The new approach for the detection of vehicles [601702]

The new approach for the detection of vehicles

Amine MAZOUZI 1, Mohamed Faouzi Bel Bachir 2
1 National Institute of telecommunications and information technologies and com munication – Oran POB
1518, El Mnouar , Oran, Algeria
2 USTO -MB, POB 1505, El Mnou ar, Oran, Algeria

Abstract –In the world , road accidents comes ahead as a cause of mortality, With these
massacre , it is out of question of just a simple ordinary road safety, but an automatic and
accurate road safety is widely desired . Our work is jus t part of this problematic and offers first
time and in terms of computer vision, a combination of different image processing approaches
and artificial intelligence.
Our method is designed to exactly detect vehicles in particular locations , it has been m ade based
on a combination of a learning process and a c orrelation measurement approach. The first one is
constituted by a secession of phases which , feature extraction types Gaussian receptive fields ,
reducing the size of the vectors that were used fo r the training phase using the technical subspace,
and learning procedure using an artificial neural network , the second one comes to refine the
results found, it baed for a correlation measurement approach.
Test our detection method was performed on the UIUC database and on video sequences having a
high detection rate.
Keywords : detection of vehicles, characteristics of gaussian susceptible field, sub-space, process
of learning, measure of correlation.

I. Introduction
Most of the road accidents are made between two cars;
it is what explains the significant number of the works,
which have for object the detection of vehicles.
Several works were made in this context; we are
particularly interested in us in Hilario's works and al. [1]
Which proposed a model geometric of vehicle. A
function of energy elaborated according to a symmetry
and a shade of vehicle is defined. Finally, a genetic
algorithm allowed to reduce the space of the parameters.
The works made by Hilario and al. [2] Improve the
function of ene rgy by adding him the form. In other
words the new function of energy takes into account a
symmetry and a shade but also a form of the vehicle. The
detection of the horizontal and vertical outline was made
by means of the filter of sobel. A genetic algor ithm is
used for a selection of parameters allowing a strong
detection.
We also were interested in the works of [3]. These, to
differentiate the static parts of the dynamic parts, they
used the notions of vast op tical flow and the clean space.
The phase of detection of vehicles was made by means of
a detection of the quantities in movement by applying
thresholds. Labels mark the objects which are in
movement. A procedure of ranging for the putting in
correspondence between objects in movement was
elaborate d. The authors of [4], proposed a method for the detection
of vehicles for a purpose of surveillance. A
spatiotemporal transformation, by basing itself on the
application of the wavelet, followed by a thresholding
adaptive to extract the zone in movement was made. To
locate them, a labeling is made.
The work [5] in for purpose the detection and the follow –
up of a vehicle. Chih -Ming and al. presented a method
using the algorithm of learning SVM (Support Vector
Machine) and the filter particulaire PF (Parti cle filtering).
The phase of detection was made by basing itself on the
process of learning SVM, the used type of the
characteristics was the outline of vehicles.
Other method [6] was presented by Luo -Wei and al. They
conceived a new algorithm of detection of vehicles, by
combining between the stage of transformation of the
color space and the classification of pixels by using the
classifier of Bayes. A phase of extraction of the
parameters of type (outline, coefficients wavelet and
corners) and then a pha se of check by means of classifiers
in waterfall are made. A rate of detection of the vehicles
of 94.9 % is obtained.
The article [7] of Hong and al. proposes a method for the
detection of vehicles. An approach was said BGF,
"Boosted Gabor Features" allow s the extraction of
the parameters used as initial parameters to the
system of learning SVM.
In [8], The authors proceed in two successive stages, the
hypothesis of generation HG and the hypothesis of check
HV. In the first stage, HG is made was by applyin g the

A MAZOUZI, M F BEL BACHIR
detector of outlines (horizontal and vertical) and this for
the purpose to select the zone presenting vehicles. The
second stage was made by taking into account the
following characteristics (the main component, the
coefficients of the wavelet and t he transform of Gabor).
A phase of learning by combining between a network of
type MLP's artificial neuron and the algorithm SVM was
made.
The researcher [9] displayed a method based in tow
approaches; the first is detection the motion by using a
diamond m ethod and the second with using a mathematic
morphology. In [10] the authors ha ve present a method
which bases of monocular (one camera) technique and
binocular (two cameras) technique. With the help of
moving object detection stage, the region of interes t
(ROI) in the acquired images is identified. The ROI is
then examined by rule -based algorithms that compute the
distance and the speed between the cars and there system.
The method that we propose in this article is a
hybridization between a process of le arning and a
measure of correlation. This approach is made by using
the characteristics of type gaussian susceptible field, the
technique said sub space and the combination enters the
network of artificial neuron (MLP) and the correlation.
Our objective i s to increase the performances in detection
in contexts and different situations, such as the presence
of vehicles in urban, rural areas, parking lots, etc.
II. Presentation of our method
Our method of detection groups a succession of phases
which we detail in what follows:
II.1. Phase of extraction of the Characteristics

In this phase [11], the procedure of extraction was
made by using three cores presented on the following
figure:

Fig. 1. The cores of Gaussian susceptible fields .

