HEIGHT PROFILES
call the vector p a pattern of a vehicle type and denote The choice of the image filtering algorithms is based on
the full set of vehicle patterns as U , = ( P I . . . p K } . the statistical properties of the measurement errors E .
These errors depend on a large number of factors, such
as antenna beam width, transmitted signal power and
It is further assumed, that A4 classes of vehicles, bandwidth, receiver noise level, radar cross-section of
Cl,...,CM, exist and that every class C, consists of the vehicle surface, etc. The results of field trials with a
K,>O vehicle type patterns 24 GHz radar sensor (detailed below) show that the
measurement errors E are characterized by a random
{P,,, ... P , + K , )E c, , rn = 1...M . (2) process that is neither Gaussian nor stationary. The
greater part of the large errors occurs when the
amplitude of the reflected signal is low. This kind of
so that Zf=,(K,)=K errors (called anomalous errors) leads to significant
outliers in height profile estimates and strongly
Our proposed vehicle classification algorithm consists influences the estimation accuracy. To suppress this
of the following steps (Nearest Neighbour approach): kind of errors, we performed median filtering of the
height estimates in the radar image (Pratt (9)).
1) Obtain the image a of a current vehicle from the
sensor. Next, the estimate of the vehicle height profile
; = ~ ( U ) E R L is obtained by reduction of the
2) Extract of a feature vector F by estimation of
dimension of a E R N . It is assumed, that the number of
vehicle the parameters height measurements N is much greater than the
dimension of vehicle type profiles L , i.e. that N >> L .
Then the height profile estimate can be obtained by
dividing the filtered image U into L (possibly
3) Calculate the distances dk = d(F,pk),p,E U,, overlapping) segments consisting of W 2 [ N / L ]
samples each. Next, for every ith segment the mean
k=I...K,where the metric &,pi) specifies the value is calculated, which is used as the ith estimate of
distance between two points with coordinatesp; and pi in the height profile:
parameter space RP .
d,;, = mind, (4) where k, is the left border of the ith segment in terms of
k
the image sample number. Note that after suppression of
and the corresponding pattern p, minimizing the anomalous errors, the distribution of the remaining
errors is close to Gaussian. Thcrcfore, the averaging ( 6 )
distance (4). The number E of class C, to which the leads to increased accuracy of the vehicle height profile
pattern pn belongs is accepted as the vehicle class estimate.
estimate.
Finally, the parameters kl...&}=
f ( u ) are estimated.
Radar Image Pre-Processing
Among the additional parameters that can be extracted
from the image i , the most important is the height h of
For a given vehicle, the image a can be represented as a
the vehicle. For the majority of vehicles, the height is
vectora =[al...aNrE R N with components measured as the distance from the road surface to the
roof of the vehicle and an estimate is obtained by
a, = h , + & , , n = l ...N , (5) choosing as the vehicle height the maximum height
obtained after median filtering of the vehicle height
profile. Furthermore, some major vehicle types, such as
where h, is the real height of the vehicle, F , is the passenger cars, minivans, vans and lorries, typically
measurement error, and N is the number of samples have almost non-overlapping vehicle heights, which
obtained for the vehicle. Due to the presence of errors simplifies the classification.
E, and variations in the number of samples of vehicle
height in (9,the image a cannot immediately he used The length of the vehicle is another important
for classification. The pre-processing step consists of parameter. It estimated by combining vehicle velocity
error filtering and subsequent extraction of the feature measurements from a Doppler radar sensor with the
vector (3). vehicle height profile.
Metric for Assignment into Classes profile estimation and by extending the template library.
High probability of correct classification is reached
Step 3) of the classification algorithm requires the when the conditions
choice of an appropriate metric & ; , p i ) . The metric
can be chosen quite freely, but the basic requirement 'is
that it has to ensure maximum distances between
patterns belonging to different classes. We use as the
metric the following weighted vector norm
pcc
Fig.1. Block-diagram of the microwave radar sensor Therefore, vehicle velocity estimates, if acceptable, can
he used in the classification algorithms. However, high
probability of correct classification is reached evcn
without the information about length. This leads us to
FIELD TRIAL RESULTS believe that a high probability of correct classification
can be achieved also in stop-go traffic, when no velocity
In order to test the performance of the proposed information is available.
approach to vehicle classification, a test-bed system has
been developed. This test-bed system consists of the The main cause of vehicle misclassification is the low
microwave radar sensor described above, as well as a performance of height profile estimation for vehicles
PC on which the signal processing algorithms have been that do not pass directly below the sensor. In this
implemented. Two examples of the height profile situation, the presence of a vehicle on the road was
estimates obtained with this test-bed system are shown detected, but the vehicle height profile estimate was
in Fig. 2 and 3. The figures show typical matching disturbed by the presence of a strong ground echo. This
photographic pictures and vehicle profiles obtained for a problem can he eliminated by installing additional the
car with a trailer and a lorry, respectively. From these radar sensors overlooking the demarcations between
figures, it is clear that the measurement errors are much lanes.
smaller than the typical dimensions of the vehicles.
