Content uploaded by Dimia Iberraken
Author content
All content in this area was uploaded by Dimia Iberraken on Jan 10, 2020
Content may be subject to copyright.
Multi-level Bayesian decision-making for safe and flexible autonomous
navigation in highway environment
Dimia Iberraken1,2, Lounis Adouane1and Dieumet Denis2
Abstract— This paper proposes an overall Multi-Controller
Architecture (MCA) for safe and flexible navigation of au-
tonomous navigation, under uncertainties in highway use-cases.
In addition to the details given about the main modules (and
their interactions) composing the proposed MCA, an important
focus of the paper is made on the definition of a robust
Two-Sequential Level Decision Network (TSLDN), which uses
both: Extended Time-To-Collision (ETTC) metric and a new
definition of a specific Predicted Inter-Distance Profile (PIDP,
between vehicles during lane changes maneuvers) in order to
estimate the maneuvers risks. The TSLDN is utilized for: the
driving situation assessment, decision-making and for safety
retrospection over the current maneuver risk. It allows us to
have the best decision to achieve the vehicle navigation task
while maximizing its safety. Several simulation results show the
good performance of the overall proposed control architecture,
mainly in terms of efficiency to handle probabilistic decision-
making even for very risky scenarios.
I. INTRODUCTION
An important challenge in the field of autonomous ve-
hicles is guaranteeing a safe and smooth navigation while
ensuring the safety of passengers. Although Advanced Driver
Assistance Systems (ADAS) such as Adaptive Cruise Control
(ACC), Lane Keeping Assistance (LKA), have successfully
improved safety, fatal car crashes still occur. One of the
main challenging maneuvers for autonomous vehicles in
highway correspond to the lane change (for overtaking or
lane change to the right) due mainly to the bad estimation
of the behavioral risk of the surrounding environment [1].
For this reason the risk management field has gained a lot
of attention in the recent years. Indicators such as the Time-
To-Collision (TTC) [2], have been extensively used for its
simplicity and its low cost computational time. However, the
common definition of the TTC is restricted for a specific
path to detect longitudinal collision and for well-defined
scenarios such as car following. In order to improve this
measure, [3]–[5] address the problem from planar perspective
where vehicles are considered in a two-dimensional plane
and the state of each vehicle is defined by a vector position,
velocity and acceleration components on X and Y direction.
This Extended TTC (ETTC) is computed at each time
step for each vehicle pair that are close enough. In this
paper, the ETTC formulation proposed in [3] is combined
with the proposed Predicted Inter-Distance Profile (PIDP cf.
Section III-A.6) during the overtaking maneuver to efficiently
The authors are with 1Universit´
e Clermont Auvergne, CNRS,
SIGMA Clermont, Institut Pascal, F-63000 Clermont-Ferrand, France.
FirstName.Lastname@uca.fr
2Sherpa Engineering Company, R&D Department La Garenne Colombe,
France. [d.iberraken, d.denis] @sherpa-eng.com
measure collision risk around the vehicle at every instant that
compensates for possible failure of the perception module or
other devices.
A common requirement for ADAS systems, is the need
of a decision strategy. In the literature, numerous methods
have been used as decision methods. In [6], authors describe
a fully automated driving algorithm that uses a dynamic
drivable area as a safety constraint for the optimal trajectory
in which the vehicle stays to ensure its safety. An energy
function based on potential field is used to assess the risk
and drive decision maneuvers. However, this method does
not take uncertainties into consideration. In [7], Schubert
uses a Bayesian network for lane change decision-making
and a deceleration to safety time (DST) as a threat measure
to assess the danger of the navigation lanes status. However,
the common definition of the DST is restricted for a specific
path to detect longitudinal collision.
In general planned trajectories assumes that obstacles
doesn’t deviate from their predefined behavior and in order to
avoid collisions, safety verification methods as reachability
analysis [8] have been used in the literature to verify the
safety of these trajectories. Reachability analysis calculates
the reachable set of positions of each vehicle in the envi-
ronment and possible future collisions are identified when
comparing the intersection of the obtained sets. However if
the trajectory is regarded as unsafe no alternative is proposed.
In this work, a Multi-level Bayesian decision-making
approach [9] through a Two-Sequential Level Decision Net-
work (TSLDN) is proposed. This approach constitutes a good
deal for handling numerous scenarios configuration, multi-
ple decision criteria while taking uncertainty into account.