Because of the v ariability of the size of the vehicles
which appear in the images, we applied this phase to an
intrinsic space and for two levels of dimension which are
[20 50], [40 100].
The figure 2 presents some results of application of the
procedure of extraction of the characteristics of type
gaussian susceptible field on both classes vehicles and no
vehicles and for [40 100].

(a)

(b)
Fig. 2. Characteristics of the type Gaussian susceptible field: (a)
positive example, (b) negative example.

II.2. Phase of sub space

Having applied the various pits to make the extraction of
the characteristics of the gaussian susceptible field, we
grouped these characteristics in two types of vectors:
vectors representing the class of vehicles possessing
examples or positive or in negative. This operation was
used se parately for every level. As an example, on an
intrinsic level of size [40 100], we can build a vector of
size 3*40*100, knowing that 3 represent the number of
the used pits.
We notice well that the size of the vector is very big what
pulled a too big comp lexity in the phase of learning.
To remedy this problem, we added for this phase a
technique named by money space out to reduce the
dimension of these vectors.
First of all, we build a matrix S which groups 500
positive vectors representing the class of v ehicles and
500 negative vectors representing the class of no
vehicles. Then, we calculate the matrix of covariance by
using the following relation:
First of all, we build a matrix S which groups 500
positive vectors representing the class of vehicles and
500 negative vectors representing the class of no
vehicles. Then, we calculate the matrix of covariance by
using the following relation:

A MAZOUZI, M F BEL BACHIR

Q=S*ST (1)

After the decomposition SVD of the matrix of
covariance, we fin d three types of matrices U, V, and U T
The reduction of the size of vectors was made by using
the first clean values of the matrix V.
Tests and tries on various sizes [5, 10, 100, and 500 ]
show that 100 is a good parameter. This technique allows
a consid erable reduction of the size of the vectors of the
positive and negative examples of learning.

II.3. Phase of the learning

This phase was made by applying a network of type
MLP's neuron which gave very good result. For that
purpose, we applied a network of art ificial neuron MLP
to a single hidden layer, besides the choice of number of
neurons of the hidden layer was made by adopting the
following law:
Nc = |Ne – Ns| (2)

Nc is the number of neurons of the hidden layer.
Ne is the numb er of neurons of the layer of inputs.
Ns is the number of neurons of the layer of output.
In our case, we have 98 neurons of the hidden layer. The
function of used activation is of sigmoid type. Weighty
matrices and vectors on the bias were fixed randomly,
The number of iterations is equal to 100, and the error of
gradient was fixed to 0.01.
After the initialization of the network of artificial neuron
MLP, comes the stage of learning to obtain by the end in
benchmark models or dictionaries.

II.4. Phase of measu re of correlation

The technique of measure of correlation is used to
measure the resemblance between two of the images. She
gives a result in absolute value which is between 0 and 1.

Corr2 (A, B) =1/NM ∑ A(x,y) B(x+u,y+v) (3)

We used this phase to refine the detection and help the
classifier MLP to make a good decision.
In this phase, we calculated first of all two representing
average images chaqu' one an used class IMvehicules and
IMnonvehicules by using them afterward to compare both
parameters Corr2 (IMvehicules, A) and corr2
(IMnonvehicules, A), In being the one embellishes with
images of test.

(a) (b)
Fig. 3. The two representatives : (a) positive representative , (b)
negative representative .

III. TEST AND PERFORMANCE APPRAISA L
To estimate our method of detection of vehicles, we
calculated the rate of gratitude by the following
formula:

RD = (TD / FD+T D) 100 (4)

RD: rate of detection.
TD: Thru detection.
FD: false detection.

The whole of the database UIUC is established of 167
images of size of 240*200.
The phase of test was made by combining between the
results of the classifier MLP and the comparison between
corr2 (IMvehicules, IMtest) and corr2 (test
IMnonvehicules, IM).

TABLE I
The representation of the various results for our hybrid method
classifier MLP and measure of correlation.

TABLE I I
The representation of the various results for the method of class ifier
MLP.

