2012 Fifth International Symposium on Computational Intelligence and Design

A Vehicle Detection Algorithm Based on Three-frame Differencing and Background

Subtraction

Haiying Zhang1, Kun Wu1

1. Software School of Xiamen University, Xiamen, China
zhang2002@xmu.edu.cn, wk6295959@163.com

Abstract—A vehicle detection algorithm based on three-frame effective and practical background model and the method to
differencing and background subtraction is presented in this update is the focus of research methods.
paper. Firstly, improved GMM is for background subtraction, According to the shortcomings of frame differencing
then the moving object region is gained using background method and background subtraction method, this paper uses
subtraction, and then the background subtraction is combined improved GMM to get background frame, then uses
with three-frame differencing to detect the motion information. background subtraction method to extract the moving object,
The simulation results show that the proposed algorithm runs again with improved frame differencing algorithm that is
veraciously and can lower the false detection rate, and fits for three-frame differencing are combined to form a new simple
real time detection.
and effective method for the detection of moving target.
Keywords-Three-frame Differencing; Background II. MOTION DETECTION ALGORITHM BASED ON THREE
subtraction; Motion detection
FRAME DIFFERENCING AND BACKGROUND SUBTRACTION

I. INTRODUCTION A. Algorithm Idea

Along with the development of technology in the digital Algorithm idea is mainly consist of background
TV and video monitoring, the effective detection and extraction and update, background subtraction, three-frame
extraction of moving object which is essential to the target differencing, motion detection and the post-processing.
tracking and identification has become a key technology. Firstly, use of the Gaussian mixture modeling background
Processing speed and reliability are the two basic problems for background image, then with the source image use
to the moving object detection, and also are the measure of background subtraction and three-frame differencing
quality related algorithm of two major indexes. The moving operation respectively, and get the middle template image
object detection method mainly divided into three categories, with the using of the two binary image for the logic “and”
including optical flow method, frame difference method and operation. Secondly, using the improved GMM is to update
background subtraction method; there are their advantages background. Finally, the template is necessary post-
and disadvantages: processed, including stop the template with small noise,
(1) Optical flow method uses the motion target of the regional, and smooth out the boundary of the objects
vector characteristics which changed with time to detect segmentation. Background update model and three frame
motion area in image sequences. Its performance is better differencing is the core and key of the algorithm, and
under the condition of camera moving, but this algorithm is background update model is the basis of background
complex and complicated computation and needs special subtraction.
hardware support, so it is difficult to meet the requirements
of real-time video processing. B. Background Extraction and Update
(2) Frame differencing method uses the two or three The static background is pixels with the gray not changed
adjacent frame based on time series image to subtract and in image sequences. In moving object detection algorithm
gets difference images, then gets moving target information based on video image, if taking background subtraction
through the threshold value. This method is simple and easy method for target detection, the establishment and update of
to implement, but it is difficult to get moving targets background will be the key to success for algorithm. This
complete contour, and easily generated “double” and “holes” background image should not contain any moving targets,
phenomenon in target internally, consequently lead to but such images in the practical application are not easy to
inaccurate detected target information. get. A kind of simple way is to take the first video frame as
(3) The basic idea of background subtraction method is to the initial background. But with the passage of time, as the
build a background firstly, and then uses the current frame light changes and the weather suddenly change, need to
subtracting background frame to detect moving object. This update the background image. This paper analyzes the
method is simple and easy to realize, and accurately extracts Gaussian mixture model and the method of background
the characteristics of target data, but it is sensitive to the updates, improved the Gaussian mixture model, in order to
change of external environment, so it is applicable to the improve the ability to adapt to the environment background
condition that the background is known. How to find more model.

