www.ietdl.

org

Published in IET Intelligent Transport Systems

Special Issue-selected papers from ITS ’07
Received on 27th August 2007
Revised on 8th April 2008
doi: 10.1049/iet-its:20070031

ISSN 1751-956X

3D vehicle detection using a laser

scanner and a video camera
S. Wender K. Dietmayer
Institute of Measurement, Control, and Microtechnology, University of Ulm, Albert-Einstein-Allee 41, Ulm D-89081, Germany
E-mail: [email protected]

Abstract: A new approach for vehicle detection performs sensor fusion of a laser scanner and a video sensor.
This combination provides enough information to handle the problem of multiple views of a car. The laser
scanner estimates the distance as well as the contour information of observed objects. The contour
information can be used to identify the discrete sides of rectangular objects in the laser scanner coordinate
system. The transformation of the three-dimensional coordinates of the most visible side to the image
coordinate system allows for a reconstruction of its original view. This transformation also compensates size
differences in the video image, which are caused by different distances to the video sensor. Afterwards, a
pattern recognition algorithm can classify the object’s sides based on contour and shape information. Since
the number of available object hypotheses is enormously reduced by the laser scanner, the system is
applicable in real time. In addition, video-based vehicle detection and additional laser scanner features are
fused in order to create a consistent vehicle environment description.

1 Introduction A more promising approach is the sensor fusion of

distance sensors and video cameras. The distance
Active safety systems such as ESP were developed to sensors measure obstacles. The corresponding
reduce the number of accidents. More complex safety positions in the video image can be calculated, if the
systems as well as driver assistant systems need more sensors are synchronous and calibrated. This approach
detailed information about the vehicle’s environment. enormously reduces the number of possible object
Vehicle detection can improve several applications sizes and positions (Fig. 3). Recent works only
with respect to robustness and low false alarm rates. concentrated on rear views of vehicles [3 –6].

There are several challenges corresponding to the Usually, the orientation of observed vehicles is only
vehicle detection task. If only a single video sensor is restricted on highways. At intersections and in urban
used for object detection, there will be a lot of object areas, vehicles can occur in all possible orientations to
hypotheses, which must be evaluated. Vehicles must be the video sensor. The vehicle’s appearance changes
expected at all possible positions and in all possible sizes with its orientation to the video sensor (Fig. 4).
in the video image, if no additional knowledge is Usually, pattern recognition concentrates on contour
available. This is usually performed with image and texture information. For this reason the detection
pyramids (Fig. 1). The complete image-scanning task becomes more complex. This fact is considered
procedure then needs a lot of processing time. There by Schneiderman and Kanade [7]. They trained a
are approaches, which reduce the available object detector with samples of different viewpoints. Thus, a
hypotheses by a flat world assumption [1, 2]. This complex detector, which can represent all views, was
approach benefits from the correlation of object necessary. Another possibility is to train several
position and size (Fig. 2). Unfortunately the flat world detectors to cover different viewpoints with different
assumption does not hold for all possible traffic scenarios. detectors. Consequently many detectors have to be

& The Institution of Engineering and Technology 2008 IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112/ 105
doi: 10.1049/iet-its:20070031
www.ietdl.org