Size of the
examples
TD
FD
RD
Examples of
size [40 100]
625
22
96.6
Examples of
size [20 50]
550
19
96.66
Examples of
size [10 25]
600
29
95.39
Total 1775 70 96.20

A MAZOUZI, M F BEL BACHIR

The refinement phase which is based on a measu re of
resemblance, "correlation", both the choice of the CRG
characteristics, makes it possible to improve considerably
our results.
If we compare our approach to [ 14] we can notice that
our detection rate is improved.

IV. The results of our method detection
The following figure that present cars of different sizes
located in different places, they are taken from the test
base of UTI , gives an idea how the detection is made
from the method, which we propose. We observe that the
hybrid method makes well result s.

(a)

(b)
Fig. 4. Presentation of the results, (a) application of classifier
MLP, (b) method crosses MLP and measures of correlation

V. Conclusion
In this article, we proposed an effective algorithm for
the detection of vehicles. For that purpose, a combination
between the clas sifier MLP and the measure of
correlation for the phase of classification gave convincing
results. A comparative study with the other existing
methods in the scientific literature showed the legitimacy
of our approach.
in future work , we want to apply thi s method in a
context of presence of adverse weather conditions such as
rain, snow, ..

Size of the
examples
TD
FD
RD
Examples of
size [40 100]
611
36
94.43
Examples of
size [20 50]
528
41
92.79
Examples of
size [10 25]
571
58
90.78
Total 1710 135 92.68

A MAZOUZI, M F BEL BACHIR
References
[1] C. Hilario, J.M. Collado, J.M. Armingol and A. Ofla Escalera,
"Pyramidal Embellish with images Analysis for Vehicle Detection,
IEEE transaction 0 -7803 -8961-1/05, on 2005.
[2] C. Hilario, J.M.Collado, J.M.Armingol, and Has. Ofla Escalera,
"Driver assistance system based one to compute vision for vehicle
detection", university Carlos 3 of Madrid Spain, on 2004.
[3] Junxian Wang, George Bebis and Ronald Miller, “Overtaking
Vehicle Detection Using Dynamic and Quasi -Static Background
Modeling”, Proceedings of the 2005 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition (CVPR’05),
2005.
[4] Yuan -Kai Wang and Shao -Hua Chen, “A robus t vehicle detection
approch”, IEEE transaction 0 -7803 -9385 -6/05, 2005.
[5] Chih -Ming Fu, Chung -Lin Huang, and Yi -Sheng Chen, “Vision –
Based Preceding Vehicle Detection and Tracking”, the 18th
International Conference on Pattern Recognition (ICPR'06), 2006 .
[6] Luo -Wei Tsai, Jun -Wei Hsieh, and Kuo – Chin Fan, “Vehicle
Detection Using Normalized Color and Edge Map“, IEEE transaction
on image, vol… 16, NO. 3, 2007.
[7] Hong Cheng, Nanning Zheng, and Chong Sun,” Boosted Gabor
Features Applied to Vehicle De tection”, Proceedings of the 18th
International Conference on Pattern Recognition (ICPR'06), 2006.
[8] Zehang Sun, George Bebis, and Ronald Miller, "Monocular
Precrash Vehicle Detection: Features and Classifiers", IEEE transaction
we embellish with image s, flight. 15, NO. 7, 2006.
[9] Djoudi Kerfa, M. F. Belbachir,’ An Efficient Real Time Moving
Object Detection Scheme Using Diamond Search Algorithm and
Mathematical Morphology ’ (2014) . International Review on
Computers and Software (IRECOS), vol 9 ,05
[10] Alshaqaqi, B., Boumehed, M., Ouamri, A., Keche, M.,
Implementation of distance and speed measurement algorithms for the
development of an automatic traffic regulation system, (2012)
International Review on Computers and Software (IRECOS), 7 (6), pp.
2804-2809.
[11] John Ruiz -Hernandez, James L. Crowley, and Nicolas Gourier,
"Detection of faces embarked for the mobile telephony", INRIA Rhône –
Alpes, France, on 2007.
[12] Yoshida Tatsuya, Kagesawa Masataka, and Ikeuchi Katsushi,
“Local -features based ve hicle recongnition system using parallel vision
board “, institute of industrial science, university of Tokio Japon,
2000.
[13] Xin Li, XiaoCao Yao, Yi L. Murphey, Robert Karlsen, and Grant
Gerhart, “A Real -Time Vehicle Detection and Tracking System in
Outdoor Traffic Scenes”, Proceedings of the 17th International
Conference on Pattern Recognition (ICPR’04), 2004.
[14] ATIBI Mohamed , and all ,’ Real-time detection of vehicles using
the haar -like features and artificial neuron networks ”, The International
Conference on Advanced Wireless, Information, and Communication
Technologies (AWICT 2015) , 2015