VEHICLE RECOGNITION USING VIBE AND SVM

Computer Science & Engineering: An International Journal (CSEIJ), Vol.6, No.4, August 2016
DOI:10.5121/cseij.2016.6402 21
VEHICLE RECOGNITION USING VIBE AND SVM
Jinlin Liu, Qiang Chen and Chen Zhang
College of Electronic and Electrical Engineering, Shanghai University of Engineering
Science, Shanghai 201620,China.
ABSTRACT
Video surveillance is becoming more and more important forsocial security, law enforcement, social
order,military, and other social problems. In order to manage parking information effectively, this vehicle
detection method is presented. In general, motion detection plays an important role in video surveillance
systems. In this paper, firstly this system uses ViBe method to extract the foreground object, then extracts
HOG features on the performance of the ROI of images. At last this paper presents Support vector machine
for vehicle recognition. The results of this test show that, the recognition rate of vehicle’s model in this
recognition system is up the industrial application standard.
KEYWORDS
ViBe , SVM, vehicle recognition, HOG
1. INTRODUCTION
With the development of the city, more and more people have their own vehicles. Meanwhile, the
speed of parking lot construction cannot compare with the rate of car growth. So there was a
contradiction between people and between the supply and demand of parking. In order to build
smart city and facilitate people’s lives, we have to solve this problem[1].
Now in many cities have been building many parking lots which have a very high level of
automation, and can remind divers how many free parking locations when they entered the parking
lots. Even some of the higher-end parking lot offers Apps, and divers can book theirs parking space
online. But for those who do not plan well site, monitoring of vehicles are very loose, or in a state of
lack of supervision.[2] There is no doubt that this situation will give some unpleasant feelings.
Therefore, in order to better build smart cities, planning and management of outdoor parking lots is
necessary.
This paper introduces how to tell the moving vehicles from parking lot surveillance video. Using
ViBe background modeling method for moving target detection to identify moving objects. This

22
system uses SVM method determines whether the object is a vehicle, and judge the state of motion
of the vehicle, and then updates the information of parking spaces.
2.BACKGROUND SUBTRACTION ALGORITHMS
Locating moving objects in a video sequence is the first step of many computer vision
applications.Many background subtraction techniques have been proposed with as many models
and segmentation strategies, and several surveys are devoted to this topic.This system uses ViBe
method detecting moving foreground objects
2.1 Brief introduction of VIBE
Background differencing is the commonly used method for the detection of moving objects in the
static background, and the ViBe algorithm is the main modeling approach. our proposed technique
stores, for each pixel, a set of values taken in the past at the same location or in the neighborhood. It
then compares this setting for the current pixel value to determine if the pixel belongs to the
background, and adapts to the model of a random selection of values instead of from the
background model. This method is different from those based on the oldest values of the classical
belief should be replaced first.
Finally, if the pixel is one part of the background, then its value will be propagated into the
background model and update the model after judgement.
ViBe algorithm using neighboring pixels in the color space determine the current pixel is the
foreground pixels or background pixels, and it establishes a background pixel samples for each
pixel value space, and save the pixel and its neighboring pixels in the sample space recent
background pixel values. The current pixel-by-pixel with the pixel values of the background
models with the most similar comparison, value as long as the current pixel-by-pixel with the
background part of the pixel values in the model, determine the background pixels pixel, pixel
background model can be expressed as
= , , . . . , (1)
In this paper, we denote by v(x) the value in a given Euclidean color space taken by the pixel
located at x in the image, and by a background sample value with an index i. Each background
pixel x is modeled by a collection of N background sample values.[3]
VIBE initialization algorithms use the first frame image background model for every sample value
of the background pixels in the sample space from the pixel and its neighboring pixels, selects a
random pixel values to initialize.Each pixel has the same probability of being chosen. The model
can be expressed as