Figure 3 Sensor fusion of distance sensors and video

Figure 1 If no additional knowledge is available, object cameras; the distance sensor measures the positions of
hypotheses at all possible positions and sizes have to be obstacles
evaluated The positions in the image and the expected object sizes can be
This is usually performed with image pyramids estimated due to sensor synchronisation and calibration
The image is scanned with a detector for a fixed object size
After that the image is scaled and the procedure starts again
camera model [8]. An example of transformed laser
scanner measurements is shown in Fig. 5.
applied to the same image, since each detector can only
cover small changes of the orientation. A transformation of a point in the video image to the
laser scanner coordinate system is also possible. In
Both solutions lead to quite high computational costs. contrast to the reversed transformation, this
Therefore this work extends the idea of hypotheses transformation is not unique. It results in a straight
selection by additionally obtaining the object’s line through the optical centre of the video camera.
orientation from a laser scanner. This orientation
allows for a reconstruction of a perspectively warped
side of a vehicle (e.g. rear end, left side, right side or
front). Thus, only four different patterns have to be 3 View decomposition
detected. For this purpose, a fast pattern classification
algorithm based on contour information and texture This section describes the three-dimensional
can be used. Experiments have shown that it is decomposition of the mixture of object side views.
possible to train one detector for combined The most visible object side is reconstructed.
recognition of left and right sides and one for front
and rear ends, respectively.
3.1 Object model
The three-dimensional shape of cars is approximated by
2 Coordinate transformation a cuboid (Fig. 6, left). This three-dimensional object box
This work fuses information of a multilayer laser scanner is localised for all vehicle candidates by the sensor
with images of a video camera. Both sensors observe the information of the laser scanner. The laser scanner
area in front of the test vehicle. The sensors work estimates contour information. This provides
synchronously and are calibrated. The sensor information about the object distance, orientation and
calibration is performed offline with special calibration a visible corner (Fig. 6, middle). The video sensor
objects [5]. This task only has to be done once, provides a mixture of the visible object sides (Fig. 6,
because the sensors are mounted with a fixed position right). The influence of a car roof on this mixture will
and orientation. Afterwards, the positions and the be quite small. For this reason, the upper side will be
orientations of the sensors relative to the vehicle are ignored by further algorithms. A decomposition and
known. Usually, a sensor provides measurements in its rectification reconstruct the look of the object side
own coordinate system. The calibration enables a with the best orientation to the video sensor. This
transformation of a point from the laser scanner reconstructed view is used to classify the object.
coordinate system to the image plane by applying
rotational and translational matrices and the pinhole In general, real objects do not exactly fit the cuboidal
object model. This results in errors in the appearance of
the reconstructed object sides. In the case of vehicles,
this error is quite small. Thus, the reconstructed side
is quite similar to the real object side. If the
reconstructed object is not a vehicle, it can appear in
any three-dimensional shape. Therefore the
reconstruction error can be much higher than for
cars. Fortunately, it seems not to be likely that the
Figure 2 Flat world assumption; the expected object reconstructed side of such an object is similar to a
size correlates with the position in the video image side view of a car.

106 / IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112 & The Institution of Engineering and Technology 2008
doi: 10.1049/iet-its:20070031
 www.ietdl.org

Figure 4 Cars significantly change the appearance with the orientation to the video sensor
Therefore the detection task becomes very difficult, if pattern recognition based on texture and contour is used

3.2 Estimation of the horizontal length of a car, because the object shape usually only
orientation defines an object orientation in a range of 0– 908.
The laser scanner measurements of an object provide In addition to the horizontal position, the laser scanner
information about the vehicle’s horizontal position and also provides a vertical position of its measurements. As
orientation. A distance-based segmentation algorithm the laser scanner measures in four horizontal layers, the
groups distance measurements, which probably vertical information is quite insufficient to determine
correspond to the same object. The contour of the the exact vertical object expansion. Nevertheless, this
object is now analysed. An algorithm fits two types of information can be used to calculate an object box,
shapes into each segment. The L-shape is typical for which certainly includes the object. The upper
rectangular objects such as cars with two observed boundary of this box is the sum of the maximum object
sides. An I-shape can be caused by a rectangular object height and the smallest z-value of all laser scanner
with one observed side or by objects such as walls or measurements. The lower boundary is the difference
crash barriers. The algorithm selects the shape, which between the largest z-value and the maximum object
fits the best. If this shape does not fit well enough, an height.
O-shape will be selected. This type describes objects
of undefined shape (Fig. 7).
3.4 Coordinate transformation
Details about the shape estimation can be found in The 3D object box side with the most visible orientation
[9]. The assigned shape allows for an extraction of the to the camera is transformed to the image coordinate
object’s orientation in cases of L- and I-shapes. system. The transformed box side describes the
Objects with an O-shape will be ignored by the position and the deformation of the corresponding
following algorithms. object side in the image. For pattern recognition
purposes, the box side is increased by 25% before it is
transformed to the image coordinate system. An
3.3 Object box fitting example for this transformation is shown in Fig. 9.
The estimation of the visible object sides is based on an
object box, which is fitted in the laser scanner 3.5 Rectification of the object sides
measurements (Fig. 8). As the transformed box side describes the position and
deformation of the original object side in the video
The horizontal position is defined precisely by the image, a perspective warping can reconstruct the
assigned shape. The visible object sides are positioned
directly at the visible surface of the box. The width
and the length of the box is defined by the maximum

