LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN

HIGHWAYS

Andr

e Gonc¸alves, Andr

e Godinho

Instituto Superior T

ecnico, Institute for Systems and Robotics, Lisbon, Portugal

ao Sequeira

Instituto Superior T

ecnico, Institute for Systems and Robotics, Lisbon, Portugal

Keywords:

Autonomous car driving, Car-like robot, Behavior-based robot control, Occupancy grid.

Abstract:

This paper presents a viability study on autonomous car driving in highways using low cost sensors for envi-

ronment perception.

The solution is based on a simple behaviour-based architecture implementing the standard perception-to-action

scheme. The perception of the surrounding environment is obtained through the construction of an occupancy

grid based on the processing of data from a single video camera and a small number of ultrasound sensors.

A ﬁnite automaton integrates a set of primitive behaviours deﬁned after the typical human highway driving

behaviors.

The system was successfully tested in both simulations and in a laboratory environment using a mobile robot

to emulate the car-like vehicle. The robot was able to navigate in an autonomous and safe manner, performing

trajectories similar to the ones carried out by human drivers. The paper includes results on the perception

obtained in a real highway that support the claim that low cost sensing can be effective in this problem.

1 INTRODUCTION

Over the last decade there has been a steady decrease

in the number of casualties resulting from automobile

accidents in the European Union roads. Still, more

than 40000 people die each year (see Figure 1) and

hence additional efforts in autonomous car driving are

socially relevant.

Figure 1: Eurostat data of sinistrality (source (EUROSTAT,

2006)).

Research on autonomous car driving systems

shows that such systems may reduce drivers stress and

fatigue resulting in a signiﬁcant decrease in the num-

ber of accidents and fatalities.

Two main strategies are possible in the develop-

ment of autonomous driving systems, one acting on

the infrastructures, i.e., the highways, and the other

on vehicles. The latter has been the one which has

gathered higher acceptance throughout the scientiﬁc

community. Developing a system that has the abil-

ity to navigate in the existing road network requires

the capability of perceiving the surrounding environ-

ment. This paper describes an architecture for au-

tonomous driving in highways based on low cost sen-

sors, namely standard video cameras and ultrasound

sensors.

Nowadays, the most advanced autonomous driv-

ing systems, such as the ones which took part in the

Darpa Grand Challenge 2006, (Thrun et al., 2006;

Whittaker, 2005) embody a large number of sen-

sors with powerful sensing abilities. Despite proving

capable of autonomous navigation even in unstruc-

tured environments such as deserts, these systems are

unattractive from a commercial perspective.

A number of autonomous driving systems de-

370

Gonçalves A., Godinho A. and Sequeira J. (2007).

LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN HIGHWAYS.

In Proceedings of the Fourth International Conference on Informatics in Control, Automation and Robotics, pages 370-377

DOI: 10.5220/0001652803700377

 SciTePress

signed speciﬁcally for highway navigation have been

described in the literature. As highways are a more

structured and less dynamic environment, solutions

can be developed with lower production and oper-

ational costs. At an academic level, (Dickmanns,

1999) developed systems that drove autonomously

in the German Autobahn since 1985, culminating

with a 1758Km trip between Munich and Denmark

(95% of distance in autonomous driving). In 1995,

Navlab 5 drove 2849 miles (98% of distance in au-

tonomous driving) across the United States, (Jochen

et al., 1995). In 1996, ARGO, (Broggi et al., 1999)

drove 2000Km (94% of distance in autonomous driv-

ing) in Italy. At a commercial level, only recently

have autonomous driving systems been introduced.

Honda Accord ADAS (at a cost of USD 46.500), for

example, is equipped with a radar and a camera, be-

ing capable of adapting the speed of the car to trafﬁc

conditions and keeping it in the center of the lane.

The system described in this paper, named HANS,

is able to perform several tasks in the car driving do-

main, namely, following the road, keeping the car

in the right lane, maintaining safe distances between

vehicles, performing overtaking maneuvers when re-

quired, and avoiding obstacles. Without loosing gen-

erality, it is assumed that there are no cars driving

faster than the HANS vehicle, meaning that no cars

will appear from behind.

