The Motion Detection Based on Background

Difference Method and Active Contour Model

WANG Tong-yao
Institute of Electronic and Information Engineering
Dalian Jiaotong University
Dalian,China
[email protected]

Abstract—For the traditional moving target detection algorithm solve this problem, two-dimensional closed curves embedded
exists some problems such as determining threshold problems in a three-dimensional surface, with the evolution of surface
and more sensitive to noise, this paper presents a rapid and curve evolution [5,6]. Geometric Active Contour Model
extraction method based on variation level set of the edge blur based on the level set method using only edge information,
moving objects. This method used the background difference difficult to get the ideal target segmentation discrete fuzzy
method extract initial outline of moving target, and then used edges[7], and Chan-Vese Mumford-Shah model[3] provided
active contour model constraint edge detection; the final by the homogeneity of image global information region can
combined variation level set method with a second evolution to better segment the fuzzy boundaries or discrete boundaries of
obtain accurate image segmentation. Experiments show that this
the target. Combined with active contour model has the
method can rapidly and accurately segment moving object, for
moving objects of different shapes and the rules of small-angle advantages of the accuracy of edge detection, background
shot has a good adaptability and robustness. difference can quickly extract the initial outline of moving
objects and the other advantages, this paper presents a rapid
Keywords- Moving Target Detection; Active Contour method for solving the automatic moving object detection and
Model; Background Difference; Variation Level Set segmentation based on background difference and active
contour model of variation level set method.
I. INTRODUCE
II. METHOD
With the development of intelligent transportation systems,
as a core technology of the vehicle detection algorithm is an A. Active Contour Model
important computer vision research. To obtain the exact Active Contour Mode also known as Snake model, which
contours of the vehicle after the target classification, tracking first proposed by Kass in 1988, is a top-down approach and
and processing is very important, affecting the entire system widely used in regional tracking and image segmentation[4].
performance. Now the most common method of vehicle It integrates the feature extraction process of initial estimates,
detection is background difference, which use Gaussian the target contour feature and knowledge-based constraints,
Mixture Model or other statistical model extract background, and through proper initialization can take the initiative of the
with the current frame difference to extract moving regions, it convergence curve of the minimum energy state. Snake model
is vulnerable to sunlight, weather, shadows and other algorithm can significantly improve the reliability and
effects[1]. The background difference method the effect of the accuracy of contour extraction and conventional compared to
target area for movement is not the frame difference image moving target detection algorithm has obvious advantages.
and the background itself, but on the exact background of the Snake model can image through the application of local and
establishment, updating, in strong light or noise, especially global information campaign to achieve the objectives of the
under the influence of the camera is difficult to obtain the exact location of the border. Snake energy minimization
background activity; Optical Flow Method[2] computation is motion curves that model the first goal to be split according to
large, some scholars have proposed improved algorithm in the the contours of the manual to determine the initial
absence of specific hardware support is difficult to achieve approximate contour, is initialized on the image, and then the
real-time processing; Frame Difference[1] in terms of energy of the curve to minimize shrinkage, the final curve of
real-time display superiority, because it is the time interval lock target in the need to split the boundary.
between two successive frames is short, by the light changes,
camera jitter is very small. But it can not completely detect Snake is seen as a parametric
the object, by the speed of moving target is easy to form curve vs ( xs , ys ), s  (0,1) , the energy function is:
voids.The traditional motion detection algorithm to determine
a threshold difficulty is more sensitive to noise and other 1
1
³ Esnake (vs )ds ³ [ E (v )  E (vs' )  Econ (vs' )]ds
' '
shortcomings. Active Contour Model [3,4] to solve the still E 'snake s image
0
and moving image segmentation and target detection is an 0
effective method, its main disadvantage is that weak topology
adaptation, that adaptive evolution can not be split or merged.
Presented by the Osher and Sethian Level Set Method is to

___________________________________
978-1-4244-8625-0/11/$26.00 ©2011 IEEE


 ­1, z t 0 ½
1
1
³E (vs' )ds ³ [ E (v )  E (vs' )  Econ (vs' )]ds
'
E 'snake snake s image Using Heaviside function H ( z ) ® ¾ , can be on
¯0, z  0 ¿
0
0
(1) the edge of the loop integral for the area in the form of
rewriting points:
Internal energy is expressed as:

1 v³ g (C )ds ³³ g ( x, y) | ’H (u ) | dxdy (5)

