- binarization

Denoising
connectivity analyzing
distinguishing
Difference f
k-1
f
k
D
k
R
k
P
k
Object Tracking Algorithm Based on Camshift Algorithm
Combinating with Difference in Frame

Hongxia Chu
1
, Shujiang Ye
1
, Qingchang Guo
2
, Xia Liu
2

Department of Electronic Engineering College of automation
Heilongjiang Institute of Technology
1
Harbin Engineering University
2

Harbin 150001, Heilongjiang Province, China Harbin 150001, Heilongjiang Province, China
[email protected] [email protected]
Abstract - Object tracking algorithm based on Camshift
algorithm which is combinating with difference in frame is put
forward to quickly and exactly track movement object. Firstly,
we confirm the movement region of object by difference in frame,
and we again confirm centroid of movement region which is
centered to initialize tracking curve. Secondly, the feature is
abstracted in the region and the target is tracked by Camshift
algorithm. The method overcomes shortcoming of artificial
orientation and divergence problem for traditional Camshift
during tracking. We prove the validity of the algorithm by
experiment at last.

Index Terms - Difference. Frame. Object tracking. Camshift

I. INTRODUCTION
The moving object detection is an important subject in
the field of application vision research. In real life, much
useful vision information is included in movement. Human
vision can see not only movement object but also static one,
but people are sometimes much more interested in the
movement object in many situations such as the detection of
traffic flow and the inspection of important site's safety and
the control and guide of aviation and warplane and assistant
driving and so on. The research target of the moving object is
sequential images which contain much information than single
frame image. So it is significative for the research of the
moving object detection and tracking system.
The research on the moving object detection and tracking
is separated into two kinds:
1) Cinema head shifts by following movement object and
always keeps target in the center of image
[1][2][3]
;
2) Cinema head is fixed and movement object only is
tracked

in vision field
[4]
.
Existing algorithms consist of the algorithm based on
template match and the algorithm based on beset picture and
optical flow and so on
[5][6]
. The object is all matched directly
in these algorithms and only the static single picture in
sequence is processed, but these algorithms don't take into
account relativity among sequential images.
The algorithm studies on moving object detection and
tacking in static scene. Namely, the method of difference in
frame is used in moving object detecting and Camshift
algorithm is applied in the moving object tracking. Different
image between two frames is still calculated during the
tacking by using Camshift algorithm in order that movement
scope of moving object is ascertained. So the divergence of
tracking window is also prevented.
II. OBJECT POSITIONING
A. Theory of Difference in Frame
Difference method is the most direct method to examine
movement information from serial image. The method uses
grey degree and grads of space-time in picking up movement
information. The method calculates difference between former
frame and back frame by comparing point-by-point grey
value.
The formula of difference in frame can be written in (1)
and (2):
D
k
(x, y)=|f
k
(x, y)-f
k-1
(x, y)| (1)

( )

=
1
0
, y x R
k

background
object

( )
( ) T y x
T y x

>
, D if
, D if
k
k
(2)

Where D
k
is image after difference, R
k
is image after
binarization, f
k
is current frame image, f
k-1
is previous frame
image, T is threshold.






Fig. 1 Illustrating difference in frame
Difference method in frame is used in object detecting
and abstracting by applying diversity between continuous two
frames or several frames in video sequence. The process of a
most essential difference method in frame is shown in Fig.1.
First, Equation (1) calculates difference between No.k frame
image and No.k-1 one and gains image D
k
after difference.
Then the image is binarizated in equation (2). When a certain
pixel value is bigger than a given threshold, we think this pixel
as foreground's pixel which is perhaps a point on the object.
Otherwise, we suggest it background's pixel. We obtain image
R
k
after image D
k
is binarizated
.
Finally, we analyze
connectivity about image R
k
and get image P
k
. When the area
of some connective region is bigger than a given threshold, we
detect object and think this region is possessed by the object.
The merit of video detection and abstracting based on
difference method in frame is shown as follows:
1) Algorithm can be easily realized.
2) Complexity degree of programmer is lower.
3) It easily implements real-time watching.
4) The method is not too sensitive to the change of beam
1-4244-1531-4/07/$25.00 2007 IEEE. 51
Proceedings of the IEEE
International Conference on Automation and Logistics
August 18 - 21, 2007, Jinan, China
in scene because interval between frames is commonly shorter.
Shortcoming of difference method in frame is shown in
the following:
1) The positioning of moving object is not too exact.
Area of detection is sportive variation district in the forward
and behind frame.
2) Movement velocity of target is a bit slow because
interval between frames is commonly shorter.
Experiment on difference in frame is shown in Fig. 2.
Where Fig. (2.a) and Fig. (2.b) is respectively continuous two
frame images. Fig. (2.c) is image after difference.