The graphical representation of the Decision Network (DN)
eases the connection between the situation assessment level
(observations level) using the collision risk measure and the
decision-making level. The overall network consists of a
situation assessment part which defines the current driving
state of safety, a decision-making strategy that makes the
control decision and a safety verification of the maneuver that
allows to propose an evasive alternative. The Probabilistic
model summarizing the TSLDN is described in Section III
and constitutes a novel manner to manage decision-making
maneuvers. In addition, it is proposed in this paper, an
adaptation to the highway case of the limit-cycle trajectories,
employed in [10] and [11] for mobile robot navigation in
cluttered environment, that will constitutes the baseline of the
used obstacle avoidance strategy. The structure of this paper
is as follows: Section II-A describes the proposed overall
MCA. Section III formalizes the decision-making problem
00
Control law
Vehicle
(𝑣, 𝛾)
Path
planning
based
local and
global
perception
Path
Homogeneous set-points
Multi-level
Bayesian
Decision
Making
Environment
Position and velocity
of obstacles
Closest hindering
obstacles
Information on the
lanes.
Optimal Decision
Threat
Measure:
Extended TTC
Probabilistic
Decision Making
Inter-vehicular distance
Localization of vehicles
Perception and Localization
Raw Sensor
Data
Sensor Uncertainty
1
2
5
(𝑥, 𝑦, 𝜃, 𝑣)𝑂𝑏𝑠𝑡𝑎𝑐𝑙𝑒
(𝑥, 𝑦, 𝜃, 𝑣)𝐸𝑔𝑜
Predicted
Inter-
Distance
Profile
Task
Achievement
validation
Risk
management:
Safety
verification
(𝑥, 𝑦, 𝜃, 𝑣)𝐸𝑔𝑜
Task activation, achievement & validation
3
ADAS
Lane Keeping
Assist (LKA)
Auto Lane
Change (ALC)
Adaptive
Cruise Control
(ACC)
4
Fig. 1. Proposed multi-controller architecture for highway navigation
where are detailed the used TSLDN structure, the risk
criteria and the decision-making strategy for deriving most
suitable maneuver decision. The simulation results will be
presented in Section IV based on experiments performed on
a Matlab/Simulink car simulator that has been implemented
to test the developed algorithms. This paper concludes with
perspectives on future research.
II. MULTI -CONTROLLER ARCHITECTURE (MCA)
It is proposed in this paper a MCA, shown in Fig. 1,
that aims at decomposing the overall complex task into a
multitude of sub-tasks to achieve [12]. Once the path plan-
ning is developed, an appropriate decision-making strategy
for autonomous navigation has to be defined that takes into
account several aspects, such as: traffic rules, passenger
safety and measurement uncertainty of perceptive modules.
In this MCA, a probabilistic decision-making block (detailed
in Section III) computes the most suitable decision according
to the environment knowledge based on perception sensors
while taking into account the presence of uncertainty to
achieve desired action. Then a selection process, based on the
Task activation, achievement and validation block, enables
the switch between different ADAS modules (block 3 and 4
in Fig. 1) to activate the corresponding ADAS that generates
dynamic target set-points (cf. Section II-A). These set-points
are fed to the nonlinear control law (block 5 represented in
Fig. 1) developed in [13] that aims to drive the vehicle toward
specific (static or dynamic) target set-points. This control
law is based on a Lyapunov function design to ensure the
convergence of the vehicle to the target. The motion of the
host vehicle is described by the so-called tricycle model [13].
A. Elementary Advanced Driver Assistance Systems (ADAS)
During autonomous navigation in highway, vehicles per-
form either an ACC behavior for driving with desired veloc-
ity while maintaining a safety distance with vehicles ahead,
or LKA or switches to an Auto-Lane Change (ALC) behavior
while guaranteeing the smoothness and the safety of the
trajectory. For these behaviors, a homogeneous target set-
points definition has been proposed in [12] [10], defined by
a pose (xT,yT,θT) and a velocity vTwhich can be constant
or variable indifferently Fig. 2.
1) Lane Keeping Assist (LKA): For the lane keeping
assist, where a global path is already defined to be the
center-line of the lane to follow, it is enough for the vehicle
to follow this path as precisely as possible without any
modification. The dynamic target set-points are extracted
then using a Frenet reference frame [10] (cf. Fig. 2(a)).