HANS uses a low resolution web camera located

in the centre of the vehicle behind the rear-view mir-

ror and a set of sixteen sonars. The camera’s objective

is to detect the road lanes and vehicles or objects that

drive ahead. The sonar is used to detect distances to

the vehicles/objects in the surroundings. HANS vehi-

cle is an ATRV robot with a unicycle-like kinematic

structure. A straightforward kinematics transforma-

tion is used such that the vehicle is controlled as if

it has a car-like kinematics. Figure 2 shows the ve-

hicle and the location of the sensors considered (the

small spheric device in the front part of the robot is

the webcam; other sensors shown are not used in this

work) The experiments were conducted in a labora-

tory environment (in Figure 2) consisting on a section

of a highway scaled down to reasonable laboratory di-

mensions. The total length is around 18 meters.

A behaviour-based architecture was used to model

the human reactions when driving a vehicle. It is di-

vided in three main blocks, common to a standard

robot control architecture. The ﬁrst, named Percep-

tion, generates the environment description. It is

responsible for the sensor data acquisition and pro-

cessing. An occupancy grid (see for instance (Elfes,

1989)) is used to represent the free space around the

robot. The second block, named Behavior, is respon-

Figure 2: The HANS vehicle (ATRV robot) in the test envi-

ronment; note the obstacle ahead in the “road”.

sible for mapping perception into actuation. A ﬁnite

automaton is used to choose from a set of different

primitive behaviors deﬁned after the a priori knowl-

edge on typical human driving behaviors. The third

block, Actuation, is responsible for sending the com-

mands to the actuators of the robot.

This paper is organized as follows. Sections 2

to 4 describe the perpection, behavioral and actua-

tion blocks. Section 5 presents the experimental re-

sults. Section 6 summarizes the conclusions drawn

from this project and points to future developments.

2 PERCEPTION

This section describes the sensor data acquisition and

processing, together with the data fusion strategy. The

environment representation method and the data fu-

sion scheme were chosen to, in some sense, mimic

those used by humans when driving a car.

2.1 Ultrasound Sensors

The space covered by the set of sonars is discretized

into an occupancy grid, where the obstacles and free

space around the robot are represented. Basic obsta-

cle avoidance is achieved by computing the biggest

convex polygon of free space in the area in front of

the robot and controlling it such that it stays in this

area.

The chosen occupancy grid, shown in Figure 3,

divides each sonar cone into a number of zones, i.e.,

circular sectors, each being deﬁned by its distance to

the centre of the robot. Each raw measurement from a

sonar is quantized to determine the zone of the grid it

belongs to. Each zone, or cell, in the occupancy grid

keeps record of the number of measurements that fell

in it. The grid is updated by increasing the value in

LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN HIGHWAYS

371

the cell where a measurement occured and decreas-

ing the cell where the previous measurement occured.

The zone with the highest number of measurements

(votes in a sense) is considered as being occupied by

an obstacle. This strategy has a ﬁltering effect that

reduces the inﬂuence of sonar reﬂexions.

Figure 3: Occupancy grid for the sonar data.

The “three coins” algorithm, (Graham, 1972), is

used to ﬁnd the convex polygon which best suits the

free area around the robot, using the information pro-

vided by the occupancy grid. This polygon represents

a spatial interpretation of the environment, punishing

the regions closer to the robot and rewarding the far-

ther ones.

Sonars are also used to detect emergency stopping

conditions. Whenever any of the measurements drops

bellow a pre-speciﬁed threshold the vehicle stops.

These thresholds are set differently for each sonar, be-

ing larger for the sonars which covering the front area

of the vehicle.

2.2 Camera Sensor

The video information is continuously being acquired

as low-resolution (320x240 pixels) images. The pro-

cessing of video data detects (i) the side lines that

bound the trafﬁc lanes, (ii) the position and orienta-

tion of the robot relative to these lines, and (iii) the

vehicles driving ahead and determining their lane and

distance to the robot.