Figure 6 Box as three dimensional object model (left), the

contour information, measured by the laser scanner
Figure 5 Laser scanner measurements, transformed to (middle) and the two object sides, which mainly influence
the image of the video sensor the appearance of the object in the video data (right)

& The Institution of Engineering and Technology 2008 IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112/ 107
doi: 10.1049/iet-its:20070031
www.ietdl.org

Figure 7 Contour analysis selects the best fitting horizontal

object shape
Figure 9 Visible side of the object box in the video image;
Three types of shapes are distinguished: O, I and L
the system takes the object side with the most
appropriate orientation to the video sensor from the laser
original view of the object side. For this purpose, a scanner data processing
perspective transformation algorithm of the Intel The side is increased by 25%
Open Computer Vision Library [10] was applied to Afterwards, it is transformed to the video image for pattern
the source ROI. It transforms the image information recognition purposes
of the trapezoidal source ROI to a rectangular image
using linear pixel interpolation. perspective warping creates object side views, which
are much more similar to the original view than the
The source ROI has a fixed length in metres in world mixture of deformed views in the video image. The
coordinates. The warping creates a rectangular image of reconstructed views can now be used for a pattern
a fixed size in pixel. This process performs a recognition algorithm, which detects objects based on
compensation of distance-based object size differences. shape and contour information.
For this reason, the later applied pattern recognition
system can benefit from a priori knowledge about the
expected reconstructed object size. Only a small range 4 Video-based object detection
of positions and sizes have to be evaluated by the
detector. The reconstructed object side view is shown 4.1 Pattern recognition system
in Fig. 10. This part applies a pattern recognition system to the
reconstructed views of the objects. The pattern
The warping only correctly reconstructs areas, which recognition system was introduced by Viola and Jones
are located directly on the box side. Consequently, [11] for the task of face detection. Several works
several types of errors occur. All objects and object already applied this system to different automotive
parts, which are not part of the corresponding object applications including traffic sign detection [12],
side, are deformed in an undefined way. Strange pedestrian detection [13], vehicle detection [1, 2] and
deformations are possible as shown in the left part multi-object detection [14, 15]. These works showed
of the warped image. In addition, there can even the impressive capabilities of real-time object detection.
be some errors at the reconstructed object side.
This is because of the fact that car sides are not The system uses a large amount of Haar-like features.
completely vertical as assumed by the object box The feature types can be applied at different image
model. Especially the front of a car often can only positions and in different sizes. Each used feature is a
be reconstructed below the hood. However, the combination of feature type, size and position.

Figure 10 Reconstruction of the most observable object

Figure 8 Horizontal contour information from the laser side
scanner (left) and the three-dimensional object box fitted Size differences caused by different distances to the video sensor
into the laser scanner measurements (right) are compensated

108 / IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112 & The Institution of Engineering and Technology 2008
doi: 10.1049/iet-its:20070031
 www.ietdl.org