They correspond to the closest position (xT,yT)in the path
(a) Frenet reference frame (b) Eucledian distance RS
Fig. 2. Set-points definition based on (a) global planned path, (b) local
planned path.
with an offset curvilinear distance, w.r.t to the origin of the
vehicle reference frame, and to an orientation θTtangent to
the defined path at (xT,yT).
2) Adaptive Cruise Control (ACC): The adaptive cruise
control follows the same homogeneous reasoning (in terms
of used set-points and control law). Dynamic target set-
points are extracted using a Frenet reference frame (cf. Fig.
2(a)). A desired velocity, that insures maintaining a temporal
safety distance with vehicles ahead, is generated using the
predefined control law.
3) Auto-Lane Change (ALC): The auto-lane change con-
troller in the other hand, is based on generated limit-cycles
trajectories which are defined in the literature according to
an elliptic periodic orbit [11] corresponding to an ellipse
of influence. These periodic orbits if well-dimensioned and
accurately followed guarantee the avoidance of any given
obstacle. These ELC trajectories are defined according to a
set of differential equations [11]:
Fig. 3. Lane Change maneuver: Overtaking yellow vehicle with auto lane
change and lane keeping assist
˙xs=mys+µxs(1−xs/a2
lc −y2
s/b2
lc +cxsys)
˙ys=−mxs+µys(1−xs/a2
lc −y2
s/b2
lc +cxsys)(1)
with m=±1 according to the direction of avoidance
(clockwise or counterclockwise). (xs,ys)corresponds to the
position of the vehicle according to the center of the ellipse,
alc and blc characterize the major and minor elliptic semi-
axes respectively (defined above), cgives the orientation
of the ellipse and µa positive constant that enable us to
modulate the convergence of the ELC trajectory toward
the ellipse of influence. In this work, an adaptation to
the highway case of elliptic limit-cycles (ELC) techniques,
previously proposed in [11] for mobile robot navigation
in cluttered environment, has been carried out, to perform
the lane change maneuver, while taking into account ve-
hicle speeds and traffic rules in the dimensioning of the
evolution trajectories of the controlled autonomous vehicle
(cf. Fig .3). To differentiate the multiple parts of the lane
change maneuver, a convention is used based on Worrall and
Bullen definition [14] in: head portion, lane change part,
tail portion (cf. Fig. 3). The ellipse of influence has been
expanded laterally and longitudinally based on the traffic
rule regulation and the vehicle dimensions. Indeed, based
on the French highway traffic regulation, a minimum lateral
distance of Ldistance =1.5mhas to be left during the tail
portion and a longitudinal safety distance equivalent to the
distance traveled over ts=2seconds. The parameters of the
ellipse are then given by the following equations:
alc =0.5lb+tsvr
blc =wb+Ldist ance (2)
with lbthe wheelbase of the vehicle, wbthe vehicle track
and vris the relative velocity.
As for the navigation, because the vehicle is already on
the defined path, the ELC takes as initial parameters the
current vehicle configuration. The extraction of set-points in
this case, is based on a heuristic defined in [10] where at
each sample time the intersection between a circle surround-
ing the ego vehicle (defined by a radius Rsand a center
corresponding to the origin of the reference frame linked to
the ego vehicle) and the pre-planned ELC is calculated (cf.
Fig. 2(b)). The intersection point (xT,yT)corresponds then
to the position of the dynamic set-point, the orientation θTis
the tangent to the ELC at (xT,yT)and the velocity vTwhich
is constant or variable indifferently.
In the next section, the probabilistic decision block con-
stituting the main contribution of this paper is detailed.
III. PROP OSE D MULTI-LEVE L DECISION NETW OR K F OR
LA NE -CHANGE ASSISTANCE
In order to clarify the used nomenclature in this paper, it
is important to recall the basic definitions of Bayesian net-
works. A Bayesian network (BN) is a directed acyclic graph
representing a set of random variables and their conditional
dependencies. Decision networks (DN) are an extension
of BNs that allow us to support probabilistic reasoning,
decision-making under uncertainty for a given system and
yield the capacity to incorporate multiple decision criteria.
A DN has three types of nodes: chance nodes, utility nodes
and decision nodes. The set of chance nodes UCrepresents
the set of random variables and their conditional probabilistic
dependencies. They are summarized in a conditional proba-
bility table (CPT) representing the conditional probabilities
P(Xi|pa(Xi)with Xi∈UC. The utility nodes UVdefines
the cost related to the decision. In the literature [15], a
normalized utility scale interval [0,1]is usually used, to
compare between some complex scenario. Both chance and
decision nodes can be parents of a utility node, and their state
directly affects the utility value. Finally, the set of decision
nodes UDrepresents the choice of action among alternatives.