The key role of the imaging system is to extract the

side lines that bound the lanes that are used for normal

vehicle motion. Low visibility due to light reﬂections

and occlusions due to the presence of other vehicles

in the road are common problems in the detection of

road lanes. In addition, during the overtaking maneu-

ver, one of the lines may be partially or totally absent

of the image.

Protection rails tend also to apear in images as

lines parallel to road lanes. Assuming that road lines

are painted in the usual white or yellow, colour seg-

mentation or edge detection can be used for detection.

The colour based detection is heavily inﬂuenced by

the lighting conditions. Though an acceptable strat-

egy in a fair range of situations, edge detection is

more robust to lighting variations and hence this is

the technique adopted in this study.

The algorithm developed starts by selecting the

part of the image bellow the horizon line (a constant

value as the camera is ﬁxed on the top of the robot)

converting to gray-scale. Edge detection is performed

on the image using the Canny method, (Canny, 1986),

which detects, suppresses and connects the edges.

Figure 4 shows the result of an edge detection on the

road lane built in a laboratory environment.

Figure 4: Edge detection using Canny’s method.

The line detection consists of horizontal scannings

starting from the middle of the road (using the line

separating the lanes detected in the previous frame) in

both directions until edge values bigger than a thresh-

old are found. This method uses the estimates of the

position of the lines obtained from previous frames.

To decide which of the edge points belong to the

road lines, they are assembled in different groups.

The difference between two consecutive points mea-

sured along the horizontal x axis is measured and if it

is bigger than a threshold a new group is created.

After this clustering step, the line that best ﬁts

each group (slope and y-intersect) is computed by

ﬁnding the direction of largest variance that matches

the direction that corresponds to the smaller eigen-

value of the matrix containing the points of the group.

Groups whose lines are very similar are considered to

be part of the same line and are regrouped together.

The number of points in each group measures the

conﬁdence level for each line. The two lines (one

for each side of the road) with the biggest conﬁdence

level are considered to be the left and right lines.

The ﬁnal step in this selection consists in com-

paring the distance between the two computed lines

with the real value (known a priori). If the error is

above a predeﬁned threshold, the line with a smaller

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

372

conﬁdence level is changed to one extrapolated at an

adequate distance from the one with the bigger conﬁ-

dence level (recall that the width of the lanes is known

a priori). The line separating the road lanes is then

determined by ﬁnding the line equidistant to the de-

tected side lines. Figure 5 shows the detection of the

lines in the laboratory environment using the above

procedure

Figure 5: Detection of the road limiting lines.

Vehicles lying ahead in the road are detected

through edge detection during the above line search-

ing procedure. The underlying assumption is that any

vehicle has signiﬁcant edges in an image.

The road ahead the vehicle is divided in a ﬁxed

number of zones. In the image plane these correspond

to the trapezoidal zones, shown in Figure 5, deﬁned

after the road boundary lines and two horizontal lines

at a predeﬁned distance from the robot. A vehicle

is detected in one of the zones when the mean value

of the edges and the percentage of the points inside

the zone that represent edges lie above a predeﬁned

threshold. This value represents a conﬁdence level

that the imaging system detected a vehicle in one of

the zones.

The two values, mean value and percentage value,

represent a simple measure of the intensity and size of

the obstacle edges. These are important to distinguish

between false obstacles in an image. For instance,

arrows painted on the road and the shadow of a bridge,

commonly found in highways, could be detected as an

obstacle or vehicle when doing the image processing.

However, these objects are on the road surface while

a vehicle has a volume. Therefore, an object is only

considered as a vehicle if it is detected in a group of

consecutive zones.

The robot’s lateral position and orientation is de-

termined from the position and slope of the road lines

in the image.

2.3 Fusing Sonar an Image Data

The occupancy grid deﬁned for the representation of

the data acquired by the camera is also used to rep-

resent the obstacle information extracted from the ul-

trasound sensors data. Obstacles lying over a region

of the occupancy grid contribute to the voting of the

cells therein.