Eint (D s ˜ | vs' |2  E s ˜ | vs'' |2 ) (2) c :
2
dH ( z )
vs' and vs'' respectively express the first and second For ’H (u ) G (u )’u ˈ G ( z ) ˈthen
dz
derivatives about vs ; D and E respectively express
control Snake the elastic force and bending forces weights,
they determine the point of Snake in a degree of extension and
E (u ) ³³ g ( x, y)G (u ) | ’u | dxdy
:
(6)
bending. Eimage express the energy of the image force, is used
to guide the Snake contour fit. Econ express force energy of Variation method can be used on the type of gradient
external constraints. Select the appropriate parameters to wu ’u
minimize the energy function, the corresponding curve is the descent flow ˖ G (u )div( g ) ˈ
final evolution of the edge of the object segmentation. wt | ’u |
B. Geodesic active contour models and Variation Level Set
Regularize G function,
Method wu ’u d
then G H (u )div( g ) , GH ( z) : HH ( z) ˈ
In 1995 V. Caselles proposed Geodesic Active Contour wt | ’u | dz
Model [7, 8]; this model is a geometric representation about
Snake model, which gives an even more perfect than the
HH ( z) as Heaviside function[7].
derivative description. Geodesic Active Contour Model is wu ’u wu ’u
based on the geodesic flow of the curve evolution, curve G (u )div( g ) and G H (u )div( g ) bec
evolution speed and related with the mean curvature. Level wt | ’u | wt | ’u |
Set Method was proposed by Osher is one of the realizations ome Variation level set method of Energy functional :
about the this model, it construct a three-dimensional surface
and the image plane intersect at Snake curve, the level set by E (C ) v³ g (C )ds . (7)
the evolution of this three-dimensional surface leaving Snake c
curve deformation[6]. Snake curve topology change will not
cause changes in the topology of three-dimensional surface, C. Building the background model to get background
which makes the level set for simultaneous detection of difference image
several objects dynamically. For the level set can detect the Background Difference because the algorithm is simple,
advantages of multiple moving targets, we use an improved fast running speed is widely used in the detection of moving
Variation Level Set Method, the following details awakened. objects [1]. Statistical average calculated during the
background image, the image sequence corresponding to each
Geodesic Active Contour Model [8] is the closed curve of pixel is calculated separately. In the image sequence
a minimum energy functional: corresponding to each pixel to calculate the average of the
pixel points to generate a background reference image frame.
E (C ) v³ g (C )ds ( 3)
From a statistical point of view of professional background
image frame gray value of each pixel can be seen as a
c
statistical result. That means a group of individual pixels in
the image sequence is statistically the most likely value.
wC
Gradient descent flow is: [ g (C )k  ’g x N ]N ˈ Generated by this principle to establish the background
wt statistical model, the simplest way is to choose average of
The corresponding function Partial Differential Equations is: background image sequence in a period of time as a frame [9].
Statistical average method is also called Average Method
[10], which is by calculating the average of a sequence of
multi-frame images in the corresponding gray level of each
wu ’u
( gk  ’g x N ) | ’u | | ’u | div( g ) (4) pixel as a background image the pixel value, the formula is as
wt | ’u | follows:


 N 1 (8) Step7. Stop the evolution;
1
Bt ( x, y )
N
¦ f ( x, y )
i Step8. Processing the evolution image by enhancement,
i 0 denoising and get the moving object.

Where that Bt ( x, y ) express the background of image III EXPERIMENT AND ANALYSIS RESULT

pixels point ( x, y ) ; f i ( x, y ) said the pixel gray value about In order to validate the algorithm, we used about moving
vehicle traffic video sequence of Anshan City as the detection
the point ( x, y ) at the frame i, N is the number of frames experiment section. In this paper, use the AVI format video as
for the image sequence. the research object, operating environment is: Inter Corei5
CPU [email protected],4GB RAM, Windows 7 Ultimate,
Background image Bt 1 ( x, y ) at time t  1 , which is Matlab 7.
from the image f t ( x, y ) at time t with the frame
background image Bt ( x, y ) at the moment of t, is
calculated by the linear combination, calculated as follows:

­ E Bt ( x, y )  (1  D ) I n ( x, y ), Qn 0
Bt 1 ( x, y ) ®
¯1, others
(9)

Where is the binary model of the mask:

°­0, ft ( x, y )  Bt ( x, y )  T
Qn ( x, y ) ® (10)
°̄1ˈothers Figure1. The 50th frame image

