehicles Symposium, pp. 914–919, 2011.
[87] M. Br¨
annstr¨
om, F. Sandblom, and L. Hammarstrand, “A probabilistic
framework for decision-making in collision avoidance systems,” IEEE
Transactions on Intelligent Transportation Systems, vol. 14, no. 2, pp.
637–648, 2013.
[88] C. Hubmann, J. Schulz, M. Becker, D. Althoff, and C. Stiller, “Au-
tomated driving in uncertain environments: Planning with interaction
and uncertain maneuver prediction,” IEEE Transactions on Intelligent
Vehicles, vol. 3, no. 1, pp. 5–17, 2018.
[89] A. Hussein, M. M. Gaber, E. Elyan, and C. Jayne, “Imitation Learning:
A Survey of Learning Methods,” ACM Computing Surveys, vol. 50,
no. 2, pp. 1–35, 2017.
[90] A. Armand, D. Filliat, and J. Ibanez-Guzman, “Modelling stop inter-
section approaches using Gaussian processes,” IEEE Conference on
Intelligent Transportation Systems, pp. 1650–1655, 2013.
[91] R. Huang, H. Liang, P. Zhao, B. Yu, and X. Geng, “Intent-
Estimation- and Motion-Model-Based Collision Avoidance Method
for Autonomous Vehicles in Urban Environments,” Applied Sciences,
vol. 7, no. 5, p. 457, 2017.
[92] D. Urun, J. Edelbrunner, and I. Iossifidis, “Autonomous driving: A
comparison of machine learning techniques by means of the prediction
of lane change behavior,” IEEE International Conference on Robotics
and Biomimetics, pp. 1837–1843, 2011.
[93] N. Motamedidehkordi, S. Amini, S. Hoffmann, F. Busch, and M. R.
Fitriyanti, “Modeling tactical lane-change behavior for automated ve-
hicles: A supervised machine learning approach,” IEEE International
Conference on Models and Technologies for Intelligent Transportation
Systems, pp. 268–273, 2017.
[94] G. Notomista and M. Botsch, “A machine learning approach for the
segmentation of driving maneuvers and its application in autonomous
parking,” Journal of Artificial Intelligence and Soft Computing Re-
search, vol. 7, no. 4, pp. 243–255, 2017.
[95] R. Gruner, P. Henzler, G. Hinz, C. Eckstein, and A. Knoll, “Spatio-
temporal Representation of Driving Scenarios and Classification using
Neural Networks,” IEEE Intelligent Vehicles Symposium, pp. 1782–
1788, 2017.
[96] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp,
P. Goyal, L. D. Jackel, M. Monfort, U. Muller, J. Zhang, X. Zhang,
J. Zhao, and K. Zieba, “End to End Learning for Self-Driving Cars,”
arXiv:1604, pp. 1–9, 2016.
[97] M. Bojarski, P. Yeres, A. Choromanska, K. Choromanski, B. Firner,
L. Jackel, and U. Muller, “Explaining How a Deep Neural Network
Trained with End-to-End Learning Steers a Car,” arXiv, pp. 1–8, 2017.
[98] S.-G. Jeong, J. Kim, S. Kim, and J. Min, “End-to-end learning of image
based lane-change decision,” IEEE Intelligent Vehicles Symposium, pp.
1602–1607, 2017.
[99] J. M. Ambarak, H. Ying, F. Syed, and D. Filev, “A neural network for
predicting unintentional lane departures,” IEEE International Confer-
ence on Industrial Technology, pp. 492–497, 2017.
[100] H. G. Seif and X. Hu, “Autonomous driving in the icity - hd maps as a
key challenge of the automotive industry,” Engineering, vol. 2, no. 2,
pp. 159 – 162, 2016.
[101] S. A. Bagloee, M. Tavana, M. Asadi, and T. Oliver, “Autonomous
vehicles: challenges, opportunities, and future implications for trans-
portation policies,” Journal of Modern Transportation, vol. 24, no. 4,
2016.
[102] P. Koopman and M. Wagner, “Autonomous vehicle safety: An in-
terdisciplinary challenge,” IEEE Intelligent Transportation Systems
Magazine, vol. 9, no. 1, pp. 90–96, 2017.
[103] ——, “Challenges in autonomous vehicle testing and validation,” SAE
International Journal of Transportation Safety, vol. 4, no. 1, pp. 15–24,
2015.
[104] G. Grimm, M. J. Messina, S. E. Tuna, and A. R. Teel, “Examples when
nonlinear model predictive control is nonrobust,” Automatica, vol. 40,
no. 10, pp. 1729 – 1738, 2004.
[105] P. J. Campo and M. Morari, “Robust model predictive control,” IEEE
American Control Conference, pp. 1021–1026, 1987.
[106] A. Bemporad and M. Morari, Robust Model Predictive Control : A
Survey. Springer, 2007.
[107] B. Kouvaritakis and M. Cannon, Model predictive control: Classical,
robust and stochastic. Springer, 2015.
[108] M. C. Steinbach, “Tree-sparse convex programs,” Mathematical meth-
ods of operations research, vol. 56, no. 3, pp. 347–376, 2003.
[109] E. Klintberg, J. Dahl, J. Fredriksson, and S. Gros, “An improved dual
newton strategy for scenario-tree mpc,” IEEE Conference on Decision
and Control, pp. 3675–3681, 2016.
John Dahl received the M.Sc. degree in Systems,
Control and Mechatronics in 2013 from Chalmers
University of Technology, Gothenburg, Sweden,
where he is currently working toward the Ph.D.
degree at the Department of Electrical Engineering.
He is currently doing research with the Depart-
ment of Electrical Engineering and Zenuity AB,
Gothenburg. His research interests include active
safety and estimation techniques, in particular, for
threat-assessment and decision-making in collision-
avoidance systems.
Gabriel Rodrigues de Campos received the Ph.D.
degree in Automatic Control in 2012 from Grenoble
University/Grenoble INP, France. He is currently
a researcher with Zenuity in Gothenburg, Sweden.
Prior to joining the Zenuity, he was a postdoctoral
fellow with the Department of Signals and Sys-
tems, Chalmers University of Technology, Sweden
and the Politecnico di Milano, Italy. His research
interests include cooperative and distributed control,
multi-agent systems, and intelligent transportation
systems.
Claes Olsson received his PhD in 2005 in Automatic
Control from Uppsala University, Sweden. The ini-
tial research focus was within the field of active
vibration isolation applied to automotive design.
Claes joined Volvo Car Corporation in 2005 where
he carried out development and research for active
safety systems and received the role as a Technical
Expert on Collision Avoidance. He is currently with
Zenuity AB acting as a technical expert within the
field of ADAS and, in particular, collision avoidance
technology.
Jonas Fredriksson received his M.Sc. in Com-
puter Science Engineering from Lule˚
a University of
Technology, Sweden in 1997 and Ph.D. in Auto-
matic Control from Chalmers University of Technol-
ogy, Sweden in 2002. Currently, he is Professor in
Mechatronics at the Department of Electrical Engi-
neering at Chalmers. His research activities include
modeling, control and simulation, with a special
interest in automotive applications.