Both the camera and ultrasound sensors systems

have associated conﬁdence levels. For the camera

this value is equal to the vehicle detection conﬁdence

level. The sonar conﬁdence level depends on the time

coherence of the detection, meaning that it rises when

a sonar detects a vehicle in the same zone in consec-

utive iterations. The most voted zones in the left and

right lanes are considered as the ones where a vehicle

is detected.

There are areas around the robot which are not

visible by the camera, but need to be monitored be-

fore or during some of the robot trajectories. For in-

stance, consider the beginning of an overtaking ma-

neuver, when a vehicle is detected in the right lane.

The robot can only begin this movement if another

vehicle is not travelling along its left side. Verifying

this situation is done using the information on the po-

sition of the side lines that bound the road and sonar

measurements. If any sonar measurement is smaller

than the distance to the side line the system considers

that a vehicle is moving on the left lane. This method

is also used to check the position of a vehicle that is

being overtaken, enabling to determine when the re-

turn maneuver can be initiated.

3 BEHAVIORS

Safe navigation in a highway consists in the sequen-

tial activation of a sequence of different behaviors,

each controlling the vehicle in the situations arising

in highway driving. These behaviors are such that the

system mimics those of a human driver in similar con-

ditions.

Under reasonably fair conditions, the decisions

taken by a human driver are clearly identiﬁed and can

be modeled using a ﬁnite state machine. The events

triggering the transition between states are deﬁned af-

ter the data perceived by the sensors. Figure 6 shows

the main components of the decision mechanism used

in this work (note the small number of states). Each

state corresponds to a primitive behavior that can be

commonly identiﬁed in human drivers.

Each of these behaviors indicates the desired lane

of navigation. The ﬁrst behavior, labeled Normal, is

associated with the motion in the right lane, when the

LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN HIGHWAYS

373

Figure 6: Finite automaton.

road is free. When a vehicle is moving in the right

lane at a smaller speed than the HANS vehicle, the

robot changes to one of two behaviors depending on

the presence of another vehicle in the left lane. If the

lane is clear, the behavior labeled Overtake is trig-

gered and the robot initiates the overtaking maneu-

ver. Otherwise the behavior labeled Follow is trig-

gered and the robot reduces its speed to avoid a col-

lision and follows behind him. The behavior labeled

Return is associated with the return maneuver to the

right lane that concludes the overtaking and is pre-

ceded by checking that the right lane is clear. The last

behaviour, labeled Emergency, not shown in Figure 6,

is activated when an emergency or unexpected situa-

tion is detected and implies an emergency stop, as this

is a safety critical system.

4 ACTUATION

The actuation block receives position references from

the currently active behavior. Assuming safe driving

conditions, the trajectories commonly performed by

vehicles moving in highways are fairly simple.

The control law considered computes the angle

of the steering wheels from a lateral position error

(the horizontal axis in the image plane), (Hong et al.,

2001; Tsugawa, 1999; Coulaud et al., 2006). The ra-

tional under the choice of such a control law is that

for overtaking, and under safe driving conditions, a

human driver sets only a horizontal position reference

while smoothly accelerates its vehicle aiming at gen-

erating a soft trajectory which allows him to change

lane without colliding with the other car. This behav-

ior leads the vehicle to the overtaking lane while in-

creasing the linear velocity relative to the vehicle be-

ing overtaken such that, after a certain time, results

in the overtaking maneuver completed. The refer-

ence for the HANS vehicle is thus a point lying a d

look-ahead-distance, in one of the road lanes. Hard

changes in direction caused by changes in the desired

lane are further reduced through the use of a pre-ﬁlter

that smooths the variations of the reference signal.

The control law is then,

ϕ = −Atan

−1

(Ke), (1)

where ϕ is the angle of the directional wheels mea-

sured in the vehicle reference frame e is the lateral

position error, A a constant parameter that is used to

tune the maximum amplitude of ϕ, and K is a constant

parameter used to tune the speed by which the control

changes for a given error. Different values for the pa-

rameters K and A yield different driving behaviors. In

a sense these two parameters allow the tuning of the

macro-behavior of the autonomous driving system.