In a Multi-Level Decision Network (MLDN) [9], the decision
nodes have a temporal order (D1,·· · ,Dn) which means the
action chosen for decision Dn−1is part of the information
available at decision Dnalong with past observations.
It is proposed in this paper a Two-Sequential Level Deci-
sion Network (TSLDN) (cf. Fig. 4). The Maneuver Decision
Level (MDL) in this network represents the choice of action
regarding the most suitable lane change maneuver in terms
of safety, based on the current situation assessment using
the ETTC (cf. Section III-A.3) while taking measurements
uncertainties into account. The Safety Verification Decision
Level (SVDL) consists of a safety checking regarding the
action chosen in the MDL, based on the definition of a
new measure, the Predicted Inter-Distance Profile (PIDP) (cf.
Section III-A.6). The PIDP is utilized in order to estimate
the current performed maneuver risks to compensate for pos-
sible failure of the perception module and/or other vehicles
intention/action lack of precision, and therefore propose the
best decision to achieve the vehicle navigation task while
maximizing its safety given the available evidence. The
purpose of a multi-level DN instead of one is the ability
to reason while accounting for a margin uncertainty (cf.
Fig. 5). As the MDL is very dynamic and the choice of
action is instantaneously taken which means if a false alarm
is triggered due to wrong information from the perceptive
module for instance, the MDL will immediately abort the
previous decision and compute another one. The SVDL, on
the other hand allows us in this case to verify the coherence
of the maneuver and thus allows a safety retrospection over
the current performed maneuver (cf. Section IV).
A. Probabilistic Situation Assessment
In the TSLDN, the lanes are numbered from right to left
by i, with i=1 denoting the rightmost lane. In this paper,
for the sake of convenience, a two lane configuration to
Maneuver Decision Level (MDL) Safety Verification Decision
Level (SVDL)
Actual Inter-Distance Profile
Extended TTC Extended TTC
Position Estimation
with uncertainty
L2: Level of dangerousness Lane 2
0 to 1
1 to 2
2 to 3
3 to 4
4 to 5
19.6
17.7
18.3
21.7
22.8
2.6 ± 1.5
OPosVeh: Observation on vehicle position
Lane1
Lane2 50.0
50.0
PosVeh: Vehicle Position Estimation
Lane1
Lane2 50.0
50.0
SL2: Status of Lane 2
Dangerous
Occupied
Free
33.3
33.3
33.3
Utility_Safety
AIDP: Actual Inter-distance Profile
Outside margin
Inside margin
No Change
33.2
34.8
32.0
SM: Status of Maneuver
Dangerous
Cautious
Safe
33.3
33.3
33.3
Utility_Check
SL1: Status Of Lane 1
Dangerous
Occupied
Free
33.3
33.3
33.3
L1: Level of dangerousness Lane 1
0 to 1
1 to 2
2 to 3
3 to 4
4 to 5
19.2
18.1
18.2
21.4
23.0
2.61 ± 1.5
Decision 1: Lane Change Maneuver
LaneChangeLeft
KeepLaneACC
LaneChangeRight
MaintainVelocity
0.76666
1.04444
0.71111
0.87777
Decision 2: Maneuver Safety Checking
Abort Maneuver
Warning Be careful
Maneuver is Safe
Fig. 4. Multi-level decision network for lane change maneuvers
present this model is considered. However, this architecture
is generic and can be extended to an N-Lane configuration.
To derive decision strategy for the most suitable maneuver
to be achieved, situation variable represented by UChave to
be defined. The chance nodes defining the structure of the
MDL are then:
1) Observation on the vehicle’s position OPosVeh:
OPosVeh is the uncertain observation denoting the estimated
position of the vehicle in the lane. This observation is
considered as the lateral control variable. In a DN when
a rough evidence on a random variable (PosVeh) is found,
the concept of virtual evidence is used [9] to propagate
it by adding a virtual node (OPosVeh), as a child of the
random variable (PosVeh). This virtual node will have a
CPT that reflects the uncertainty of the observation. The
candidate lane is selected by checking the closest distance
of the vehicle to the center-line of one of the lane based
on the definition of the Frenet reference frame [10] (cf.