Two methods were used for label generation. The

labelling of front and rear views is supported by a video
and laser-scanner-based tracking [3]. The human user
has to label the first appearance of the car. The
algorithm then remembers the labelled pattern and the
distance to the object. The pattern is resized
corresponding to the measured distance in the next
frame. The resized pattern is located in the new frame
by a maximisation of the cross-correlation value. The
Figure 11 Classifier cascade; each cascade stage rejects second method is for side views of cars. Side views are
some of the object candidates labelled by the wheel positions. The tool later performs
The hypotheses, which pass the last stage, are usually used as a perspective reconstruction of the side view and also
detected objects transforms the positions of the wheels to the new image.
Afterwards, a rectangular area of the reconstructed
The system takes advantage of the features’ low and size- image is used as sample. The wheels are located at fixed
independent computational costs. Useful combinations positions in this area and define the area’s size (Fig. 13).
of few features are selected by a training procedure
with AdaBoost. This boosting algorithm creates a The labelling of the wheels is supported by a wheel
binary classifier based on the selected features. The detector, which is applied to each object separately.
complete pattern recognition system is a cascade of This detector applies a Gaussian and a Laplace filter to
several binary classifiers. Each stage of the cascade calculate an edge image of an appropriate ROI, which
reduces the number of vehicle candidates. The output is calculated using the object’s laser scanner
of the final stage is used as the detection result (Fig. 11). measurements. Afterwards a pair of wheels is located
using the Hough transform for ellipses (Fig. 14). The
human user only has to confirm or correct the labels.
4.2 Training procedure
The parameters of the classifier cascade are estimated by
a training procedure with AdaBoost. This training is 4.4 Application
performed offline based on labelled sample data. For The application of the trained cascades on new unlabelled
the purpose of vehicle detection, two classifier data is the object detection. Both cascades are applied to
cascades are trained. The first cascade detects vehicle reconstructed object sides. The pattern recognition
front and rear ends. The other cascade detects left and system is applied to sub-windows of the reconstructed
right sides. Fig. 12 shows some examples of the view. Several positions and sub-window sizes (scales)
training data. have to be processed. The sizes of reconstructed sides
of all vehicles are quite similar since the warping also
compensates for size differences, which are caused by
4.3 Labelling of the sample data different distances. For this reason, the object detection
The detector parameter estimation procedure (training) only needs to be performed at a small amount of scales
is based on a large set of labelled sample data. Additional and positions. The scales compensate for differences of
data are required for system evaluation purposes. The the real vehicle size. The positions mainly compensate
necessary ground truth data are often labelled by for the uncertain vertical object position.
hand. For this reason, the preparation of the labelled
data is quite time and cost consuming. The necessary An object will be classified as a vehicle, if at least one
effort can be reduced by semiautomatic labelling. This of the cascades detects a vehicle. Since a single detector
approach still needs a human user to confirm the is applied several times, positive objects are often
labels on the data, but the label generation is mostly detected multiple times in the reconstructed view.
performed by computer vision algorithms. False alarms usually cause only one or few detections.

Figure 12 Sample data for rear ends (left) and left sides (right)

& The Institution of Engineering and Technology 2008 IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112/ 109
doi: 10.1049/iet-its:20070031
www.ietdl.org

to an object, several positions and scales of the

template in the calculated ROI will be evaluated. All
these object hypotheses are considered by the feature
extraction. The maximum stage smax of the cascade,
which is passed by at least one hypothesis, defines the
main rating of this feature. If the last stage slast is
passed, the feature value will be 10. If only the stage
before the last is passed, the feature value will be 9
and so on. The minimum feature value of an applied
Figure 13 Layout of side view samples; the wheels are cascade is 0. An offset to the feature value is
located at a fixed position in the reconstructed side view calculated by considering multiple detections, which
For this reason, it is sufficient to label the wheels are usually generated by the cascaded object detector
in cases of positive objects. The number of hypotheses
h, which pass the maximum stage, defines an offset
Therefore false alarms can be reduced by requiring a between 0 and 1. Because of experimental results, a
fixed number of detections for each positive output of maximum of 20 passing hypotheses is considered.
the detector. Fig. 15 shows the detections on the Each hypothesis, which passes the maximum stage,
reconstructed view of the example. increases the feature value by 0.05. The extracted
feature value f can be given by
5 Fusion of the video and laser 8
scanner features >
>
1, if occluded or
>
>
< not in the FOV
The object detection with two cascades already f ¼ max (0, 10
slast þ smax
performs excellent vehicle detection. Unfortunately, it >
>
is subject to several restrictions. The video sensor’s >
> þ min (1, 0:05  h)), if detector was
:
field of view is rather small and the detection system applied
cannot classify objects, which are occluded or outside
the ROI. In addition, the detector only performs the The fusion of these new video-based features with the
object detection for one object class. In order to laser scanner and object-tracking-based features
create a consistent vehicle environment representation enables an improved object classification. For this
with multiple classes, a fusion of the cascaded vehicle purpose, a classification framework is applied to each
detectors’ output with the laser scanner features object (Fig. 16). A pattern classifier (i.e. a neural
seems to be necessary. network) evaluates all available features and calculates
a membership value between 0 and 1 corresponding
The fusion is performed at the feature level. The laser to each class. The membership values can be
scanner provides features in terms of floating point restricted by rule-based a priori knowledge (i.e.
values such as size, velocity or detected reflectors maximal object size of each class). A temporal
[16]. This section describes an additional features stabilisation is performed by an additional filter.
extraction based on an applied classifier cascade (i.e. Finally, the selected class is chosen corresponding to
vehicle side view detector), which is necessary to the maximal class membership value. Details about
enable the fusion with the laser scanner. the used classification framework can be found in [16].