(2.a) Previous frame image (2.b) Current fame image









(2.c) Image after difference
Fig. 2 Difference experiment
Noise is existing in the binarization image because of the
interference of noise. Morphologic method is introduced to
wipe off noise. Center of mass of movement object is
confirmed at last.
B. Morphologic
A little noise appears in target because of the interference
of noise. Noise need to be removed in order that object
positioning can be well done. The method of eroding and
expanding on binary image in morphologic is used to
accomplish the task.



Expanding: { } X B E z
z
|
Eroding: { } X B E z
z
|
Where X is binary image, B is grid of view picture, Bz is
a image which is obtained by moving structure element B
from center(which is red) to position z.
Experiment on morphologic is presented in Fig.3. Fig.3 is
the denoised image.









Fig. 3 Image after denosing Fig. 4 Centorid calculating image
C. Confirming Centroid
Centroid of movement region is confirmed by moment
method in the paper. Formula is shown as follows:
The zero
th
moment is:
00 ( , )
x y
M I x y =

(3)
The first moment for x and y:
10 ( , )
x y
M xI x y =

(4)

01 ( , )
x y
M yI x y =

(5)

The mean search window location is in (6)
10 01
00 00
, c c
M M
x y
M M
= = (6)
Experiment about centroid computing can be seen in Fig.
4. The position of centroid is marked by the red cross point.
Coordinate of centroid is (130, 198) in the experiment. Long
shaft and short shaft of ellipse is ascertained according to two
th

moment of binarization image. Long shaft and short shaft is
reduced by percent 20 because variational region of
binarization image is relatively larger.
III. CAMSHIFT ALGORITHM
Recently, tracking algorithm based on Camshift
algorithm is more and more noticed by means of its favorable
performance in reality and robust. Camshift algorithm is
widely used in face tracking about apperceiving user interface
now. Camshift algorithm makes use of color information in
region to track object and adopts a non-parameter technique
and searches movement target by clustering method
[7]
.
To be brief, Camshift algorithm applies target's hue
character to find the size and location of movement object in
video image. In the next frame video image, search window is
initialized by current position and size of movement object.
The process is repeated to realize continuous object tracking.
Initial value of search window is set as current position and
size before search.
52
Search window searches near the area that movement
object perhaps appears, so a mass of search time may be saved
and Camshift algorithm has favourable reality. At the same
time, Camshift algorithm finds movement object by colour
matching. Color information has no much change during
movement object moving, so it also has favorable robustness.

A. Probability Distributing Chart
Transformation formula between RGB space and HSV
space is shown in (7a)-(7c):

( ) ( )
( ) ( )
( ) ( )
max( , , )
3 max , , min , ,
max( , , )
2
3 max , , min , ,
max( , , )
4
3 max , , min , ,
b r
if g r g b
r g b r g b
if b r g b b r
H
r g b r g b
if r r g b
b r
r g b r g b

=
= +

+




(7a)

( ) ( )
( )
max , , min , ,
max , ,
r g b r g b
S
r g b

= (7b)

max( , , ) V r g b = (7c)
Where, H represents hue heft, S represents hue saturation
heft, V represents brightness heft, r, g, b respectively
represents red and green and blue heft.
For each image, the method of calculating hue histogram
is
[8]
: the total of image's pixel is given as n. Histogram's hue
level is m. We also define c(x
i
*
) is histogram bin index
associated to the pixel No.x
i
, where i is from 1 to n. The value
q
u
of unweighted histogram is computed in (8), where u is
from 1 to m:

( )
1
n
u i
i
q c x u
=
=

(8)

Where, is unit impulse transfer function.
The process of histogram normalization is shown in (9):

( )
1,...,
255
min , 255
max
u u
u m
p q
q
=


=



(9)

According to (9), value of the histogram and each chroma
level is normalized from [0, max(q)] to the new range [0, 255].
Then back-projection is used to associate the pixel values in
the image with the value of the corresponding histogram bin.
Namely, pixels with the highest probability of being in the
histogram will be mapped as visible intensities in the image.
So color probability distribution image can be seen in Fig. 5.
Where Fig. (5.a) is Original image, Fig. (5.b) is H heft image,
Fig. (5.c) is color histogram of refered area, Fig. (5.d) is
probability distribution image.







