IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO.
2, JUNE 2007
Collision Avoidance for Vehicle-Following Systems
Stefan K. Gehrig and Fridtjof J. Stein
Abstract—The vehicle-following concept has been widely used
in several intelligent-vehicle applications. Adaptive cruise control
systems, platooning systems, and systems for stop-and-go traffic
employ this concept: The ego vehicle follows a leader vehicle
at a certain distance. The vehicle-following concept comes to its
limitations when obstacles interfere with the path between the ego
vehicle and the leader vehicle. We call such situations dynamic
driving situations. This paper introduces a planning and deci-
sion component to generalize vehicle following to situations with
nonautomated interfering vehicles in mixed traffic. As a demon-
strator, we employ a car that is able to navigate autonomously
through regular traffic that is longitudinally and laterally guided
by actuators controlled by a computer. This paper focuses on
and limits itself to lateral control for collision avoidance. Pre-
viously, this autonomous-driving capability was purely based on
the vehicle-following concept using vision. The path of the leader
vehicle was tracked. To extend this capability to dynamic driving
situations, a dynamic path-planning component is introduced.
Several driving situations are identified that necessitate responses
to more than the leader vehicle. We borrow an idea from robotics
to solve the problem. Treat the path of the leader vehicle as an
elastic band that is subjected to repelling forces of obstacles in the
surroundings. This elastic-band framework offers the necessary
features to cover dynamic driving situations. Simulation results
show the power of this approach. Real-world results obtained with
our demonstrator validate the simulation results.
Index Terms—Computer vision, intelligent vehicle, robotics,
stereo vision.
I. I NTRODUCTION
C OLLISION avoidance has been the subject of extensive
research both in the fields of robotics and intelligent vehi-
cles. Since research on autonomous robots has been conducted
for more than two decades, the lessons learned in that field also
benefit the intelligent-vehicle research. A tremendous benefit
is assessed for reducing collisions with automated systems in
regular traffic [1]. The detection capabilities of vision-based
intelligent vehicles are mature enough to perform such a task
(see, e.g., [2]).
For lateral vehicle guidance on highways, vision-based lane
following is a promising and well-tested strategy (see, e.g., [3]).
Recent work on lane recognition in urban areas is described in
[4]. However, lane following relies on lane markings and tends
to fail in cluttered scenes and in urban areas. Hence, a popular
approach for automated vehicle guidance is to follow a leader
Manuscript received March 1, 2004; revised October 13, 2005, June 7, 2006,
October 7, 2006 and October 11, 2006. The Associate Editor for this paper was
S. Darbha.
The authors are with the DaimlerChrysler Research and Technology AG,
71059 Sindelfingen, Germany (e-mail: stefan.gehrig@dcx.com; fridtjof.stein@
dcx.com).
Digital Object Identifier 10.1109/TITS.2006.888594
1524-9050/$25.00 © 2007 IEEE
233
vehicle both longitudinally and laterally. This intuitive behavior
comes to its limitations when other traffic participants interfere
with the leader-vehicle’s path. Dynamic path modification be-
comes necessary. These dynamic driving situations motivated
the work of this paper.
How is this paper organized? Section II gives an overview
of related work both in the fields of intelligent vehicles and
robotics. In Section III, the elastic-band framework is intro-
duced. Section IV describes the necessary adaptations to the
elastic-band framework in order to tune the avoidance behavior
toward that of human beings. In Section V, the extension of
the elastic-band approach to nonholonomic vehicles is briefly
discussed. The impact of the elastic-band framework on the
planning and decision module is described in Section VI.
Results are detailed in Section VII. Conclusions and future
work comprise the final section.
II. R ELATED W ORK
A. Potential Field Approaches
The most prominent idea for collision avoidance in robotics
applications is artificial potential fields (e.g., [5] and [6]). The
obstacles are modeled as potentials and the gradient of the
superimposed potential field yields the direction command for
the mobile robot. Potential fields are an elegant way to model
obstacles and can be analyzed globally. A formal verification of
the obstacle-avoidance behavior is analytically feasible.
Generalized potential fields depend—not only—on distance
to obstacles but can also depend, e.g., on obstacle velocity.
Therefore, these fields can determine irrelevant obstacles mov-
ing away from the ego vehicle. This method was first applied
using velocity-dependent potentials in [5].
A popular approach to avoid local minima, i.e., trapping
situations, in potential field applications is the use of harmonic
potentials (see, e.g., [7] and [8]).
B. Approaches Using Physical Models
Other physical models besides potential fields are also popu-
lar for collision avoidance.
An alternative approach to the potential field method is
presented in [9], where analogies of this problem to hydro-
mechanics problems are shown. Fluid dynamics equations are
used. The fluid starts at the starting point toward the goal point,
and obstacles obstruct the flow. From the resulting flow field,
the planned path can be computed.
In [10], an approach to behavioral control of robots is de-
scribed that uses the model of dynamical systems. Obstacles
generate a stimulus that repels the autonomous robot. The
path for navigation and obstacle avoidance is generated solving
234 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007
dynamical differential equations. The approach is rather time-
consuming due to the necessary relaxations. For this approach,
a low-level sensory information as opposed to symbolic obsta-
cle information is used.
Quinlan and Khatib present another approach to obstacle
avoidance that uses the model of an elastic band for a robot
(see, e.g., [11] and [12]). The elastic-band framework is used
for our application. An initial path has to be supplied. This path
is modeled as an elastic band that is subjected to forces and
exerted by obstacles, which are represented as potential fields.
This approach has also been extended to nonholonomic mobile
robots [13]. Brock extends that idea to the many-dimensional
configuration space of mobile manipulators [14] and obtains
speeds of up to 2 m/s in indoor robotics environments [15].
Other approaches often include rule-based systems or sys-
tems executing behavioral patterns. These systems are not
covered here due to space limitations.
C. Approaches for Intelligent Vehicles
A potential field application in driver assistance applications
is described in [16]. Vehicle dynamics is taken into account, and
a combination of different drive–assist systems is feasible based
on force control. Simulation results are presented. Extensions to
this approach, including an experimental validation on a sports
car, are presented in [17]. Similarly, lateral control for vehicle
following is presented in [18], where dynamic aspects for heavy
trucks are considered.
A completely autonomous car with obstacle-avoidance capa-
bility is presented in [19]. The application of potential fields
to determine the risk yielded oscillatory behavior in overtaking
situations. As a consequence, behavioral patterns were imple-
mented for obstacle-avoidance maneuvers. Longitudinal and
lateral maneuvers are presented.
A very interesting approach both for robotics and intelligent-
vehicle applications is presented in [20]. In contrast to most
other approaches, obstacle avoidance is performed in configu-
ration and in velocity space. Constant velocity is assumed for
all obstacles. However, this constraint is not always met.
Obstacle velocity is also considered in [21]. The basic idea
stems from aerospace research dealing with collision achieve-
ment in missile guidance. Arbitrary shapes of the obstacles
are considered. Again, constant velocity is assumed for all
obstacles. This assumption leads to undesirable behavior in the
vicinity of intersections, where most vehicles slow down and
avoidance space is limited to the paved street.