One feature is extracted from each applied cascade 6 Results

for each object. The feature value will be 21, if the
cascade cannot be applied to an object because of The system was evaluated with labelled test data. This
occlusions or size restrictions. If the cascade is applied data consist of several sequences with different views

Figure 14 Algorithm calculates an edge image and localises pairs of wheels by means of a Hough transform in an
appropriate ROI of the original image
Human user only confirms or corrects the labelled data

110 / IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112 & The Institution of Engineering and Technology 2008
doi: 10.1049/iet-its:20070031
 www.ietdl.org

Figure 15 Multiple detections: since the detectors are

Figure 17 Results of the video-based object detector for
applied to a small range of positions and sizes, positive
labelled test data
objects are often detected multiple times

shows the detection rate over the false detection rate.

of cars. All sequences together consist of approximately The tradeoff between these two measures was
10 000 images/laser scans. Two measures were used to performed by requiring different numbers of detections
describe the performance; the detection rate is the ratio for a positive detected car and by ignoring some of the
of correctly detected cars to available cars. The false classifier stages at the end of the classifier cascades.
detection rate is the ratio of wrongly detected cars to
all detected cars. The complete system was applied in
real time with the sensor measurement frequency 6.2 Fusion of video-based object
of 12.5 Hz on a standard Intel Pentium 4 3.2 GHz detection and laser scanner features
desktop computer.
This second evaluation analyses the performance of the
complete system, which fuses video object detection
6.1 Video-based vehicle detection and additional laser scanner features. All objects in the
The first evaluation describes the performance of the field of view of the laser scanner are evaluated.
video-based cascaded object detection system. The Fig. 18 shows three operating point curves. The
laser scanner is only used to determine the ROI (3D dotted curve shows the performance of the system,
object box). which only uses laser scanner features. The results of
the feature level fusion system are given for all objects
The evaluation uses only cars, which are (partly) in the laser scanner’s field of view (solid curve) and
visible to the video sensor. All occluded cars, all cars at for all objects, which are in the field of view of both
distances higher than 80 m and all cars, which are sensors (dashed curve).
located outside the field of view, are ignored. Fig. 17

Figure 16 Feature level sensor fusion: features of different

sensors are used in terms of floating point values
A pattern classifier uses all features to calculate a membership
value corresponding to each class for each object Figure 18 Improvement of the fusion: the performance of
Rule-based knowledge can be applied to restrict the classifier the pure laser-scanner-based approach (dotted line), the
output
A temporal filter stabilises the output feature level fusion approach for all objects (solid line)
Afterward the assigned class can be chosen corresponding to the and for all objects in the camera’s field of view (dashed
maximal membership value line) is given

& The Institution of Engineering and Technology 2008 IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112/ 111
doi: 10.1049/iet-its:20070031
www.ietdl.org