Fig. 2(a)). This estimation is augmented with uncertainty,
as the situation assessment needs to account for erroneous
estimation. This is modeled using an ellipse of uncertainty
around the vehicle, propagated laterally and longitudinally in
terms of: the distance traveled from position Posk−1to Posk
and the vehicle’s dimension. The overflow of this ellipse near
the adjacent lane constitutes the uncertainty accounted in the
(OPosVeh) node’s CPT.
2) Vehicle position estimation PosVeh:This parameter
denotes the vehicle position in the lane. The possible states
are Lane1 and Lane2.
3) Observation on the level of danger of the lanes Li: Li
is an observation node that describes the level of danger of
the lanes. It is based on an extended formulation of the TTC
(ETTC), that addresses the problem from planar perspective
where vehicles are considered in a two-dimensional space.
A quartic equation with one unknown the ETTC (given in
[3]), take as parameters the state of each vehicle defined
by their position, velocity and acceleration component on
X and Y directions. The smallest root value of this quartic
equation is the ETTC value. The ETTC is computed at each
time step for each vehicle pair that are close enough. The
most dangerous vehicle in each of the lane characterized by
a small ETTC is used as input to the MDL. This observation
is considered as the longitudinal control variable. In the
MDL, the conditional probability distribution related to the
ETTC measure under the condition of the status of the
lanes P(Li|SLi)is approximated by a normal distribution.
This is justified by the fact that the ETTC is an estimation
based on uncertain sensor observations and only a probability
distribution is known with certainty. This is known as soft
evidence [9]. Thus, the likelihood function of the ETTC will
be:
P(Li|SLi) = N(µET T C ,σ2
ET T C ) = 1
σET T C √2πexp−1
2
(ET TCLi−µE T T C )2
σ2
ET TC (3)
with i∈(1,2),µET T C is the mean and σET T C is the standard
deviation.
Based on a five-level discretization of the likelihood function
of the ETTC and given uncertain evidence, equation (4) is
obtained:
P(Li|SLi=Dangerous) = N(µET TC =E T T Cdan,σ2
ET T C )
P(Li|SLi=Occupied ) = N(µE T TC =E T TCocc ,σ2
ET T C )
P(Li|SLi=Free) = N(µET T C =E T TC f ree,σ2
ET T C )
(4)
The value ET T Cstate with state = [Danger-
ous,Occupied,Free] represents the fixed threshold for
determining the occupancy of the lane.
4) Status of Lane SLi:These nodes describe the status
of occupancy of the lane. The possible states are Dangerous
(vehicles present on the lane at a critical distance to the
ego vehicle), Occupied (denoting the uncertainty and the risk
outside of the critical zone) and Free (no vehicles present on
the lane until a certain distance).
5) Utility SafetyUS:Utility nodes UVdefines the cost
related to the decision. In the MDL (cf. Fig. 4), Usa f ety
is the utility related to the safety of each of the maneuver
alternatives given the observations.
Fig. 5. Predicted inter-distance profile with velocity uncertainty
The SVDL in the other side is represented by the following
nodes:
6) Observation on Actual Inter-distance Profile AIDP:
We propose in this paper, a criteria-based on a predicted
inter-distance profile (PIDP) during overtaking built off-
line during normal conditions, to manage the maneuver
risk during overtaking (from the head portion to the tail
portion) and we compare it on-line to the actual inter-distance
profile (output of a laser range-finder sensor for example).
Indeed, the assumption considered is that if nothing changes
in the initial configuration, the predicted evolution of the
inter-distance between vehicles during lane change is not
supposed to change. The PIDP is proposed while following
the definition on a complete lane change maneuver (cf.
Fig. 3) and is shown in Fig. 5. The PIDP is constructed as the
euclidean distance between the ego vehicle and the candidate
obstacle-vehicle for the overall lane change maneuver.
An Upper Safety Boundary (USB) and a Lower Safety
Boundary (LSB) is allowed, that is fixed while taking into
account the vehicles’ perceptive uncertainties. Multiple ve-
locity configurations of the ego vehicle and the obstacle-
vehicle during the overtaking are considered. A ±percentage
uncertainty over the velocities is selected as uncertainty
value. We have chosen to propagate the uncertainty in this
way instead of setting a fixed value because uncertainty is
increasingly rising as the velocity increases. The LSB is
considered in this algorithm as a threshold of danger at every
time step. Indeed, the risk of collision increases when the
progress of the distance between vehicles is not conform
to the expected one. The following errors are calculated to
assess the dangerousness of the situation (cf. Fig. 5):
•Err1is the lateral error between the PIDP and the AIDP.