978-0-7695-4811-1/12 $26.00 © 2012 IEEE 148

DOI 10.1109/ISCID.2012.45
 1) Background Modeling
    N 
; W
 (2.9)
The basic idea of Gaussian background modeling is, from N 
the time axis see, every pixel in a period of time inside of From equations (2.7)-(2.9) is not difficult to find that
observations can be made of K independent components of learning rate α direct influence on the algorithm of the ability
the Gaussian weighted and composition. Calculation of each to respond to the environment. When α and ρ is big, Xt as
Gaussian distribution weights and according to weights from current frame is in a larger share, so applicability to the
big to small sorting, the greater the weights that represent the environment, and a strong ability to quickly reflect
likelihood of a background, and then select a Gaussian environment change, but at the same time easily affected by
distribution before B building background model [1, 2]. noise and not quite stable, and will likely judge slowly
xt instead of the value of each pixel at the time t in video moving target as the background; When α is small, to a
image, like the expression given by equation (2.1).
; ; ;   [ \  
certain extent to reduce the noise, but too slow to update,

  
 , W
 L    L  W (2.1)

when the new target suddenly appeared, it cannot be detected.

Each pixel is modeled respectively by GMM with the K 3) Improved GMM
Gaussian distribution as detailed in equation (2.2). Analyzing GMM, we can get that the selection of

; 
background model updating rate directly influence the model
;   
.
S   (2.2) capacity of adaptation to environment. From (2.7)-(2.9)
W
L 
L W W L W L W found that μt and σt2 is the same update factor ρ. But proved
Among them, the K for GMM in the number of Gaussian by the experiment, actually mean value and variance to the
distribution, generally take 3 to 5; ωi,t represents the estimate last of the discriminant influence is different. Mean value
of weight coefficient in the i Gaussian distribution; μ i,t influences the sensitivity of detection, and variance
represents mean value covariance matrix and Σi,t represents significantly influence model on adapted speed, so mean
vector covariance matrix; η represents Gaussian distribution value and variance use the same update factor is not
probability density function. reasonable. In order to solve this problem, mean value and
    variance with a different update factor ρ1 and ρ2, like the
 0 (2.3) expression given by equation (2.10) and (2.11) [4, 5].
 
N W N W  N W

;   

; W  W  7  
 ; W 

(2.4)
W
 
 W 
 ;
 W
(2.10)
  H   W

      ; W   
W Q  7


 
    (2.11)
W ;
 
W  W   W W
Among them, ρ1 and ρ2 is a constant and is concentrated
N ,

N W
 (2.5) on the [0, 1] interval, in addition ρ1 > ρ2. ρ1 is larger in order
Each pixel will be made up of the K Gaussian to improve the detection sensitivity, ρ2 is smaller in order to
distribution according to the weights from big to small reduce the extent of model, and don’t adapt the target into
sorting, then most likely to describe the process of stable the background.
background Gaussian distribution will be located at the top C. Background Subtraction
of the sequence, and transient disturbance by background
After get the background image Bi(x, y), with the current
from the distribution will sliding into the bottom sequence,
frame Ii and background Bi(x, y) differencing, when the
and finally by the new assignment of the Gaussian
pixel grayscale value d of the difference image DBi is greater
distribution replaced. So, choose the front B Gaussian
than the threshold T, the detection is judged as the moving
distribution as background pixel model, like the expression
targets, like the expression given by equation (2.12) and
given by equation (2.6).
(2.13) [8, 9].

, [ \   % [ \ 
 E 
%  DUJ PLQ   N  7  (2.6)
G  (2.12)
E
 N    L
2) Background Model Update  G  7
To each frame image, compared with the pixels of each '% [ \  
L 
 G  7
(2.13)
Gaussian model, if meet | Xt-μt-1 |≤2.5 σt-1, it is said the pixel
matches background model and belongs to the background; Among them, Ii(x, y) represents current frame pixel
otherwise the pixel belongs to foreground [3]. grayscale value, B(x, y) represents background frame pixel
For not match the Gaussian distribution, not to update, grayscale value, d represents the difference image DBi pixel
for matching of the Gaussian distribution, according to the grayscale value, and DBi(x, y) represents the binary image
equations (2.7)-(2.9) to update the mean value and the pixel grayscale value, T set for the threshold.
variance of the model, ρ as the update factor, α as learning D. Three -frame Differencing
rate.
    