An investigation on string instability for vehicle follow-
ing has been conducted in [22]. This paper especially shows
limitations for such approaches, especially for high dynamic
scenarios.
Summarizing, there has been a tremendous amount of work
performed especially concerning vehicle dynamics for vehicle-
following scenarios. However, most of them considered only
one input to follow—either the lane boundaries or the leader
vehicle. In this paper, we focus on the necessary planning
component, when more than one input from the environment
has to be considered, and neglect the vehicle dynamics aspects
for the most part.
Fig. 1. Representation of an elastic band as a series of particles with springs
in between.
III. E LASTIC B ANDS
A. Introduction to Elastic Bands
The original elastic-band approach for collision avoidance
was proposed by Quinlan and Khatib [11]. Similar physical
properties are used for snakes in computer vision [23]. An
initial feasible path must be delivered by a path planner. This
path is dynamically modified by treating the path as an elastic
band that is able to change its shape. The starting and end points
are kept fixed. The energy of the elastic band is determined by
an internal energy similar to a spring energy and an external
energy due to obstacles modeled as potentials. The total en-
ergy of the elastic band is minimized, yielding smooth paths
(see Section IV-E). Forces acting on the elastic band are
computed by taking the gradient of the potential energy at
discrete path points. The repelling forces on the elastic band are
produced by obstacles in the vicinity of the path. In total, three
types of forces are acting on the elastic band. These forces are
described in the following sections, where we base our analysis
on [24]. The treatment of the continuous case does not provide
any additional insights and can be found in [24] as well.
For the underlying theory of elastic bands, see books on
Classical Mechanics (e.g., [25]).
B. Internal Contraction Force
The elastic band is modeled as a series of particles with a
series of springs in between in the discrete case (see Fig. 1 for a
representation of an elastic band). Hooke’s law would suggest
a force proportional to the amplitude of the displacement.
However, with defining zero as the unstrained state, the internal
force would become larger as the length of the elastic band
increases. This would yield an avoidance behavior that depends
on the length of the elastic band, which is undesirable. Hence,
we follow Quinlan’s approach, normalize the force to be only
dependent on band curvature, and let the internal potential be


n−1
Vint = kc qi+1 − qi  (1)
i=1
where qi represents the ith configuration particle along the
elastic band, kc is the constant of elasticity, and Vint is the
internal potential of the elastic band. The elastic band consists
of n particles with n − 1 springs in between.
GEHRIG AND STEIN: COLLISION AVOIDANCE FOR VEHICLE-FOLLOWING SYSTEMS
Consequently, the internal force fint is computed via
 
 qi+1 − qi qi−1 − qi
fint (qi ) = −∇q Vint = kc + . (2)
qi+1 − qi  qi−1 − qi 
This normalized force ensures the same avoidance behavior
independent of the band length.
C. External Force
The external force is due to obstacles that are modeled as
potentials in the scene. Any potential shape that repels the
elastic band from obstacles is conceivable. We decided to use
Quinlan’s position-dependent potential [24] superimposed by
Krogh’s velocity-dependent potential [5]. The gradient of this
potential fext yields the external force. Details on the external
force follow in Section IV-E.
D. Constraint Force
The elastic band potentially reduces its internal energy by
moving particles along the elastic band. This is an undesired
property since the band might thin out at some parts. To
constrain the motion along the elastic band, a constraint force
fconstr is introduced with the direction t
t = qi+1 − qi + qi − qi−1 (3)
qi+1 − qi  qi − qi−1 
tangential to the elastic band. The force is computed by projec-
tion of the external force along the elastic band
fext , t  
fconstr = − ·t (4)
|t 2 |
where ,  denotes the scalar product.
E. Elastic-Band Algorithm
The elastic band is simulated with reduced dynamics since
only the equilibrium state is of interest for obstacle avoidance.
For every time step, the algorithm is repeated. The path is
represented by particles qi , and these particles are subjected to
the total force
ftot = fint + fext + fconstr . (5)
With our pseudostatic simulation, the particles are moved
along the total force to the new position
qi,new = qi,old + α · ftot (qi ) (6)
with α = (dt2 )/(2m), all assumed constant. Higher order
terms beyond particle acceleration are neglected here. This
procedure represents a very simple gradient descent method
performing Euler integration of the underlying partial differ-
ential equations with a constant step size. Applying this pro-
cedure iteratively yields a linear convergence and proved to be
sufficiently fast for our real-time application. The new configu-
rations are recomputed until the total force of all configuration
235
points is sufficiently small, i.e., until the force equilibrium is
reached.
Pathological situations such as two obstacles located closely
to the left and right of the initial path might yield an oscillatory
behavior. We reduce the moving factor α after a certain number
of iterations to reduce these oscillations.
Brent introduces an optimization algorithm that minimizes a
function of one variable without computing derivatives and per-
forms very few evaluations of that function [26]. This procedure
is used in the original approach to find the force equilibrium,
but it is less efficient for situations where only very few path
modifications occur.
In the original elastic-band approach, the moving procedure
is supplemented with a procedure of adding and removing par-
ticles in order to maintain a collision-free path at all times (see
Section IV). So-called bubbles model the available free space
around a configuration. Adding particles becomes necessary
when two adjacent bubbles no longer overlap. Bubbles are
circles with a radius of the minimum distance to the closest
obstacle for our 2-D case.
To speed up convergence of the elastic-band algorithm, the
final configuration from the last time step can be used as the
initial configuration for the next time step. Path transformations
have to be performed to account for the interframe egomotion.
The required modifications of the original approach for the
vehicle-following scenario are detailed in the next section.
IV. A DAPTATIONS FOR V EHICLE F OLLOWING
A. Basic Idea
The elastic-band approach is adapted to the vehicle-
following scenario here. The path of the leader vehicle con-
stitutes the initial path. Obstacles in the environment exert
forces on the band and move it into a final configuration.
This is the path the ego vehicle tracks subsequently. Due to
the motion of the leader vehicle, the end point of the elastic
band moves as well. Moving obstacles are accounted for by
modeling the obstacle potential to be velocity-dependent. There
are configurations where the final path is still infeasible, in
contrast to the original elastic-band approach. In order to find
these configurations, a final clearance check along the envi-
sioned ego-vehicle path is performed. If the clearance check
fails, the autonomous-driving mode is switched off, and the
driver has to take over the control of the vehicle manually.
Also, when the lateral acceleration exceeds a comfortable limit,
the autonomous-driving mode is switched off as well. This
procedure allows us to keep longitudinal and lateral control
decoupled.
The timing for the “switching off” of the autonomous-
driving mode is a very involved and tedious task. A lot of
research has been conducted in that area (see, e.g., [27] for
reaction time statistics). However, these human-machine-
interface issues are beyond the scope of this paper.