The primary concern when designing such a sys-

tem must be safety, i.e., the system must operate such

that it does not cause any trafﬁc accident. Under

reasonable assumptions (e.g., any vehicle circulating

does not move on purpose such that is causes an acci-

dent), safety amounts to require system stability. The

overall system is in fact a hybrid system, with the ac-

tive behavior representing the discrete part of the hy-

brid state and the position and velocity of the vehicle

the continuous one. From hybrid systems theory it

is well known that continuous stability of individual

states does not imply stability of the whole system,

i.e., it is not enough to ensure that each behavior is

properly designed to have the global system stable.

Lyapunov analysis can be used to demonstrate that

(1) yields a stable system. Given the HANS vehicle

kinematics (the usual reference frame conventions are

adopted here)

˙x = vsin(θ)cos(ϕ) (2)

˙y = vcos(θ)sin(ϕ)

θ =

sin(ϕ),

where x,y, θ stand for the usual conﬁguration vari-

ables, v is the linear velocity and L the distance

between the vehicle’s rear and front axis, and the

lateral position and orientation errors, respectively,

e = x + d sin(θ), with d the look-ahead distance, and

ε = θ − θ

, after some straightforward manipulation

yields

˙e = V sin(θ+ ϕ)

ε =

sin(ϕ)

(3)

Substituting in the Lyapunov function candidate

V(e,ε) = 1/2



+ ε



yields for

V(e,ε)

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

374

V = e˙e+ ε

ε = evsin(ε+ ϕ) + ε

sin(ϕ) (4)

Given that V is positive deﬁnite and radially un-

bounded, to have the Lyapunov conditions for global

assymptotic stability veriﬁed amounts to verify that

is negative deﬁnite. Table 5 shows the conditions to

be veriﬁed for (4) to be negative.

Case

State variables 1st term 2nd term

1 e > 0, ε > 0 ϕ < −ε ϕ < 0

e > 0, ε < 0 ϕ < −ε ϕ > 0

e ><,ε > 0 ϕ > −ε ϕ < 0

e ><,ε < 0 ϕ > −ε ϕ > 0

(5)

Figure 7 shows the surfaces corresponding to the

third and fourth columns in table 5.

Figure 7: Lyapunov analysis.

Clearly, there are subspaces of the state space for

which the control law (1) does not results in the can-

didate function being a Lyapunov function. However,

the state trajectories when the HANS vehicle lies in

these regions can be easily forced to move towards

the stability regions. For instance, in cases 2 and

3 the vehicle has an orientation no adequate to the

starting/ending of an overtaking maneuver. By sim-

ply controlling the heading of the vehicle before start-

ing/ending the overtaking maneuver, the system state

moves to the subspaces corresponding to situations

1/4, from which (1) yields a stable trajectory.

The analysis above does not demonstrate global

assymptotic stability for the hybrid system that glob-

ally models the HANS system. However, assuming

again reasonable conditions, the results in (Hespanha

and Morse, 1999) can be used to claim that, from

a practical point of view, the system is stable. The

rational behind this claim is that driving in a high-

way does not require frequent switching between be-

haviors, which amounts to say that the dwell time

between discrete state transitions tends to be large

enough to allow each of the behaviors entering in a

stationary phase before the next switching occurs. Us-

ing Figure 7, this means that once the system enters a

stability region it stays there long enough to approach

the equilibrium state (e,ε, ˙e,

ε) = (0,0,0,0). Alter-

natively, global assymptotic stability can be claimed

by using the generalized version of Lyapunov sec-

ond method (see for instance (Smirnov, 2002)) which

only requires that the last behavior triggered in any

sequence of maneuvers is stable.

5 RESULTS

The system was tested in two different conditions.

The imaging module was tested in real conditions,

with the camera mounted in front, behind the car’s

windshield. Figure 8 shows the results.

Figure 8: Testing the imaging module in real conditions.