In the process of image processing, frame differencing is
  ; W
(2.7) calculated difference of two adjacent frames Ii(x, y) and Ii +
W W 

; W  W  ; 
1(x, y) to get moving region, through the difference image
      W   
7


 
W W
 (2.8) can rapidly detect the range of movement caused by moving
W

149
objects in adjacent images, but it's moving object extraction
bigger than the actual, and often has "double" phenomenon,
in addition, because of the detected object is two frames
before and after relative parts, can't detect overlap, lead to
detect target has "empty" phenomenon. And three frame
differencing is improved on the basis of the frame
differencing, it takes the three adjacent images as a group
frame and differences again up to detect the shape of the
moving object contour of the intermediate frame better [10].
Set image sequence as: Ii(x, y), Select three consecutive
frames image Ii-1(x, y), Ii(x, y), Ii+1(x, y) in image sequences,
and respectively calculate the difference value image
between two frames image, like the expression given by
equation (2.14).

G L L  [ \   , L [ \   , L  [ \ 

(2.14)

G L L 
 [ \   , L  [ \   ,L [ \  Figure 1. From left to right: source image, background subtraction, three-
frame differencing, the proposed method.
Through select proper threshold to get the binary image
according to equation (2.15). From the experiment results can be seen, with the traffic
 G L L  [ \   7
 monitoring video influenced by the change of light, resulting
E [ \   
in background subtraction method extracted moving targets


L L 
 G L L  [ \   7 (2.15)
more noise, and vehicle outline is incomplete; And the
detection with “double” phenomenon is relatively obvious in
 
  G L L  [ \   7
E L L  [ \   
three-frame differencing; This paper combines two kinds of
  G L L  [ \   7 methods, can avoid light effect, and eliminate the “double”
phenomenon, obtain more accurate and complete test results.
With binary image use logic "or" operation for each pixel This paper also contrasts three kinds of methods on the
(x, y) getting other binary image Li(x, y) according to processing time and detection accuracy and mistake
equation (2.16). detection rate. Table 1 to Table 4 details the obtained results.

 E [ \   E L [ \    (2.16)
/ [ \  
L L    L 

L


 EL L  
 
 [ \  E L  L
[ \   TABLE I. THREE METHODS OF PROCESSING TIME CONTRAST
(UNIT:S)
Necessary treatment is taken on binary image Li(x, y) to Time needed
remove the noise point of the images and the hole of targets, Algorithm 33s test 58s test 130s test
this can through the morphological filtering on to achieve, video video video
such as open and close operations. Background subtraction 64 90 250
E. Moving targets detection Three-frame differencing 45 79 189
Three-frame differencing adapts effectively to the change The proposed method 65 99 290
of environment light, overcomes the shortcoming of
background subtraction; and background subtraction does
not exist "double" and "holes" phenomenon, can make up the TABLE II. GMM BACKGROUND SUBTRACTION EXPERIMENTAL
defects of three-frame differencing. So, with background RESULTS

subtraction and three-frame differencing will get more Statistical results

Video
accurate complete results. With binary images which got sequences Traffic Correct Fall- Fallout Accuracy
Omission
from background subtraction and three-frame differencing volume detection out ratio ratio
use logic "and" operation to get results binary image Ri(x, y), Video 1 10 9 1 0 0 90%
like the expression given by equation (2.17). Video 2 49 45 4 2 4% 91.8%
 [ \   /L [ \    (2.17)
5 L [ \    '%
 L

'% L [ \   /L [ \  
Video 3 97 88 9 4 3% 90.7%

Video 4 138 125 12 6 4.3% 90.5%

III. EXPERIMENTAL RESULTS

This paper uses Visual Studio 2008 and Open CV 2.3.1 TABLE III. THREE-FRAME DIFFRERNCING EXPERIMENTAL RESULTS
for software development tools to test the sample of highway
traffic video. Background subtraction method, three-frame Statistical results
Video
Traffic Correct Fall- Fallout Accuracy
differencing method and the proposed method had been test sequences
volume detection
Omission
out ratio ratio
respectively, testing results as shown in figure 3-9.