Driving situations that necessitate this dynamic modification
of the leader-vehicle path are
1) passing a cyclist or pedestrian along the road while fol-
lowing a leader vehicle;
236 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007

2) driving too close to parking cars due to the smaller size the x−z-plane and cannot be adequately described by rectan-
of the leader vehicle; gles. The convex hull is computed with a recursive algorithm
3) driving around a car that comes to a stop slightly inside described in [31] that works in O(n · log(n)), where n is the
the road intersection; and number of points.
4) driving around a pedestrian that moved one step into The relative velocities of the objects with respect to the
the road. ego vehicle are computed based on the center of the obstacles
Many more scenarios can be identified. In addition, further ex- bounding boxes using an extended Kalman filter, assuming
tensions of the autonomous-driving capability of such a vehicle rigid objects. One of the extracted objects is the leader vehicle
definitely need this dynamic path-planning component. supplying the initial path. The other detected objects constitute
The following sections cover the sensor system, details on obstacles.
the distance computation to obstacles in the scene, algorithmic
modifications to the elastic-band approach, potential shapes for C. Efficient Distance Computation
obstacles and lanes, and the actual integration of the approach
The distance of the ego-vehicle path to the obstacles is the
in our demonstrator.
most important quantity for the determination of the external
force on the elastic band. Since the distance is computed
B. Sensor System thousands of times in one cycle, it is of uttermost importance
to compute it fast.
Our range sensor delivers a list of objects in the scene. We
If the ego-vehicle path is not influenced by the potential
use a calibrated stereo camera system to obtain 3-D information
generated by the bounding box of the obstacle, no refined
of the car’s surroundings. However, the elastic-band algorithms
distance computation is necessary, and the distance between the
in this paper are applicable to any sensor delivering range data
bounding boxes is sufficient. The algorithm described below is
with a “suitable” angular resolution.
only necessary for obstacles being close to the ego vehicle. This
The calibrated stereo camera system delivers 3-D measure-
coarse-to-fine strategy saves a significant amount of computa-
ments of significant points resulting in a sparse 3-D-point cloud
tion time.
(less than 500 3-D measurements). The significant points of the
Distance computation between two convex polyhedra is de-
left image are matched in the right image by correspondence
scribed, e.g., in [32]. This algorithm is comparatively fast in
analysis along the epipolar line similar to [28]. With a baseline
general but solves the distance computation in more than two
of about 30 cm, we are able to detect obstacles robustly up to
dimensions. Lin introduces a good incremental algorithm for
40 m with decreasing accuracy for faraway objects. The lateral
the distance computation, but the initial distance computation
position precision of the leader depends on many circumstances
is slow [33]. Faster implementations are available in two di-
(visibility conditions, shape of the leader vehicle, . . .). Empir-
mensions. A very simple method in robotics to account for
ically, we compared the position against hand-labeled leader-
the dimensions of the ego vehicle and the obstacles in distance
vehicle positions and determined a standard deviation of about
computation is obstacle growing. There, the obstacles grow by
20 cm. To further increase lateral position accuracy, one could
the size of the ego vehicle. However, this works only accurately
also apply classification schemes such as described in [29]. It
for circular ego vehicles.
is also possible to employ modern dense stereo techniques to
In our approach, the distance computation between a rectan-
obtain dense 3-D clouds [30].
gle (our ego vehicle) and a convex hull (obstacle) is reduced
To extract objects from 3-D measurements, we apply a spatial
to a distance computation between two line segments. The ego
clustering method to all 3-D points, except the ones on or below
vehicle’s orientation is parallel to the tangent on the elastic
the ground. Here, we assume a constant orientation between the
band. When the ego vehicle and the obstacle do not intersect,
camera system and the flat road.
the distance is computed from all points of the ego vehicle to
One difficulty in combining range data from different frames
all obstacle points. The points with the minimum distance are
is the matching between consecutive frames. How can we
taken, and the distance of the two adjacent line segments on
match data from one frame to the next? On the symbolic
both sides is computed subsequently. Thus, a total of four line-
object level, this is straightforward. We match objects from
segment-to-line-segment distance computations are necessary.
different frames by requiring them to be close together in space.
The distance between the obtained obstacle line seg-
This procedure works well for small scene changes between
ments and the ego-vehicle line segments is computed using
consecutive frames.
Lumelsky’s algorithm [34]. This algorithm is about five times
We assume a flat environment and use only the x and z
faster than the straightforward approach.
coordinates of the objects for collision avoidance.
One object corresponds to one-clustered 3-D-point cloud
D. Modifications to the Original Elastic-Band Approach
projected to the x−z-plane represented by its convex hull for
collision avoidance purposes. A convex hull of a set of points Taking the elastic-band framework over to the intelligent-
is defined as the smallest convex polygon containing the points. vehicle realm, the equations introduced in Section III stay the
In our algorithm, the convex hull represents the outline of an same. The distance computation of the previous section is used
obstacle, which is better suited than the rectangular description to compute the external force. Since our vehicle-following sys-
such as a bounding box. Moreover, objects such as a crowd of tem operates at small velocities, no vehicle dynamics are taken
people standing close to each other exhibit arbitrary shapes in into account. In addition, including vehicle dynamics would
GEHRIG AND STEIN: COLLISION AVOIDANCE FOR VEHICLE-FOLLOWING SYSTEMS
also suggest a coupling of longitudinal and lateral control,
which we wanted to avoid to make the algorithm applicable and
independent of the chosen longitudinal control strategy.
The vehicle-following application of the elastic-band frame-
work differs from the original idea in the following ways.
1) The initial path cannot be guaranteed to be collision
free. Since the initial path is created by the path of the
leader vehicle, changes in the scene might have occurred
since the leader vehicle has passed. It is possible that
another traffic participant has approached that path in
the meantime. In addition, the leader vehicle might be
smaller in width than the autonomous vehicle and might
have chosen a path very close to an obstacle.
2) The equilibrium position of the configuration particles is
the leader-vehicle path. We want the autonomous vehicle
to follow exactly this path in the absence of obstacles.
Hence, the internal forces of the initial configuration
receive an offset yielding zero in the absence of obsta-
cles. The internal force used for our application is thus
computed as
 the repulsion factor for obstacles can be varied according
 qi+1 − qi qi−1 − qi
fint (qi ) = kc +
qi+1 − qi  qi−1 − qi 

qinit,i+1 − qinit,i qinit,i−1 − qinit,i
− −
qinit,i+1 − qinit,i  qinit,i−1 − qinit,i 
where qinit,i is the initial configuration of the ith particle.
This also results in an offset term for the total energy of
the band. This behavior was chosen to maintain the intu-
itive vehicle-following behavior. Hence, the best strategy
is to follow the leader-vehicle path unless obstructions
occur. In the original approach, the elastic band was
assumed to have zero length and would always collapse
to a straight line in the absence of obstacles.