(5.a) Original image (5.b) H heft image






(5.c) Hue histogram image referred (5.d) probability distribution image
Fig. 5 Color probability distribution image
B. Camshift Tracking Algorithm
Camshift tracking process is shown in the following:
1) Regions that target is tracked are confirmed by difference
method in frame.
2) Calculating the zero
th
moment and first moment
The zero
th


moment is calculated in (10)

00 ( , )
x y
Z I x y =

(10)

The first moment for x and y is computed in (11a) and
(11b)
10 ( , )
x y
Z xI x y =

(11a)

01 ( , )
x y
Z yI x y =

(11b)

Where, I(x, y) is pixel value at point(x, y). Change scope
for x and y is the range of search window.
3) Compute the centroid( ) ,
c c
x y of search window location
in (12).
Z
01
,
Z
00
Z
10
= =
Z
00
c c x y (12)

4) Reinstall the size of search window S as a color probability
distribution function the region of search window.
5) Repeat step 2), 3), 4) to converge(the change of centroid is
smaller than given threshold)
Zero
th
moment describes distribution area of target in
53
image. Color probability distribution image is discrete grey
image that maximum value is 255. Therefore, relation in size
S and 255 for search window is written in (13):

S = 2 Z /256
00
(13)

Considering symmetry, S is selected as odd number
which approaches result. Long axis and short axis and
orientation angle of tracking object is calculated by computing
second moments. Second moments for x and y is defined in
(14a) and (14b) and (14c).

2
20 ( , )
x y
Z x I x y =

(14a)

2
02 ( , )
x y
Z y I x y =

(14b)

11 ( , )
x y
Z xyI x y =

(14c)

The orientation angle of the object Long axis is
confirmed in (15):
11
00
2 2
20 02
00 00
2
1
arctan
2
c c
c c
Z
x y
Z
z z
x y
Z Z

=






(15)

Supposed parameter a and b and c in (16):
2 20
00
11
00
2 02
00
c
c c
c
z
a x
Z
Z
b x y
Z
z
c y
Z
=
=
=
(16)

The distance of Long axis and short axis from the image
is given in (17a) and (17b):
( ) ( )
2
2
2
a c b a c
l
+ + +
= (17a)

( ) ( )
2
2
2
a c b a c
w
+ +
= (17b)

IV. EXPERIMENT
We validate the validity of the algorithm by tracking
movement hand. Result is shown in Fig. 6 and Fig. 7. Fig. 6 is
the result of tracking algorithm based on Camshift algorithm
combinating with the difference in frame. Fig. 7 is the result of
conventional tracking algorithm.


Fig. 6.1 Fig. 6.2


Fig. 6.3 Fig. 6.4
Fig. 6 Tracking algorithm based on Camshift algorithm combinating with
the difference in frame


Fig. 7.1 Fig. 7.2


Fig. 7.3 Fig. 7.4
Fig. 7 Conventional tracking algorithm
We prove the validity of the algorithm by upper
experiment. The algorithm can quickly and exactly track
movement object. It doesn't lose target information and
tracking region is also convergence.
V. CONCLUSION
We proposal object tracking algorithm based on Camshift
algorithm combinating with the difference in frame.
Movement object is accurately positioned by difference in
frame, which overcomes shortcoming of artificial orientation
for traditional Camshift. Approximate sport scope of object is
confirmed by frame difference during tracking. The algorithm
54
also overcomes radiation's shortcoming during tracking. But
there are some aspects to need to be improved in the future,
such as target moving too quick and tracking information
losing, etc.
REFERENCES
[1] Pan feng, Wang xuanyin, Xiang Guishian, Liang dongtai.
A kind of new tracking algorithm of movement object [J].
optics engineering, 2005,32(1):43-46,70
[2] P.Nordlund, T.Uhlin, Closing the loop: detection and
pursuit of a moving object by a moving observer [J]. Image
and Vision Computing. 1996, 14: 265~275.
[3] Clarke J C, Zisserman A. Detection and tracking of
independent motion [J]. Image and Vision Computing.
1996, 14: 565~572.
[4] St.Edmunds College. Active contour models for object
tracking [J]. Computer Science Tripos, Part II, 2003,(5)
[5] Hogg D, Model-based vision: A program to see a walking
person [J]. Image and Vision Computing. 1995, 1: 5~20.
[6] Rowe S, Blake A. Statistical mosaics for tracking [J].
Image and Vision Computing. 1995, 14: 549~564.
[7] Harrington S. Computer graphics, a programming
approach [M]. New York, USA: McGraw Hill Company,
1987: 376-388P.
[8] Allen John G, Xu Richard Y D, J in Jesse S. Object
tracking using Camshift algorithm and multiple quantized
feature spaces [A]. In: Pan-Sydney Area Workshop on
Visual Information Processing VIP2003[C], Sydney,
Australia, 2003: 1-5P.

55