•Err2is the lateral error between the AIDP and the LSB.
These errors are fed into the SVDL. The node AIDP (cf.
Fig. 4) has then three states:
•Outside Margin means that the input AIDP is strongly
different then the PIDP (Err16=0) and its values
goes beyond the limit boundaries defined by the LSB
(Err2<0).
•Inside Margin means that the input AIDP is different
than the PIDP (Err16=0), however its values are within
the limit boundaries (Err2>0).
•No Change means that the input AIDP is equivalent to
the PIDP (Err1=0)
The same concept of virtual evidence is used in here through
the AIDP node, to propagate the uncertainty of the observa-
tions. The velocity uncertainty in this case is reflected in the
CPT of this node.
7) Status of maneuver (SM):This node describes the sta-
tus of the engaged maneuver based on the observations that
the node AIDP provides. The possible states are Dangerous
(for the case where the brought evidence on the AIDP is
outside the margin), Cautious (denoting the uncertainty and
the risk when the AIDP falls inside the margin), Safe (the
observation AIDP does not endanger the situation).
8) Utility Check UCh: UCheck is the cost related to the
safety verification during the lane change maneuver based
on the PIDP.
B. decision-making strategy for Lane-change maneuver
In this network, two decision nodes are represented. For
Decision1four possible maneuvers are defined: Lane Change
Left (LCL) and Lane Chane Right (LCR) for lane change
maneuvers , Keep Lane ACC (KLACC) for staying in the
considered lane while keeping a temporal safety distance
ET T Cocc with the obstacle-vehicle in front and Maintain
Velocity (MV) which is an alternative decision allowing to
stay in the current lane while maintaining previous velocity
configuration. This state allow us to ensure passenger safety,
smooth navigation and energy saving. Decision2in the other
side, has 3 states: Abort Maneuver (AM) that allow us to
react to a dangerous change in the PIDP by canceling the
previous decision effect on the system, Warning Be Careful
(WBC) state represents an additional safety level, where a
warning is issued if any change in the PIDP is detected and
Maneuver is Safe (MS) that comfort our previous decision-
making in node Decision1regarding to safety. The ultimate
goal of the proposed cascade decision-making strategy is
deriving the most suitable decisions given the available
evidence. To identify the most suitable decision, we compute
the Expected Utility (EU) of each decision state and the final
decision is the alternative maximizing this EU. A MLDN is
a representation of a joint expected utility function due to
the chain rule:
EU (UD) = P(UC|UD)∑
w∈UV
U(Xpa(w))(5)
Following the temporal order for this network UObsv =
(SL1,SL2,Posveh,SM)≺D1≺D2, the EU of D1given
observations will be then:
EU (D1|UObsv) = ∑
UObsv
P(UObsv)US(SL1,SL2,PosVeh,D1)
+UCh(SM ,D2)=∑
UObsv
P(SL1)P(SL2)P(Posveh)P(SM)
USaf ety (SL1,SL2,PosVeh,D1)+ UCheck(SM,D2)
(6)
The EU for the second decision D2given past observations
and decisions (D1=d1) is:
EU (D2|UObsv,D1=d1) = ∑
UObs
P(UObsv |D1,D2)UCh(SM,D2)+
US(SL1,SL2,PosVeh,D1)=∑
UObsv
P(SL1)P(SL2)P(PosVeh)P(SM)
USaf ety (SL1,SL2,PosVeh,D1)+ UCheck(SM,D2)
(7)
The most suitable decisions will be:
ρ1=max
D1
EU (D1)(8)
ρ2=max
D2
EU (D2|UObs,D1)(9)
IV. SIMULATION RESULTS
To demonstrate the robustness of the proposed approach
for handling safe highway maneuvers (cf. Fig. 6), let us
show in what follows simulation examples. In Fig. 7 and
8(a) the first set of simulation results highlights the efficiency
of the proposed MCA in normal situations. The other set
of simulation results depicted on the Fig. 8(b), 9 shows
capability of the overall MCA for anomaly detection and its
capacity to reconfigure and react by taking the appropriate
decision in emergency situations.