3) The concept of bubbles of free space disappears automat-
ically by allowing initial trajectories that are not collision
free. By giving up on the bubble concept, we need an
additional algorithm to reassure a collision-free path after
the equilibrium position has been found. This is done by
geometrically checking for overlaps along the final con-
figuration of the elastic band. Since we model our leader-
vehicle path as a polygon and our obstacles as convex
hulls, a simple line segment intersection algorithm serves
the purpose of the overlap check. In our application, we
keep the number of particles constant and equally spaced
(typically 1–5 particles/m, the results are presented with
2 particles/m). The procedure of adding and removing
particles becomes superfluous as well. More generally,
a collision-free path can never be guaranteed in the
presence of moving obstacles. A drastic example would
be a faster car than the one that intentionally wants to
provoke a collision. Provided sufficient driver skills in the
“hunter” car, a collision is unavoidable.
4) A regular car is often treated as a nonholonomic ve-
hicle. The original algorithm applies only to holonomic
robots. This limitation has not been explicitly taken into
account. Hence, the paths created by the elastic band
could be infeasible for nonholonomic cars. Since uncer-
237
tainties in the 3-D measurement of objects and uncertain-
ties in the vehicle control require additional slack around
the planned path, this is not considered a serious problem.
In addition, the nonholonomic constraint also applies to
the leader vehicle and is reflected in the leader-vehicle
path. However, extending this algorithm to scenarios such
as parallel parking, the nonholonomic constraint has to
be considered. Work on elastic bands for nonholonomic
vehicles has been presented in [13]. Section V also dis-
cusses this problem.
5) In regular traffic situations, lane markings are also used
for vehicle guidance. The lane markings are detected
and are also modeled as virtual obstacles with repelling
forces pointing away from the lane boundaries toward
the lane center. Lanes are modeled as polygons so the
same distance computation algorithm as for obstacles is
used. More details on the potential shape are presented in
Section IV-F.
6) When a classification for the obstacle into pedestrian, car,
truck, bicyclist, or other traffic participant is available,
to the classification result. This way, an additional safety
envelope can be applied to pedestrians.
7) The external force of the elastic band must be shaped to
(7) comply with natural driving behavior. Drivers keep more
distance from obstacles at high speed than at low speed
(e.g., parking situations). Therefore, a velocity-dependent
potential shape must be used. The human-like avoidance
behavior is obtained by empirical parameter tuning of the
parameters appearing in (1)–(13). Details of the external
potential follow in the next section.
E. Potential Shape for the Obstacles
We have experimented with a lot of potential shapes and
came to the conclusion that potentials with limited reach best
model human driving behavior. Usually, events and objects
further away than a certain distance from our planned path do
not have an impact on our actually driven path.
The original approach for elastic bands is designed for
mobile robots that rarely exceed 2 m/s. Therefore, a position-
dependent potential is sufficient. We modified this approach in a
way that the effective reach of Quinlan’s potential deff becomes
a function of velocity
 2
Vext1,j (q) = 2 kr (deff − dj (q)) , if dj (q ) < deff (8)
1
0, otherwise
deff = d0 + dinc · vego (9)
where Vext1,j denotes Quinlan’s potential of obstacle j at
position q. For the presented results, we pick deff = 0.3 m and
dinc = 0.02 s. dj (q) denotes the distance between obstacle j
and particle q of the elastic band; kr is the repulsion gain.
dinc increases the effective reach of the potential for increasing
velocities.
In addition, we superimpose a velocity-dependent potential
introduced by Krogh [5]. Here, we take the velocities of the
obstacles into account. The magnitude of the potential depends
238 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007

on the relative velocities between obstacle and ego vehicle.

Since we already know an initial path, only obstacles close to
that path are considered. A fading function for every obstacle
j, ffad,j is multiplied with the Krogh potential Vext2,j (q). This
limits the impact of approaching obstacles to a vicinity of the
leader-vehicle path. In real traffic situations, only approaching
vehicles in adjacent lanes are relevant for our driving behavior
(this neglects the intersection case, but for that, a very large field
of view would be necessary)
 avrel,j
 ffad,j 2·dj ·a−v 2 , Tj > τj , vrel,j > 0
rel,j
Vext2,j (q, vrel,j ) = 0, vrel,j ≤ 0

∞, otherwise
(10)
 2
dkrogh − dj,min
ffad,j = . (11)
dkrogh

vrel is the projection of the relative velocity on the direction

vector between ego vehicle and obstacle. a represents the
maximal possible deceleration. dj,min denotes the closest
distance of the jth obstacle to the initial elastic-band path. Tj
and τj are minimum (τj ) and maximum (Tj ) avoidance time,
as defined in [35]:

2dj (q )
Tj =
vrel,j
vrel,j Fig. 2. Visualization of the chosen obstacle potential shape. For the situa-
τj = . (12)
a tion depicted in the upper part, the resulting potential can be viewed in the
lower part.
Note that dj,min is only available for obstacle-avoidance
algorithms that plan the whole path. Algorithms only yielding
where dj (q) is the direction vector from obstacle j to particle
current heading and accelerations do not deliver the information
q. The total external energy is computed from
to compute that quantity.
The idea behind Krogh’s velocity-dependent potential is the
Vext = Σj Σi (Vext1,j (qi ) + Vext2,j (qi )) . (14)
following: The minimum avoidance time represents a braking
maneuver with maximum deceleration a. The maximum avoid-
The total energy is easily computed by Etot = Vint + Vext with
ance time uses the full distance to the obstacle and brakes
the constraint force producing no energy.
with a constant deceleration. The reserve avoidance time Tres
When an obstacle is unavoidable for or inside a particle of
is Tres = Tj − τj . The inverse of that reserve time to respond
the elastic band, a maximum force is exerted toward a direction,
to an obstacle represents the potential computed in (10). Note
where obstacle clearance is obtained on the shortest way. De-
that the velocity-dependent potential described here inherits its
tails on how to find this direction are presented in [36]. Roughly,
name from the fact that the potential magnitude and shape de-
this direction is obtained in a simple minimization step, where
pend on the relative velocity between ego vehicle and obstacle.
the obstacle convex hull is approximated by a rectangle.
However, within one time step, this velocity is constant. We
Several restrictions regarding building the gradient of the
use 5 m/s2 as maximum deceleration, which reflects somewhat
potential field apply, e.g., the potential might not even be
suboptimal road conditions.
differentiable at some positions [24]. This might cause the
Fig. 2 shows an example of the chosen potential. In that
algorithm to compute a wrong force in one iteration, but be-
scenario, the obstacle on the right is stationary, whereas the left
cause of a movement of most particles in the elastic band, the
obstacle moves at half the speed of the ego vehicle, which is
situation is resolved in the next iteration. Pathological cases
10 m/s (upper part). The superposition of Quinlan’s position-
might occur where the minimum energy configuration is not
dependent potential and Krogh’s velocity-dependent potential
achieved (local minima, symmetrical situations). In these rare
is depicted in the lower part of Fig. 2.
situations, the final geometrical clearance check assesses the
The total external force per particle is thus
risk of a collision and a potentially unnecessary “switch off”
of the autonomous system is performed. It is not desirable to
dj (q )
fext,j (q ) = kr (deff − dj (q )) · find arbitrary collision-free paths since we want to stick close
dj (q )
to the vehicle-following principle. This is the “advantage” of
Vext2,j (q, vrel,j )2 dj (q ) designing an autonomous system with a human fallback option
+2· · (13)
vrel,j dj (q ) in reserve.