23
0
= 0 ∈ , (2)
y is one pixel selected in the neighborhood of x randomly，and is the collection of
neighborhoods. v(y) pixel values may appear more than once in the background model.
ViBe algorithm starts to detect foreground objects from second frame. In order to classify a pixel
value v(x) on the basis of its corresponding model M(x), we compare it to the closest values within
the set of samples by defining a sphere of radius R centered on v(x)[3].If thegiven
threshold ♯min is less than thecount of the set junction of sphere and the collection of model
samples M(x), the pixel value v(x) will be classified as background. More formally, we compare
♯min to
♯ ∩ , , . . . , (3)
If M(x) is larger than ♯min, the pixel x is background pixel, otherwise x is foreground pixel [4]. We
can learn that directly from this figure.
Fig.1. Comparison of a pixel value with a set of samples in a two dimensional Euclidean color space (C1,C2).
To classify v(x), we count the number of samples of M(x) interesting the sphere of radius R centered on v(x).
2.2 Experimental Results
For the experiments , the proposed method was implemented using C++ with OpenCV library. In
this experiment, we set the radius R of the sphere used to compare a new pixel value to pixel
samples, the time subsampling factor Φ,the number N of samples stored in each pixel model, and
the number ♯min of close pixel samples needed to classify a new pixel value as background. In our
experience , we set radius R = 20, and time subsampling factor Φ=16，N=20，and set ♯min=20.

24
Our dataset is made of videos from parking lot, we pick a pieceof video for experiment to test the
effect of ViBe algorithm. In order to highlight the prospects for detection of objects and have a
good performance, we add the code of merge window.
(a) (b)
(c) (d)
Fig.2. Input video is (a), after the process of foreground objects detection is (b),After morphological
processing and edge processing is (c). Locating moving objects region in the picuture(d)
3.VEHICLES RECOGNITION
When after receiving the moving image, the next we are going to determine whether objects are
vehicles. There are many ways to classify, such as logistic regression analysis, kernel logistic
regression, Support Vector Machine (SVM). Support vector machine is a new and rapidly
developing methods of machine learning and classification, are widely used in many types of
classification and pattern recognition.
3.1 Briefintroduction of SVM
SVMs are primarily two-class classifiers that have been proved thatit is an effective and more
robust method to handle linear or non-linear decision boundaries [5] [6] . SVM will find the
hyperplanebya set of points, which belong to either of two classes. This hyperplanehas the largest
distance to each class,and the largest possible fraction of points of the same class on the same

25
side.This is equivalent to the implementation of structural risk minimization to achieve good
generalization [5] [6]. Assuming l examples from two classes.
In this section, We suppose we have a set S which has N points of training samples, ∈ ℝ"
withi =
1, 2, … , N. Each point belong to one of two classes and is given a label ∈ −1,1 .Our
purposeis to build a hyperplane equationthat divides S into two classes correctly, and maximize the
distance between the two classes and the hyperplane [6]. A hyperplane can be described as the
equationin the feature space.
< & , > +) = 0 (4)
where w ∈ ℝ"
and b is a scalar. When the training samples are linearly separable, Optimal
hyperplane of SVM, which can separate the two classes, no training mistakes and maximize the
minimum distance from the sample to the hyperplane.
3.2 Histograms of Oriented Gradients (HOG)
Histogram of oriented gradients (HOG) feature is a descriptor for computer vision and image
processing todetect object. Through calculations and statistical characteristics of gradient
orientation histogram of image local area to form. Hog combination of SVM classifiers have been
widely used in image recognition, especially in the detection of pedestrians has a great
performance.In practice, the orientation range is divided into + bins, and each pixel in the region
votes for its corresponding bin.
3.3 Experimental Results
Framework of the vehicle recognition as shown in the figure: including learning and testing phases.
During the learning phase, mainly is the integration of collected data, get these features, then
entered into a learning classifier, to generate the target classifier. After entering the testing phase,
you need to use the classification obtained in the previous step, by comparing the scan window and
then click the input image, is judged to be identified in accordance with the classifier of the small
area [7][8].