To ensure there is no update in the background there is a

moving vehicle, motor vehicles, in part by the time frame of
the background image the same area instead. Time constant
D  [0,1] which represents the background image update
rate, the experiment is taken as 0.5.
D. Improved algorithm steps
With Background Difference method, Active Contour
Model and Variation Level Set Method for moving target
detection advantages, to overcome the shortcomings of these
three methodsDŽIn this paper, we propose a more robust
method of moving target detection about these three methods
integrated application, the method specific described as
follow:
Step1.Using statistical averaging method for background Figure2. Background image of the 50th frame image
image generation;
Step2.Using background difference method, take the
current frame image difference with the background image to
get the foreground image which obtain the differential image;
Step3.Using active contour model treat the difference
image to obtain the initial outline of moving objects.
Step4. Using the initial outline to initialized level set
curve;
Step5. Using the level set equation to level set curve
evolution for moving target detection;
Step6. Determine the value of the edge indicator function
is the maximum, if it is then to Step7, otherwise go to Step 5,
continue to curve evolution;


 Figure3. Background Difference Image Contour Model initial contour detection make the target
within the local limits, greatly reduces the detection level set
by the target number of iterations, Improve the time efficiency
of the algorithmˈFrom the experimental results, the algorithm
is accurate and effective, but not the shape of a moving target
can be rapid and accurate detection, the algorithm greatly
improve the applicability and robustness.
IV CONCLUSION
This thesis uses statistical background image generation
method, and then extracted using background difference of
the initial outline of moving objects, then the main motion
active contour model to extract the initial contour bound
targets, and finally with the initial level set contour evolution
Figure.4 Initial outline of the difference image method to obtain accurate segmentation. The method
considers the application of light to the world the changes in
external factors such as the impact of moving images, on the
environment has a robust, compared to the conventional use
of background subtraction and other traditional methods for
moving object detection, this method advantage of using the
active contour model contour of moving objects quadratic
constraint evolution, segmentation results obtained are more
accurate, complete and clear. Experimental results show that
for a fixed camera or mobile camera for small angle, the use
of the second evolution, making the movement of objects of
different shapes and angles, small changes in circumstances
have had good adaptability.
REFERENCES
[1] EL GAMMAL A, DURAISWAMI R , HARWOOD D , etal.
Figure.5 Initializing level set curve Background and foreground modeling using nonparametric kernel
density estimation for visual surveillance [ J ].Proceedings of the IEEE,
2002 , 90( 7) : 1151-1163.
[2] Dougherty, L., Asmuth, J.C., Blom, A.S., Axel, L., Kumar, R..
Validation of an optical flow method for tag displacement estimation
[J].IEEE Transaction on Medical Imaging, 1999, 18(4): 359 – 363.
[3] Chan T F, Vese L. Active contours without edges. CAM Report98-53.
UCLA, 1998.
[4] Chan T F, Sandberg B Y, Vese L A. Active contours without edge for
vector-valued images. J. of Visual Communication and Image
Representation, 2000.11: 130-141.
[5] Osher Sˈ Sethian Jˊ Fronts propagating with curvature-dependent
speed ˖ Algoritmhs based on Hmailton-Jacobi formulation ˈ
3ˊComputˊPhysˈ1998:79-1249.
[6] Osher S ˈ Fedkiw R ˊ Level Set Methods and Dynamic Implicit
Surfaces[M]ˊNew York˖Springer-Verlagˈ2002˖34-67.
[7] CASELLES V, KIMMEL R, SAPIRO G. Geodesic active contours[J].
Fig.6 The result of moving target detection International Journal of Computer Vision,1997, 22(1):61-79.
[8] Paragios N.Deriche R.Geodesie active contours and level sets for the
The Figure.1 shows that the 50th frame image of the video detection and tracking of moving objects[J].IEEE Transactions on
image, there is three moving target in the image. The Figure.2 Pattern Analysis and Machine Intelligence2000,3(22)˖266—280.
shows that we use statistical averaging method for generating [9] Song Gao, Bui, T.D.. Image segmentation and selective smoothing by
the background image. Then we use background difference using Mumford-Shah model[J]. IEEE Transaction on Image Processing,
method to take the current frame image (Figure.1) difference 2005, 10(14): 1537 – 1549.
with the background image (Figure.2) to get the foreground [10] M SanjeeArulampalamˈSimon MaskellˈNeil Gordon, and Tim Clapp.
A Tutorial on Particle Filters for Online Nonlinear/Non-Gaussian
image (Fig.3) which obtains the differential image. The
Bayesian TrackingˈIEEE Transactions on Signal Processingˈ2002ˈ
Figure.4 shows that we use Active Contour Model to treat the 50(2)174-188.
Figure.3 to obtain the initial outline of moving objects, and as
the level set curve initialization show in the Figure.5.Figure.6
shows that the moving target detection results about this
method.
Using background subtraction to extract the moving
objects quickly and accurately, then through by the Active