GEHRIG AND STEIN: COLLISION AVOIDANCE FOR VEHICLE-FOLLOWING SYSTEMS
Fig. 3. Visualization of the chosen lane potential shape. For the situation
depicted in the upper part, the Quinlan potential for lanes is depicted in the
lower part.
The parameters for the potentials have been selected based on
a unit mass (= 1 kg) spring-mass model. The elasticity constant
is 80 N/m.
F. Potential Shape for the Lanes
Lane markings provide an important guidance for driving.
Hence, we use lanes to support the vehicle-following procedure
by modeling them as virtual obstacles.
To detect lanes, the 3-D points already determined for object
detection are used. 3-D points below a certain height are
accounted for lane markings. Using some additional models
such as lane width, a lane is fitted through the 3-D points.
The lane model estimates the current lane position, orientation,
curvature, and change of curvature with a Kalman filter (see,
e.g., [3]). This result is sampled at discrete points, which yields
a polygon. The polygon is input to the elastic-band algorithm.
Only one lane to the left and one lane to the right (if existent)
of the ego position are used for the algorithm. The potential for
lanes is modeled with the Quinlan potential. Being outside the
lane yields a maximum potential, the potential in the middle
between the lanes is zero. Fig. 3 shows an example of a lane
potential.
When driving close to the lane center, no force results from
the lane potential. Only when driving close to the marking or
239
even outside the lane, a force toward the lane center is exerted.
The maximum lane force is smaller than the maximum obstacle
force to account for the effect that crossing a lane line is less
dangerous than hitting a real obstacle.
Taking both lane and obstacle information into account
constitutes a fusion of lane-following and vehicle-following
behavior resulting in a humanlike driving behavior.
G. Control With the Elastic Band
The result of the elastic-band algorithm is a path that has
to be tracked. Considering the leader-vehicle path as the initial
path to be tracked, the elastic-band framework delivers a mod-
ified path. Path tracking is performed by selecting a path point
at a certain lookahead distance. For all results presented here,
we used a minimum lookahead distance of dl = 5m + 0.7s ·
vego . This path point is approached taking vehicle kinematics
into account (see [37] for a detailed description). The control
behavior of negotiating a path at a certain lookahead distance
can also be found when observing humans driving through a
curve and tracking their fixation point [38].
The lateral controller to track the aforementioned path point
is a simple tractor controller (see, e.g., [39]) that follows that
point on a straight line and driving through curves with constant
curvature.
Since the elastic-band algorithm operates statically on the
current scene snapshot and is rerun at every time step, no
connection between the elastic band results from one frame to
another exists. Sudden changes in the controller input might
cause an oscillatory behavior. Hence, a low-pass filter is applied
to the resulting desired steering angle delivered by the elastic-
band path.
V. V EHICLE D YNAMICS A SPECTS
The elastic-band framework presented so far only delivers
feasible paths for holonomic vehicles. A car obviously is often
treated as a nonholonomic vehicle since it has a minimum turn-
ing radius that limits the heading changes. Vehicle dynamics
limits the maximum curvature of feasible paths even more.
This nonholonomic restriction is taken into account in [13].
The induced metric not only considers the Euclidian distance
to an obstacle but also takes the current heading direction into
account. This metric satisfies the requirement of a minimum
turning radius and also applies to the internal force. It tends to
contract, but the zero position is modified: A straight line in an
obstacle-free environment cannot be obtained. The result is a
circular arc. The external force has a nonconvex potential shape
around the border of the minimum turning radius, which leads
to stability problems.
A special interpolation and smoothing algorithm is needed
in this implementation since the number of bubbles of free
space must be kept small. The real-time implementation is
made feasible through elimination of configuration points
along the path. Therefore, interpolation and smoothing are
needed.
The result is an algorithm taken far away from the intu-
itive elastic-band idea. The backbone of this approach is the
240 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007

representation of free space, which is not applicable to the

vehicle-following scenario.
Therefore, we decided to allow crooked paths implying
turns below the minimum turning radius. While following the
path, our demonstrator with the resulting vehicle dynamics
smoothes out the path. With sensible parameterization of the
internal elastic-band force, a feasible path for all typical traffic
situations is obtained. In addition, tight turns are suppressed
by the lateral controller anyway to provide sufficient ride
comfort. When the vehicle is forced to an infeasible turn due
to obstacles, the deviations to the planned path are noticed
immediately since the elastic-band algorithm runs in real-time.
As a consequence, the control of the vehicle is given back to
the driver.

VI. P LANNING AND D ECISION FOR E LASTIC B ANDS Fig. 4. Simulation result of the elastic-band algorithm (bird view) utilizing
both lanes and obstacles.
In the standard vehicle-following approach, a planning and
decision module selects the leader vehicle and sends the leader-
vehicle position to the controller. A leader vehicle is initially
selected by projecting the planned corridor forward using the
current steering angle, and the first vehicle to intersect with this
corridor is the leader vehicle. The driver has to confirm this
leader-vehicle selection manually.
The elastic-band framework can be integrated easily using
a path-based approach for vehicle following [37]. A reference
frame at rest is created by motion-integration of the own vehicle
motion. The leader-vehicle position is transformed into that
coordinate system for every time step, which yields the leader-
vehicle path. Only the leader-vehicle path has to be exchanged
with the elastic-band path. Due to measurement noise of the
leader-vehicle position, the final elastic-band path is smoothed
using a low-pass filter for the curvature.
The longitudinal control is implemented in such a way that Fig. 5. Simulation result of the elastic-band algorithm using lanes only (bird
we follow the leader vehicle at a safe distance (see, e.g., [40] view).
for such a controller). The same functionality can also be
implemented using fuzzy logic [41]. The elastic-band algorithm VII. R ESULTS
only affects the lateral control. For the rather small avoidance
A. Simulation Results
maneuvers we want to cover within our vehicle-following sys-
tem, this is considered sufficient. For a completely autonomous 1) Simulation Results Without Lateral Control: The initial
system, lateral and longitudinal control must obviously be verification of the implemented algorithms was performed in
coupled. an offline environment, where only a snapshot of a scene was
The corridor to be searched for the leader vehicle becomes analyzed. Parameter tuning was performed in this environment.
the corridor around the elastic-band path. Hence, when small In Fig. 4, a situation is shown, where the leader vehicle is
intersections with an obstacle occur, the elastic-band algorithm slightly driving out of lane and other cars drive to the left and
bends the path in such a way that no intersection with the right. Hence, a slight path modification to the right occurs due
obstacle occurs. As a safety measure, the bent of the elastic- to the obstacles as well as slightly due to the lane markings
band path is monitored, and when deviations beyond a certain with the elastic-band algorithm. Obstacles are modeled as
threshold occur, the control is returned to the driver. Although rectangles for simplicity. The equilibrium state is found within
the elastic band is still feasible in such a case, the behavior 18 iterations.