26
Learning Phase
Collecttraining data
set
Create feature setof
training D ata set
Build a binary classifier
Detection Phase
D etection w indow
scans
Classifierclassifies the
detection w indow
Create a training data
set
Learning Phase
Collecttraining data
set
Create feature setof
training D ata set
Build a binary classifier
Detection Phase
D etection w indow
scans
Classifierclassifies the
detection w indow
Create a training data
set
Fig.3 Flow chart of SVM
In our experiment, we collect training data from ImageNet, we download the pictures, and resize
them 32*32, so the feature dimensions of each image is 1764.
Fig.4 positivesample images
After the system completes the learning phase, the system will judge the foreground objects. Flow

27
chart is as follows.
Input ROI and resize Gradient computation
Gradient weight
projection
Scan window to get the
regional image feature
Predict the SVM classifier
Determination of the
merge window
Output results
Contrast normalization
Fig.5 detction process flow chart
In this section, the HOG feature of foreground objects will be sent to SVM. Every window will be
tested and predicted by SVM. The SVM will output “-1” or ”1”. The”1” presents the vehicle,”-1”
presents non-vehicle. This information will show in a new window.
Fig.6 After passing vehicle recognition, we remove the vehicle pictures and displays
Fig.7 This design can detect more than 2 vehicles at the same time.

28
4. CONCLUSION
We have considered the problem of vehicle detection from video surveillance of parking lots. The
HOG can describe vehicle information, handle within-class variations, and global illumination
insensitive to changes in.
The first part of the study is devoted to the analysis of the performance of ViBe for foreground
objects detection. SVM based on HOG features is shown to achieve the good results. However, the
outcome of the study shows that, results with high accuracy, but the efficiency problem happens in
certain situations. The outcome of experiments reveals that SVM can work well in normal
situation, but in some extreme situationssuch as tInclement weather, the effectofrecognition of a
little bad.In order to enhance the robustness and accuracy of the system, we plan to integrate other
features (for example, LBP feature), and use optical flow method to optimizethe moving object
detection and background modelling.
REFERENCES
[1] S. Suryanto, D.-H. Kim, H.-K. Kim, and S.-J. Ko, “Spatial colorhistogram based center voting method
for subsequent objecttracking and segmentation,” Image and Vision Computing, vol.29, no. 12, pp.
850–860, 2011.
[2] J. Hwang, K. Huh, and D. Lee, “Vision-based vehicle detection and tracking algorithm design,” Optical
Engineering, vol. 48, no.12, Article ID 127201, 2009.
[3] O. Barnich and M. Van Droogenbroeck, “ViBe: a powerful random technique to estimate the
background in video sequences,” in Int. Conf.on Acoustics, Speech and Signal Processing (ICASSP),
pp. 945–948,April 2009.
[4] C. Burges, “Tutorial on support vector machines for pattern recognition,” Data Mining and Knowledge
Discovery, vol. 2, no. 2, pp. 955–974, 1998.
[5] L.-W. Tsai, J.-W. Hsieh, and K.-C. Fan, “Vehicle detection using normalized color and edge map,”
IEEE Transactions on Image Processing, vol. 16, no. 3, pp. 850–864, 2007.
[6] V. Vapnik, The Nature of Statistical Learning Theory. Springer Verlag, 1995.
[7] M. Van Droogenbroeck and O. Barnich, “Visual background extractor.”World Intellectual Property
Organization, WO 2009/007198, 36 pages, January 2009.
[8] Joshua Gleason, Ara V. Nefian, Xavier Bouyssounousse, Terry Fong and George Bebis 2011
[9] G´erardBiau, Luc Devroye, and G´abor Lugosi. Consistency of random forests and other averaging
classifiers. J. Mach. Learn. Res., 9:2015–2033, 2008.
[10] C. N. Anagnostopoulos, I. Giannoukos, T. Alexandropoulos, A. Psyllos, V. Loumos, and E. Kayafas,
"Integrated vehicle recognition and inspection system to improve security in restricted access areas," in
Intelligent Transportation Systems (ITSC)", 2010 13th International IEEE Conference on, 2010, pp.
1893-898. [3] L. Bottou, C. Cortes , J.

29
AUTHORS
Jinlin Liu is currently studying in Mechanical and Electronic Engineering fromShanghai
University of Engineering Science, China, where he is working towards theMaster
degree. His current research interests include image processing, machine learning

VEHICLE RECOGNITION USING VIBE AND SVM

More Related Content

What's hot (19)

Similar to VEHICLE RECOGNITION USING VIBE AND SVM (20)

Recently uploaded (20)

VEHICLE RECOGNITION USING VIBE AND SVM