The testing of the system including the motion be-

haviors was conducted in a laboratory environment

using the ATRV vehicle shown in Figure 2. This

robot was controlled through the kinematics transfor-

mation, u = vcos(ϕ), ω =

sin(ϕ).

In the ﬁrst test the robot is moving in the right lane

and, after detecting a car travelling in the same lane

at a lower speed, initiates the overtaking maneuver.

Figure 9 shows the resulting trajectory.

While travelling in the right lane, the Normal be-

havior is active. Once the car in front is detected

the robot switches to the Overtake behavior and initi-

ates the overtaking maneuver moving to the left lane.

Since the velocity of the robot is greater than that of

the car being overtaken, after a while the Perception

block detects that the car is behind the robot and de-

cision system switches to Return. When the centre of

the right lane is reached, the active behavior returns

to Normal completing one cycle of the decision ﬁnite

state machine.

LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN HIGHWAYS

375

Figure 9: Overtaking a single vehicle.

Figure 10: Steering wheels angle during overtaking.

Figure 10 shows the angle of the steering wheels.

It reaches a maximum around 20

◦

slightly after

switching to Overtake (t = 40), making the robot turn-

ing left, towards the left lane. After reaching the max-

imum direction angle, the lateral error and directional

angle starts decreasing.

Approximately at t = 60 the steering wheels an-

gle is null, meaning that the position of the lookahead

point is at center of the left lane, i.e., the robot is

aligned with the axis of left lane. When ϕ drops to

negative values the robot turns right aiming at reduc-

ing the orientation error and keeping the alignment

with the axis of the left lane. The lane change is com-

pleted at approximately t = 120, when the steering

angle stabilizes at 0, meaning the robot is moving in

the centre of the left lane.

At t = 125 the system switches to the Return be-

havior. The trajectory towards the right lane is similar

(tough inverse) to the one carried out previously from

the right to left lane.

In both lane changes the steering wheel angle is

noisier in the zones where the maximum values are

reached (t = 40 and t = 130). At these moments the

orientation of the robot is high and therefore one of

the road lines (the one farther to the robot) has to be

extrapolated because it is not visible in the camera im-

age. This results in some oscillation in the line’s po-

sition resulting in a noisier position error and steering

wheel angle.

The second experiment aims at verifying the sys-

tem’s response when it is necessary to interrupt an on-

going maneuver to overtake a second vehicle. Figure

11 shows the resulting trajectory.

Figure 11: Double overtaking.

While circulating in the right lane, the robot de-

tects a car in front and initiates the overtaking ma-

neuver following the routine already described in the

previous experiment. When switching to Return the

robot does not detect a second car which is circulating

ahead in the road, in the right lane, at a lower speed.

This situation occurs if the car is outside the HANS’

detection area but inside the area it needs to complete

the return maneuver. If the car was inside the detec-

tion area, HANS would continue along the left lane

until passing by the second car and then return to the

right lane. In this situation, however, the car is only

detected when the robot is approximately in the mid-

dle of the return phase (and of the road). Immediately,

the decision system changes the behavior to Overtake,

the robot initiates a new overtaking maneuver. After

passing by the car, the robot returns to the right lane

to conclude the double overtaking maneuver.

Figure 12 shows the evolution of the steering

wheels angle. The switching from Return to Over-

take) is visible around instant 125 through the drastic

change in the control signal from -18 to 15 degrees.

Still, the system performs a smooth trajectory.

6 CONCLUSION

Despite the low resolution of the chosen sensors, the

perception block was able to generate a representation

of the surrounding environment with which decisions

could be made towards a safe autonomous navigation.

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

376

Figure 12: Steering wheels angle during the double over-

taking.

Obviously, the use of additional sensors, such as more

cameras and laser range ﬁnders, would give the sys-

tem enhanced capabilities. Still, the experiments pre-

sented clearly demonstrate the viability of using low

cost sensors for autonomous highway driving of auto-

mobile vehicles.