of the ego vehicle would already be far from that of typical Fig. 5 shows an example how the algorithm works in the
vehicle-following system. After such a situation occurred, a absence of obstacles but with lanes. When the leader vehicle
new leader vehicle must be selected. slightly drives out of lane, the ego vehicle is guided back into
The same situations occur when the leader vehicle disappears the lane. No modification of the leader-vehicle path occurs in
from the field of view of the camera system. Control is returned the absence of obstacles and lanes.
to the driver, and a new leader vehicle must be selected. With 2) Simulation Results With Lateral Control: Collision
the use of car-to-car communication and GPS localization, one avoidance maneuvers are dangerous to conduct in the real
could circumvent this limitation. world. It is not safe to interfere with a car’s path when it
GEHRIG AND STEIN: COLLISION AVOIDANCE FOR VEHICLE-FOLLOWING SYSTEMS
Fig. 6. Bird view for a situation where a pedestrian slightly enters the driving
corridor, which necessitates a swerve maneuver using a faster actuator. A
steering actuator with about 150-ms delay is used. The ego vehicle and the
leader drive at 10 m/s, and they are about 20 m apart. An avoidance maneuver
with a comfortable safety margin results.
moves. Maneuvers with especially little clearance to obstacles
are critical. Therefore, we decided to perform some of these
maneuvers in simulation first.
In Fig. 6, an avoidance maneuver is presented, where a
pedestrian slightly overlaps with the driving corridor. Hence, a
swerving maneuver is performed. This maneuver is performed
in a simulation where the full controller code and a simple
vehicle model of the ego vehicle are employed. Our vehicle
model considers the bicycle model for lateral dynamics includ-
ing the self-steering gradient. Tire forces are not modeled, since
we are not interested in highly dynamic maneuvers here. A
fast actuator is necessary to keep the deviation of the actual
path to the desired path small. The desired path as output
from the elastic-band algorithm is depicted in a dashed line for
every time step. These dashed lines are visible when notable
deviations between the desired and the actual path occur.
B. Real-World Results
1) Real-World Results Without Lateral Control: We in-
tegrated our elastic-band framework in the basic vehicle-
following system. Computing the modified path runs in real
time without much optimization. With our current parame-
terization, the elastic band finds its equilibrium state within
20 iterations for most situations.
Fig. 7 shows a typical traffic scene with counter traffic on the
left and a parked obstacle (a dredger) on the right.
The leading vehicle drove very close to the dredger (see
initial path in Fig. 8) and has a smaller car than our research
vehicle. Hence, following the initial path brings us dangerously
close to the dredger. The elastic-band approach yields the
dashed path in Fig. 8, which avoids the dredger and keeps a
safety distance from the car on the left at the same time. Lane
markings were not used here.
As can be seen from this example, approaching vehicles on
adjacent lanes have little impact on the ego path unless these
vehicles come very close to the ego path.
241
Fig. 7. Traffic scene with a dredger at the right side of the street and with
counter traffic.
Fig. 8. Bird view for the dredger scene. The leader vehicle is located at the
end (z = 17.5 m) of the path and is smaller than the ego vehicle.
2) Real-World Results With Simulated Obstacles: Further
experiments have been conducted using the elastic-band path
as input to the lateral control. Results using simulated obstacles
in the research vehicle are described in the following.
Above simulation results encouraged us to perform the test in
our demonstrator: a Mercedes Benz E-class 420. It is equipped
with electronic gas, brake, and steering wheel, enabling au-
tonomous motion. On-board sensors deliver vehicle velocity
and steering angle that are used to integrate the ego motion
for ego-position estimation. This is necessary to determine the
leader-vehicle path [37]. In addition, a yaw rate sensor could be
used to account for effects such a side-wind and tilted roads.
Furthermore, ego-position estimation can be enhanced using
natural landmarks observed in the traffic scene [42].
We simulated obstacles in the research vehicle to test the
algorithm safely. The experiments can be reproduced exactly
this way. Fig. 9 shows a scene, where a small leader vehicle
drives in a straight way very close to an obstacle on the left.
The ego vehicle avoids this obstacle by performing a swerve
maneuver. The desired path as output from the elastic-band
algorithm is depicted in a dashed line for every time step. These
lines can only be seen when deviations between the desired and
the actual path occur. Fig. 10 shows the sequence of the swerve
maneuver in three snapshots. The desired path as output from
the elastic-band algorithm is depicted in a dashed line. Note
that the ego vehicle does not follow exactly the elastic-band
path due to nonholonomic constraints and due to the simple
controller design.
242 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007
Fig. 9. Bird view for an actually driven situation where a pedestrian barely
enters the driving corridor, which necessitates a swerve maneuver. The ego
vehicle and the leader vehicle drive at 6 m/s, and they are about 15 m apart.
Fig. 10. Snapshots for the scene depicted in Fig. 9, showing the swerve
maneuver in three steps.
Fig. 11. Bird view for a situation where a car slightly enters the driving
corridor, which necessitates a swerve maneuver. The ego vehicle and the leader
vehicle drive at 8 m/s, and they are about 18 m apart.
In Fig. 11, the leader vehicle drives straight. Another vehicle
in the right lane recognizes an obstacle ahead and starts to
change its lane. When recognizing the autonomous vehicle, it
Fig. 12. Bird view for a situation where the leader vehicle passes a bicycle
and the ego vehicle must avoid the bicycle not to cut it. The ego vehicle and
the leader vehicle drive at 8 m/s, and they are about 18 m apart. The snapshots
show that the bicycle is avoided successfully.
Fig. 13. Deviation plot for the scene in this figure. The frame number roughly
corresponds to the z position in the global reference frame (constant velocity
assumed). Compared to the image data, the cardboard box location is depicted
in the graph. Assuming the leader-vehicle path runs along the z axis at x = 0,
the deviation curve also shows the path of the ego vehicle.
brakes but overlaps slightly with the other lane. Hence, a swerve
maneuver of the autonomous vehicle will yield results similar
to a maneuver a human would conduct in such a situation.
In Fig. 12, the leader vehicle passes a bicyclist and cuts back
in too early for pure vehicle following. Due to the elastic-band
framework, enough room is left for the bicyclist. Fig. 12 shows
three snapshots of the scene that are 0.8-s apart each. The ego
vehicle leaves sufficient room for the bicyclist when returning
to its lane.
3) Real-World Results Using Stereo Vision: After extensive
tests with simulated obstacles, we performed avoidance maneu-
vers with real image processing data in the research vehicle. A
correct mapping of the whole avoidance maneuver in an inertial
reference frame is difficult to establish due to localization and
measurement uncertainties. In our chosen scenario, a human as
a thin leader vehicle runs on a straight path at about 5 m/s.