150
 Statistical results same time can overcome light change to the influence of
Video
sequences Traffic Correct
Omission
Fall- Fallout Accuracy background subtraction and eliminate three-frame
volume detection out ratio ratio differencing bringing “double” phenomenon, anti-jamming
Video 1 10 10 0 1 10% 100% is strong, can get moving targets more complete regions, and
Video 2 49 46 3 5 10% 93.8% then identify moving objects from video series. The
experimental results show that this method is accurate and
Video 3 97 92 5 9 9.2% 94.8% rapid, and mistake detection rate is low, and has broad
Video 4 138 130 8 13 9.4% 94.2% application. Be used in such as neighborhood monitoring,
traffic monitoring, and so on. That can reduce manpower and
material resources, and get a good economic benefit.
TABLE IV. THE PROPOSED METHOD EXPERIMENTAL RESULTS
REFERENCES
Statistical results
Video [1] Tang Z, Miao Z J. Fast background subtraction and shadow
sequences Traffic Correct Fall- Fallout Accuracy
Omission elimination using Improved Gaussian mixture model[C]. IEEE
volume detection out ratio ratio
International Workshop on Haptic, Audio and Visual Environments
Video 1 10 10 0 0 0 100% and Games, 2007: 38-41.
Video 2 49 47 2 1 2% 95.9% [2] Ueng S K, Cheng H P, Lu R Y. An adaptive gauss filtering method
visualization symposium[J]. IEEE, 2008, (3): 127-134.
Video 3 97 94 3 2 2% 96.9%
[3] Gutehess D, Trajkovies M, Cohen Solal E. A background model
Video 4 138 133 5 4 2.8% 96.3% initialization algorithm for video surveillance, Computer Vision[C].
Proceedings, Eighth IEEE International Conference, 2001, 11(1):
733-740.
[4] Lipton A, Fujiyoshi H, Patil R. Moving target classification and
From the time, the proposed method is the longest for the tracking from real-time video[C]. IEEE Workshop on Applications of
combination of the two algorithms, but within the acceptable Computer Vision, Princeton: IEEE Press, 1998: 8-14.
range, and meets the real-time. From fallout ratio and [5] Valera M, Velastin S.A.Intelligent distributed surveillance systems[J].
IEEE Proc. Vision Image Signal Process, 2005, 152(2):192-204.
accuracy ratio, GMM background subtraction affected by
environmental impact is bigger, so its accuracy ratio is 90%, [6] Arseneau S, Cooperstock J. Real-time image segmentation for action
recognition [C]. Proc IEEE Pacific Rim Conference on
and three-frame differencing has higher accuracy ratio at Communications. Computers and Signal Processing, Canada: Victoria,
94%, otherwise fallout ratio of three-frame differencing is 1999: 86-89.
10% because of “double” phenomenon, and the proposed [7] Barron J, Fleet D, Beauchemin S. Performance of optical flow
method improved obviously at fallout ratio and accuracy techniques[J]. International Computer Vision, 1994, 12(1):42-77
ratio, by fallout ratio decreasing below 3% and accuracy [8] X.H.Zhou, B.Liu, and H.Q.Zhou. A motion detection method based
ratio also improving to more than 95%. on symmetrical differencing and background subtraciton[J]. The
computer simulation, 2005, 4: 117-120.
IV. SUMMARY [9] H.Wen, T.Hong. Based on background subtraction and symmetrical
difference target detection[J]. Micro computer information, 2007,9(1):
The innovation point of this paper is put forward a 99-101.
moving object detection algorithm which is based on three- [10] G.L.lv, S.G.Zhao, J.Zhao. A new method for motion detection based
frame differencing and background subtraction. This method on three-frame differencing and connectivity testing of target
can be applied to complex environment, allow the presence image[J]. LCD and Display, 2007, 22(1): 87-93.
of moving objects when background modeling, and at the

151