Obviously, the pure laser-scanner-based classification [4] MAEHLISCH M., SCHWEIGER R., RITTER W. , DIETMAYER K.:
already performs very well. But there is still a significant ‘Sensorfusion using spatio-temporal aligned video and
improvement of the feature level fusion. Several facts lidar for improved vehicle detection’. Proc. Intelligent
should be considered, while comparing the results. Vehicles Symposium 2006, Tokyo, Japan, 2006
First, the field of view of the video camera is much
smaller than the laser scanner’s field of view. Thus, [5] KAEMPCHEN N.: ‘Feature-level fusion of laser scanner and
the classification can only be improved for some of the video data for advanced driver assistance systems’, PhD
objects. Second, the laser scanner already performs a Thesis, Ulm University, 2007
good classification of moving objects. Detailed analyses
of the test sequences have shown that the achieved [6] TAKIZAWA H., YAMADA K., ITO T.: ‘Vehicles detection using
benefit is primarily on non-moving cars, but the sensor fusion’. Proc. 2004 IEEE Intelligent Vehicles
evaluated scenarios contain a much higher amount of Symposium, Parma, Italy, 2004
moving cars than non-moving cars. However, the
small amount of cases, which are improved by the [7] SCHNEIDERMAN H., KANADE T.: ‘A statistical model for 3D
video camera, is nevertheless of importance. Non- object detection applied to faces and cars’. Proc. IEEE
moving objects, which are poorly classified by the Conf. Computer Vision and Pattern Recognition, 2000
pure laser scanner approach, can especially cause
dangerous situations, if they are located in front of the [8] FAUGERAS O.: ‘Three-dimensional computer vision’ (MIT
test vehicle. Fortunately, these objects are in the field Press, 2001)
of view of the video camera.
[9] WENDER S., FUERSTENBERG K.CH., DIETMAYER K.C.J.: ‘Object
tracking and classification for intersection scenarios
7 Conclusion using a multilayer laser scanner’. Proc. 11th World
A new real-time vehicle detector was introduced. The Congress on Intelligent Transportation Systems, Nagoya,
system fuses laser scanner measurements with images of a Japan, 2004
video sensor. The system can handle the infinite number
of different views of a car, which are generated by [10] Intel Research Labs: ‘Open Source Computer Vision
different orientations to the observer. A three-dimensional Library’, Online available at: http://www.intel.com/
object model is used. A perspective transformation is technology/computing/opencv/, accessed October 2007
applied to one of the object sides. In addition, scaling
effects caused by different distances to the video sensor [11] VIOLA P., JONES M.: ‘Fast and robust classification using
are compensated. This allows for the application of shape- asymmetric AdaBoost and a detector cascade’, in
and contour-based pattern recognition algorithms at a ‘Advances in Neural Information Processing Systems 14’,
small number of sub-window positions and scales. The (MIT Press, 2002)
system performance was evaluated with labelled test data.
[12] LINDNER F., KRESSEL U., KAELBERER S.: ‘Robust recognition of
8 Acknowledgment traffic signals’. Proc. 2004 IEEE Intelligent Vehicles
Symposium, Parma, Italy, 2004
A previous version of this paper was presented at the
ITS ’07 European Congress held in Aalborg, Denmark [13] MAEHLISCH M., OBERLAENDER M., LOEHLEIN O., GAVRILA D., RITTER W.:
in June 2007. ‘A multiple detector approach to low-resolution FIR
pedestrian recognition’. Proc. 2005 IEEE Intelligent Vehicles
9 References Symposium, Las Vegas, USA, 2005

[1] PONZA D., LOPEZ A., LUMBRERAS F., SERRAT J., GRAF T.: ‘3D vehicle [14] KALLENBACH I., SCHWEIGER R., PALM G., LOEHLEIN O.: ‘Multi-class
sensor based on monocular vision’. Proc. IEEE Conf. object detection in vision systems using a hierarchy of
Intelligent Transportation Systems, Vienna, Austria, 2005 cascaded classifiers’. Proc. 2006 IEEE Intelligent Vehicles
Symposium, Tokyo, Japan, 2006
[2] KHAMMARI A., NASHASHIBI F., ABRAMSON Y. , LAURGEAU C.:
‘Vehicle detection combining gradient analysis and [15] WALCHSHAEUSL L., LINDL R., VOGEL K.: ‘Detection of road
AdaBoost classification’. Proc. IEEE 8th Int. Conf. users in fused sensor data streams for collision
Intelligent Transport Systems, Vienna, Austria, 2005 mitigation’. Proc. AMAA: Advanced Microsystems for
Automotive Applications 2006, Berlin, Germany, 2006
[3] KAEMPCHEN N., DIETMAYER K.: ‘Fusion of Laserscanner and
video for advanced driver assistance systems’. Proc. 11th [16] WENDER S., WEISS T., FUERSTENBERG K., DIETMAYER K.C.J. :
World Congress on Intelligent Transportation Systems, ‘Feature level fusion for object classification’, PReVENT
Nagoya, Japan, 2004 ProFusion e-J., 2006, 1, pp. 31– 36

112 / IET Intell. Transp. Syst., 2008, Vol. 2, No. 2, pp. 105 – 112 & The Institution of Engineering and Technology 2008
doi: 10.1049/iet-its:20070031