A cardboard box is located next to that path. The ego vehicle
has to avoid the box while following the human. Fig. 13 shows
the deviation between leader vehicle and ego vehicle path. The
GEHRIG AND STEIN: COLLISION AVOIDANCE FOR VEHICLE-FOLLOWING SYSTEMS
Fig. 14. Avoidance scene with a human as a leader vehicle passing close to a
cardboard box on the left. Due to the larger width, the ego vehicle performs a
swerve maneuver to avoid the box. The elastic-band path is depicted projected
to the ground. The leader vehicle and the ego vehicle are about 12 m apart.
Fig. 15. Traffic scene with trucks to the left and right of the leader vehicle.
Back light, wet street, and a large distance deteriorate detection.
deviation increases at the cardboard box location due to the
avoidance maneuver. Fig. 14 shows a snapshot of the scene.
The depicted corridor is the corridor modified by the elastic-
band algorithm. In a simple vehicle-following scenario, the car
would have touched the cardboard box.
We performed further tests in real traffic using the elastic-
band framework. The desired avoidance maneuvers were prop-
erly performed in all encountered dynamic situations. We only
found experimental systems with similar performance in the
robotics community with speeds well below 5 m/s (e.g., [15]).
Other lateral control algorithms for low-speed following show
only simulation results (e.g., [43]).
4) Results Under Adverse Visibility Conditions: Feasibil-
ity studies are usually concerned with default situations. The
system presented here is designed to keep a high performance
level under all circumstances.
Fig. 15 shows a scene where the leader car is far away
and the vehicle center determination delivers a slightly wrong
measurement due to adverse visibility conditions. Since the car
is driving between two trucks and the elastic-band framework
delivers a path with sufficient obstacle clearance, the erroneous
center determination is corrected, and the ego vehicle follows
the car accurately (see Fig. 16).
In all realistic scenarios, the elastic-band algorithm found an
equilibrium state within 20 iterations. The average computa-
tion time for the algorithm with interfering obstacles is about
10 ms on a 400-MHz Intel Pentium II PC. Even in cluttered
and complex scenes, the maximum computation time always
stays below 50 ms.
VIII. C ONCLUSION AND F UTURE W ORK
In this paper, a dynamic collision avoidance component for
the standard vehicle-following approach has been introduced.
243
Fig. 16. Bird view for the scene in Fig. 15. The leader-vehicle path becomes
pushed toward a straight path, which is desirable for that scenario.
The elastic-band framework is used to modify the initial path
of the leader vehicle.
Modeling human driving behavior requires a lot of context
knowledge that has to be represented as a rule base in some
way. However, we consciously skipped that step and tried to
model human driving behavior with a physical model in order
to keep things intuitive and to analyze global properties of the
planned path.
The results show that for standard vehicle-following sit-
uations, no modification is necessary. In dynamic situa-
tions, the intuitive obstacle-avoidance behavior is achieved.
Computation-time measurements show that real-time perfor-
mance is achieved without much optimization effort. By con-
sidering the leader vehicle, all obstacles in the scene, and lane
markings, an intelligent fusion of lane-following and vehicle-
following behavior has been achieved.
Note that this approach comes to its limitations when se-
vere deviations from the leader-vehicle path occur. Although
a feasible collision-free path is still obtainable, the required
maneuver of the ego vehicle would lead to uncomfortable
swerve maneuvers, which are not desired. Systems performing
emergency swerve maneuvers clearly need to take vehicle
dynamics into account. For this, the elastic-band approach
also has the potential to model dynamic aspects such as tire
forces.
In the future, we will investigate the use of a discretized
depth map in contrast to the obstacles as basis for the potential.
This would yield a more reactive response that is more closely
related to the sensor data. The number of distance computations
would increase significantly, but at the same time, the distance
computation degenerates to a distance computation between the
point and the line segment.
An interesting future area of research is the choice of the lat-
eral controller to follow the elastic-band path. In the intelligent-
vehicle community, lateral vehicle guidance is often achieved
performing lane following. A lot of robust lateral controllers
have been presented for that task (e.g., [44]). These controllers
could be used to track the elastic-band path by treating that path
as a lane center.
244 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 8, NO. 2, JUNE 2007
R EFERENCES
[1] Y. Iwasaki, “Japan’s policies on research and development of the advanced
cruise-assist system,” in Proc. 6th World Congr. Intell. Transp. Syst.,
Toronto, ON, Canada, 1999.
[2] U. Franke et al., “Autonomous driving goes downtown,” IEEE Intell.
Syst., vol. 13, no. 6, pp. 40–48, Nov./Dec. 1998.
[3] A. Gern, U. Franke, and P. Levi, “Advanced lane recognition—Fusing
vision and radar,” in Proc. Intell. Vehicles Symp., 2000, pp. 55–60.
[4] Y. He, H. Wang, and B. Zhang, “Color-based road detection in
urban traffic scenes,” IEEE Trans. Intell. Transp. Syst., vol. 5, no. 4,
pp. 309–318, Dec. 2004.
[5] B. H. Krogh, “Feedback obstacle avoidance control,” in Proc. 21st Aller-
ton Conf. Commun., Control, Comput., 1983, pp. 325–334.
[6] O. Khatib, “Real-time obstacle avoidance for manipulators and mobile
robots,” in Proc. IEEE Conf. Robot. Autom., 1985, pp. 500–505.
[7] J. Guldner, V. Utkin, and R. Bauer, “A three-layered hierarchical path
control system for mobile robots: Algorithms and experiments,” Robot.
Auton. Syst., vol. 14, no. 2/3, pp. 133–147, May 1995.
[8] R. A. Grupen et al., “Toward a path co-processor for automated vehicle
control,” in Proc. Intell. Vehicles Symp., 1995, pp. 164–169.
[9] J. Decuyper and D. Keymeulen, “A reactive robot navigation system based
on a fluid dynamics metaphor,” in Proc. Conf. Parallel Problem Solving
Nature, 1991, pp. 356–362.
[10] A. Steinhage and G. Schöner, “Self-calibration based on invariant view
recognition: Dynamic approach to navigation,” Robot. Auton. Syst.,
vol. 20, no. 2, pp. 133–156, Jun. 1997.
[11] S. Quinlan and O. Khatib, “Elastic bands: Connecting path planning and
control,” in Proc. IEEE Conf. Robot. Autom., 1993, pp. 802–807.
[12] S. Quinlan and O. Khatib, “Towards real-time execution on motion
tasks,” in Experimental Robotics 2. New York: Springer-Verlag, 1993,
pp. 241–254
[13] M. Khatib et al., “Dynamic path modification for car-like nonholo-
nomic mobile robots,” in Proc. IEEE Conf. Robot. Autom., 1997, vol. 4,
pp. 2920–2925.
[14] O. Brock and O. Khatib, “Executing motion plans for robots with many
degrees of freedom in dynamic environments,” in Proc. IEEE Conf.
Robot. Autom., 1998, pp. 1–6.
[15] O. Brock and O. Khatib, “High-speed navigation using the global dynamic
window approach,” in Proc. IEEE Conf. Robot. Autom., 1999, vol. 1,
pp. 341–346.
[16] J. C. Gerdes and E. J. Rossetter, “A unified approach to driver assistance
systems based on artificial potential fields,” in Proc. ASME Int. Mech.