For the different simulations scenarios (cf. Simulation
video - https://goo.gl/aEpBSk), an exemplary overtaking
maneuver scenario is shown. The scene is constituted of three
vehicles: two vehicles on the right lane including the ego-
vehicle (that we name respectively ego-vehicle and obstacle-
vehicle 1) and one vehicle on the left lane named obstacle-
vehicle 2. The ego vehicle is traveling on the right part of
a two-lane highway. The ahead vehicle (which maintains a
constant velocity) is considered as an obstacle by the ego
vehicle. In the first part of the simulation, as the ego vehicle
(a) Head Portion at t=1.5s(b) Lane Change Part at t=3s
(c) Lane Change Part at t=4.5s(d) Tail portion at t=6s
Fig. 6. Simulation results of the overtaking maneuver. (See. Simulation
video - https://goo.gl/aEpBSk)
Fig. 7. Results of TSLDN during normal conditions.
(a) PIDP for normal conditions (b) PIDP for emergency situations
Fig. 8. Predicted Inter-Distance Profiles respect to the vehicle evolution
approaches the obstacle-vehicle, it performs a Lane change
maneuver to the left starting at t=0.85sand ending at t=
4.85s. (cf. Fig. 6). Fig. 8(a) shows the evolution of the AIDP
wit respect to the PIDP. In this set of simulation no change is
noticed regarding to the PIDP. Fig. 7 shows the EU of each
decision alternatives for both decisions: at the beginning of
the scenario the ego-vehicle approaches the obstacle-vehicle
1 (cf. Fig.6(a)) which leads the EU of the Keep Lane ACC
state to decrease while the EU of state Lane Change Left
increases. This is justified by the fact that the distances are
shrinking and SL1is approaching the state Occupied. At t=
0.85s, since Decision1is in the state Lane Change Left the
decision in node Decision2is simultaneously computed while
taking into account the AIDP. In this case, it can be noticed
that the most suitable decision is Maneuver is safe and
remains constant during the whole lane change maneuver.
After the lane change part is achieved (cf. Fig.6(b)), the
vehicle is now on the left lane, a decision to Maintain the
velocity is activated to distance the obstacle-vehicle 1 that is
overtaken (cf. Fig.6(a)). Simultaneously, while maintaining
the actual velocity, the TTC value calculated between the ego
vehicle and the obstacle-vehicle 2 initially present in the left
lane is dropping. At t=7.5s, the state Keep Lane ACC is
calculated to be the most suitable decision with the highest
EU for this configuration. At t=10.5sthe ego-vehicle is at
a convenient TTC value from the obstacle-vehicle 1 located
in the right lane and a Lane Change Right is performed as
(a) Decision 1: Lane Change Maneu-
ver
(b) Decision 2: Maneuver Safety
Checking
Fig. 9. Results of TSLDN during emergency situations
it is the alternative with the highest EU.
In the second set of simulation Fig. 9 and 8(b), we have
selected a dangerous scenario that can occur in a highway
environment where the obstacle-vehicle 1 in front suddenly
brake, while the ego vehicle is trying to perform a lane
change maneuver. In this case, if the ego vehicle does not
react quickly, a collision can occur. This scenario tests the
decision-making algorithm during the lane change maneuver
where the inter-distance is the closest. At t=2.25s, the
obstacle-vehicle 1 in front brakes strongly which results in
the modification of the AIDP evolution shown in Fig. 8(b).
The SVDL network issues then a warning, through the most
suitable decision alternative Warning Be Careful in Fig. 9(b).
This is justified by the fact that the AIDP profile is still
inside the margin. At t=2.75s, the AIDP profile goes outside
the lowest safety margin, which results in the shift from the
state Warning Be Careful to the state Abort Maneuver. This
Decision2directly affects the EU (Fig. 9(a)) of the first level
decision node namely Decision1, since the abortion of the
lane change maneuver is achieved and the MDL assess again
the risk on the lanes and the decision to Keep Lane ACC is
activated to ensure that no collision occurs with the obstacle-
vehicle 1 and any new coming obstacle-vehicle.