When measuring the height profile sections
corresponding to the flat surfaces of the vehicles (e.g. CONCLUSION
vehicle's roof), the estimation accuracy is very high. As
a result, a high probability of correct vehicle A new vehicle classification system has been
classification is achieved. considered. The system is based on the use of vehicle
height profiles as the basic characteristic of vehicles.
In order to estimate the empirical probability of correct The main advantages of the proposed classification
classification of vehicles, statistical experiments have system are high classification accuracy, possibility to
been carried out. In these experiments, the vehicles were classify a theoretically unlimited number of vehicle
separated into the following classes: motorcycle, car, types, high performance under most weather and
car with trailer, van, van with trailer, bus, lorry and lorry illumination conditions, as well as low weight, size and
with trailer. The radar sensor was mounted on a bridge cost. As the classification process makes moderate
(mounting height 5.4 m) overlooking a motorway with demands on computational capacity, real-time operation
free flow traffic: Classification of vehicles was then can be achieved with low-cost hardware. The system
carried out with (w'=l) and without ( ~ " 0 ) making use has also proven to he capable of performing
the information about a given vehicle's length (from classification during stop-and-go situations, thus further
vehicle speed measurements). The values of the other enhancing its value for tolling applications. Further
weights were w*=l and d=0,02. The number of development of the system is expected to be performed
vehicles to be classified NY as well as empirical in the areas of optimising the antenna beam pattern,
probability of correct classification PCc are presented in increasing the size and accuracy of the template library,
Tab. 1. and in improvement of the signal processing algorithms.
The field trial results show, that using the information The proposed classification system is intended primarily
ahout vehicle length increases the performance of the for toll-collection applications, hut suitable also for
classification algorithm. Major improvements are found other vehicle classification applications.
in the classification of van and lorries that are
characterized by similar heights but differ in length.
413
Height Profile
4.5, I I
Height Profile
. , . , , . ,
3.51
4t 1
3
F 25
1=:2
5 15
1
05
0
0.51 ’ I 4.51 ” ” ’ ” ‘ 1
1w MO 300 4c? 5w 600 7w 8w 100 MO 300 m 5w WO 7w WO
I i
Fig.2, (b) Fig.3, (b)
Fig. 2,3. Vehicle pichues (a) and corresponding vehicle system”, Proceedings of the IEEE Svmposium on
height profiles (b). Profiles are shown as obtained from Intelligent Transportation Svstems, 460 4 6 5 .
the sensor (before pre-processing).
5 . C h u m i Z , Siyal M, 2000, “A new segmentation
technique for classification of moving vehicles”,
Vehicular Technology Conference Proceedings,
REFERENCES 2000-Spring T o h o , 1 , 3 2 3 -326.
1. Taylor M, Young W, 1988, “Traffic analysis: new 6. Burden M, Bell M, 1997, “Vehicle classification
technology and new solutions”, Hargreen, Melbourne, using stereo vision”, Sixth International Conference on
Australia. Image Processing and Its Applications, 2,881 -885.
2. Gajda J, Sroka R, Stencel M, Wajda A, Zeglen T, 7. Roe H, Hobson G, 1992, “Improved discrimination
2001, “A vehicle classification based on inductive loop of microwave vehicle profiles”, Microwave SvmDosium
detectors”, Proceedings of the 18th IEEE Conference IEEE MTT-S International, 717-720.
IMTC 2001,1,460 4 6 4 .
8. Walton E, Theron I,,Gunawan S, Cai L, 1997,
3. Shiquan P, Harlow C, 1996, “A system for vehicle “Moving vehicle range profiles measured using a noise
classification from range imagery”, Proceedings of the radar”, Antennas and Propaeation Societv International
Twentv-Eighth Southeastern Symposium on Svstem Svmposium, IEEE Digest, 3,2597 -2600.
m, 327-331.
9. Pratt W, 1978, “Digital Image Processing”, John
4. Abdelbaki H, Hussain K, Gelenbe E, 2001, “A laser Wiley & Sons.
intensity image based automatic vehicle classification