Eng. Congr. Expo., Nashville, TN, 1999, pp. 622–630.
[17] E. J. Rossetter, J. P. Switkes, and J. C. Gerdes, “A gentle nudge
towards safety: Experimental validation of the potential field driver
assistance system,” Int. J. Automot. Technol., vol. 5, no. 2, pp. 95–108,
2004.
[18] R. White and M. Tomizuka, “Autonomous following lateral control of
heavy vehicles using laser scanning radar,” in Proc. Amer. Control Conf.,
Arlington, VA, Jun. 2001, pp. 2333–2338.
[19] D. M. Reichardt, “Kontinuierliche Verhaltenssteuerung eines autonomen
Fahrzeugs in dynamischer Umgebung,” Ph.D. dissertation, Universität
Kaiserslautern, Kaiserslautern, Germany, 1996.
[20] P. Fiorini and Z. Shiller, “Motion planning in dynamic environments
using velocity obstacles,” Int. J. Robot. Res., vol. 17, no. 7, pp. 760–772,
Jul. 1998.
[21] A. Chakravarthy and D. Ghose, “Obstacle avoidance in a dynamic envi-
ronment: A collision cone approach,” IEEE Trans. Syst., Man, Cybern.,
vol. 28, no. 5, pp. 562–574, Sep. 1998.
[22] V. K. Narendran and J. K. Hedrick, “Autonomous lateral control of
vehicles in an automated highway system,” Veh. Syst. Dyn. J., vol. 23,
pp. 307–324, 1994.
[23] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour
models,” Int. J. Comput. Vis., vol. 1, no. 4, pp. 321–331, Jan. 1988.
[24] S. Quinlan, “Real-time modification of collision-free paths,” Ph.D. disser-
tation, Stanford Univ., Stanford, CA, 1994.
[25] F. Kang and S. Zhong-Ci, Mathematical Theory of Elastic Structures.
Beijing, China: Science Press, 1995.
[26] R. P. Brent, Algorithm for Minimization Without Derivatives, Prentice-
Hall Series in Automatic Computation. Englewood Cliffs, NJ: Prentice-
Hall, 1973.
[27] D. M. Reichardt, “Using automated assistance systems—putting the
driver into focus,” in Proc. Intell. Vehicles Symp., 1998, pp. 413–418.
[28] U. Franke and I. Kutzbach, “Fast stereo object detection for stop and go
traffic,” in Proc. Intell. Vehicles Symp., 1996, pp. 339–344.
[29] T. Kato, Y. Ninomiya, and I. Masaki, “Preceding vehicle recognition
based on learning from image samples,” IEEE Trans. Intell. Transp. Syst.,
vol. 3, no. 4, pp. 252–260, Dec. 2002.
[30] W. van der Mark and D. Gavrila, “Real-time dense stereo for intelligent
vehicles,” IEEE Trans. Intell. Transp. Syst., vol. 7, no. 1, pp. 63–77,
2006.
[31] H. Edelsbrunner, Algorithms in Combinatorial Geometry. New York:
Springer-Verlag, 1987.
[32] J. E. Bobrow, “A direct minimization approach for obtaining the distance
between convex polyhedra,” Int. J. Robot. Res., vol. 8, no. 3, pp. 65–76,
1989.
[33] M. C. Lin and J. F. Canny, “A fast algorithm for incremental distance
calculation,” in Proc. IEEE Conf. Robot. Autom., 1991, pp. 1008–1014.
[34] V. J. Lumelsky, “On fast computation of distance between line segments,”
Inf. Process. Lett., vol. 21, no. 2, pp. 55–61, Aug. 1985.
[35] B. H. Krogh, “A generalized potential field approach to obstacle avoid-
ance control,” in Proc. Conf. Robot. Res.: Next Five Years and Beyond,
Bethlehem, PA, Aug. 1984, pp. 1–15.
[36] B. Rieger, “Entwurf, Simulation und Implementierung von Hindernisver-
meidungsstrategien für ein Autonomes Fahrzeug,” M.S. thesis, Fakultät
für Physik, Universität Tübingen, Tübingen, Germany, 2000.
[37] S. K. Gehrig and F. J. Stein, “A trajectory-based approach for the lateral
control of vehicle following systems,” in Proc. Intell. Vehicles Symp.,
1998, vol. 1, pp. 156–161.
[38] M. F. Land and D. N. Lee, “Where we look when we steer,” Nature,
vol. 369, no. 6483, pp. 742–744, Jun. 1994.
[39] U. Franke et al., “Truck platoon in mixed traffic,” in Proc. Intell. Vehicles
Symp., 1995, pp. 1–6.
[40] M. Hunt et al., “Speed control design for an experimental vehicle using a
generalized gain scheduling approach,” IEEE Trans. Control Syst. Tech-
nol., vol. 8, no. 3, pp. 381–395, May 2000.
[41] J. E. Naranjo, C. Gonzalez, R. Garcia, and T. dePedro, “Acc and stop-and-
go maneuvers with throttle and brake fuzzy control,” IEEE Trans. Intell.
Transp. Syst., vol. 7, no. 2, pp. 213–225, Jun. 2006.
[42] S. K. Gehrig and F. J. Stein, “Dead reckoning and cartography using stereo
vision for an autonomous car,” in Proc. Int. Conf. Intell. Robots Syst.,
Kyongju, Korea, 1999, pp. 1507–1512.
[43] S. Mammar and M. Netto, “Integrated longitudinal and lateral control for
vehicle low speed automation,” in Proc. Int. Conf. Control Appl., 2004,
pp. 350–355.
[44] J. Kosecka et al., “Vision-based lateral control of vehicles,” in Proc. IEEE
Conf. Intell. Transp. Syst., Boston, MA, 1997, pp. 900–905.
Stefan K. Gehrig was born in Germany in 1968.
In 1991, he received the Dipl.-Ing. (B.A.) degree in
computer science from Berufsakademie Stuttgart,
Stuttgart, Germany. Subsequently, he continued to
study physics at the University of Stuttgart, Stuttgart,
Germany, at University of Tuebingen, Tuebingen,
Germany, at San Jose State University, San Jose, CA,
and at University of California, Berkeley. In 1997,
he received the Dipl.-Phys. degree (summa cum
laude) from University of Tuebingen, Tuebingen,
Germany, where he received the Ph.D. degree in
2000, in cooperation with DaimlerChrysler.
His main research interests are intelligent vehicles, computer vision, and
robotics. Currently, he is a Staff Researcher with the DaimlerChrysler Research
and Technology, Sindelfingen, Germany.
Fridtjof J. Stein was born in Germany in 1962. In
1988, he received the Dipl.-Inf. degree in computer
science from University of Karlsruhe, Karlsruhe,
Germany. Subsequently, he continued to study com-
puter science and received the Ph.D. degree from
University of Southern California, Los Angeles,
in 1992.
His main research interests are intelligent vehicles
and computer vision. Currently, he is a Principal
Researcher with the DaimlerChrysler Research and
Technology, Sindelfingen, Germany.