V. CONCLUSION
In this paper, an overall multi-controller architecture
(MCA) for safe and flexible autonomous vehicle navigation
has been proposed. It is highlighted first the adaptation of
the previous work given in [10] [11] (control architecture
and used obstacle avoidance technique, initially dedicated
for cluttered environment) in order to cope with high-
way navigation constraints. Inside this MCA, an important
module corresponding to a Two-Sequential Level Decision
Network (TSLDN) has been proposed that corresponds to
the main contribution of the paper. This module is designed
to manage several highway maneuvers under uncertainties
(which are due mainly to perceptive and/or other vehicles
intention/actions lack of precision). The TSLDN is utilized
for: the driving situation assessment, decision-making and for
safety retrospection over the current maneuver risk. A dual-
safety criteria is proposed as a measure of risk to assess the
traffic situations and other traffic participants behaviors. It is
based on an Extended Time-To-Collision (ETTC) metric and
a specific Predicted Inter-Distance Profile (PIDP) during the
lane change maneuver that improves the safety of the ego-
vehicle maneuvers. Several simulation results show the good
performance of the overall proposed control architecture,
mainly in terms of efficiency to handle probabilistic decision-
making even for very risky scenarios. Future work will be
carried out to evaluate the overall proposed approach in real-
time experimentation, mainly in collaboration with the R&D
Department of Sherpa Engineering Company.
ACKNOWLEDGMENT
This work has been sponsored by Sherpa Engineering
and ANRT (Conventions Industrielles de Formation par
la Recherche). This work has been sponsored also by
the French government research program Investissements
d’avenir through the RobotEx Equipment of Excellence
(ANR-10-EQPX-44) and the IMobS3 Laboratory of Ex-
cellence (ANR-10-LABX-16-01), by the European Union
through the program Regional competitiveness and employ-
ment 2014-2020 (FEDER - AURA region) and by the AURA
region.
REFERENCES
[1] S. E. Lee, E. C. Olsen, W. W. Wierwille, et al., “A comprehensive
examination of naturalistic lane-changes,” tech. rep., United States.
National Highway Traffic Safety Administration, 2004.
[2] J. C. Hayward, “Near miss determination through use of a scale of
danger,” 1972.
[3] J. Hou, G. F. List, and X. Guo, “New algorithms for computing the
time-to-collision in freeway traffic simulation models,” Computational
intelligence and neuroscience, vol. 2014, p. 57, 2014.
[4] F. Jim´
enez, J. E. Naranjo, and F. Garc´
ıa, “An improved method to
calculate the time-to-collision of two vehicles,” International Journal
of Intelligent Transportation Systems Research, vol. 11, no. 1, pp. 34–
42, 2013.
[5] J. R. Ward, G. Agamennoni, S. Worrall, A. Bender, and E. Nebot, “Ex-
tending time to collision for probabilistic reasoning in general traffic
scenarios,” Transportation Research Part C: Emerging Technologies,
vol. 51, pp. 66–82, 2015.
[6] K. Kim, B. Kim, K. Lee, B. Ko, and K. Yi, “Design of integrated risk
management-based dynamic driving control of automated vehicles,”
IEEE Intelligent Transportation Systems Magazine, vol. 9, no. 1,
pp. 57–73, 2017.
[7] R. Schubert, “Evaluating the utility of driving: Toward automated
decision making under uncertainty,” IEEE Transactions on Intelligent
Transportation Systems, vol. 13, no. 1, pp. 354–364, 2012.
[8] M. Althoff, D. Althoff, D. Wollherr, and M. Buss, “Safety verification
of autonomous vehicles for coordinated evasive maneuvers,” in In-
telligent vehicles symposium (IV), 2010 IEEE, pp. 1078–1083, IEEE,
2010.
[9] F. V. Jensen, “Bayesian networks and decision graphs. series for
statistics for engineering and information science,” 2001.
[10] L. Adouane, “Reactive versus cognitive vehicle navigation based on
optimal local and global pelc,” Robotics and Autonomous Systems,
vol. 88, pp. 51–70, 2017.
[11] L. Adouane, A. Benzerrouk, and P. Martinet, “Mobile robot navigation
in cluttered environment using reactive elliptic trajectories,” IFAC
Proceedings Volumes, vol. 44, no. 1, pp. 13801–13806, 2011.
[12] L. Adouane, Autonomous Vehicle Navigation: From Behavioral to
Hybrid Multi-Controller Architectures. Taylor & Francis CRC Press,
2016.
[13] J. Vilca, L. Adouane, and Y. Mezouar, “A novel safe and flexible
control strategy based on target reaching for the navigation of urban
vehicles,” Robotics and Autonomous Systems, vol. 70, pp. 215–226,
2015.
[14] R. Worrall and A. Bullen, “An empirical analysis of lane changing on
multilane highways,” Highway Research Record, no. 303, 1970.
[15] S. J. Russell and P. Norvig, Artificial intelligence: a modern approach.
Malaysia; Pearson Education Limited,, 2016.