The sonar data processing uses crude ﬁltering

techniques, namely the occupancy grid and the vot-

ing system, to ﬁlter out data outliers such as those

arising due to reﬂection of sonar beams. Similarly

with image processing, supported in basic techniques

that nonetheless where shown effective, with the line

search based on the edge detection showing robust-

ness to lighting variations.

The small number of behaviours considered was

shown enough for the most common highway driv-

ing situations and exhibit a performance that seems to

be comparable to human. Nevertheless, the decision

system can cope easily with additional behaviors that

might be found necessary in future developments. For

instance, different sets of A,K parameters in (1) yield

different behaviors for the system.

Future work includes the use of additional low

cost cameras and the testing of alternative control

laws, for instance including information on the veloc-

ity of the vehicles driving ahead.

ACKNOWLEDGEMENTS

This work was supported by Fundac¸

ao para a

encia e a Tecnologia (ISR/IST plurianual fund-

ing) through the POS

Conhecimento Program that in-

cludes FEDER funds.

REFERENCES

Broggi, A., Bertozzi, A. Fascioli, A., and Guarino, C.

(1999). The Argo autonomous vehicles vision and

control systems. Int. Journal on Intelligent Control

Systems (IJICS), 3(4).

Canny, J. (1986). A computational approach to edge de-

tection. IEEE Trans. Pattern Analysis and Machine

Intelligence, 8(6):679–698.

Coulaud, J., Campion, G., Bastin, G., and De Wan, M.

(2006). Stability analysis of a vision-based control

design for an autonomous mobile robot. IEEE Trans.

on Robotics and Automation, 22(5):1062–1069. Uni-

versit Catholique de Louvain, Louvain.

Dickmanns, E. (1999). Computer vision and highway au-

tomation. Journal of Vehicle System Dynamics, 31(5-

6):325–343.

Elfes, A. (1989). Using occupancy grids for mobile robot

perception and navigation. Computer, 22(6):46–57.

EUROSTAT (Accessed 2006).

http://epp.eurostat.ec.europa.eu.

Graham, R. (1972). An efﬁcient algorithm for determin-

ing the convex hull of a ﬁnite planar set. Information

Processing Letters, 1:132–133.

Hespanha, J. and Morse, A. (1999). Stability of switched

systems with average dwell-time. In Procs of 38th

IEEE Conference on Decision and Control, pages

2655–2660.

Hong, S., Choi, J., Jeong, Y., and Jeong, K. (2001). Lateral

control of autonomous vehicle by yaw rate feedback.

In Procs. of the ISIE 2001, IEEE Int. Symp. on Indus-

trial Electronics, volume 3, pages 1472–1476. Pusan,

South Korea, June 12-16.

Jochen, T., Pomerleau, D., B., K., and J., A. (1995). PANS -

a portable navigation platform. In Procs. of the IEEE

Symposium on Intelligent Vehicles. Detroit, Michigan,

USA, September 25-26.

Smirnov, G. (2002). Introduction to the Theory of Differ-

ential Inclusions, volume 41 of Graduate Studies in

Mathematics. American Mathematical Society.

Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D.,

Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny,

M., Hoffmann, G., Lau, K., Oakley, C., Palatucci, M.,

Pratt, V., Stang, P., Strohband, S., Dupont, C., Jen-

drossek, L., Koelen, C., Markey, C., Rummel, C., van

Niekerk, J., Jensen, E., Alessandrini, P., Bradski, G.,

Davies, B., Ettinger, S., Kaehler, A., Neﬁan, A., and

Mahoney, P. (2006). Stanley: The robot that won the

DARPA Grand Challenge. Journal of Field Robotics,

23(9):661–692.

Tsugawa, S. (1999). An overview on control algorithms for

automated highway systems. In Procs. of the 1999

IEEE/IEEJ/JSAI Int. Conf. on Intelligent Transporta-

tion Systems, pages 234–239. Tokyo, Japan, October

5-8.

Whittaker, W. (2005). Red Team DARPA Grand Challenge

2005 technical paper.

LOW COST SENSING FOR AUTONOMOUS CAR DRIVING IN HIGHWAYS

377