TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.1, Jan
uary 20
14
, pp. 734 ~ 7
4
0
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i1.3688
734
Re
cei
v
ed
Jun
e
27, 2013; Revi
sed Aug
u
st 12, 2013; Accepted Aug
u
s
t 25, 2013
Target Trackin
g
Feature Selection Algorithm Based on
Adaboost
Chen Yi
Sichu
an Un
iver
sit
y
of Arts and
Science, Pin
g
c
han
g, Sichu
a
n
, PR Chin
a, 6350
00
e-mail: 1
685
65
0@q
q
.com
A
b
st
r
a
ct
W
i
th the d
e
v
e
lo
p
m
ent
of i
m
a
ge
process
i
ng
tec
h
n
o
lo
g
y
and
po
pul
a
r
i
z
at
io
n of co
mp
uter
techno
lo
gy, intelli
ge
nt mac
h
in
e vision tec
h
n
o
lo
gy has a
w
i
de ran
ge of ap
plicati
on i
n
the
med
i
cal,
mil
i
ta
ry,
ind
u
strial and other
fie
l
ds.
T
a
rget
tracki
ng
feature se
lecti
on a
l
gor
ith
m
is
one
of resear
ch focuses
in
th
e
mac
h
i
ne i
n
tell
i
gent vis
i
on tec
hno
logy. T
h
ere
f
ore, to
desi
g
n
the target trac
king fe
ature se
lectio
n al
gorit
h
m
w
i
th hig
h
accur
a
cy a
n
d
g
ood
stability
is
extr
emely
nec
e
ssa
ry. This p
a
p
e
r
prese
n
ts a
targ
et trackin
g
fe
atur
e
selecti
on
al
go
rithm b
a
sed
on A
d
a
boost.
It incl
udes
Adab
oost
alg
o
rith
m’
s
pri
n
ci
ple
an
d A
dab
oost
alg
o
rith
m'
s ap
p
licatio
n i
n
vid
e
o
obj
ect trackin
g
. Exper
i
m
enta
l
results sh
ow
that
the pr
opos
ed al
gor
ith
m
h
a
s
the character
i
s
t
ics of real
-time
,
accuracy and
stability.
Ke
y
w
ords
: int
e
lli
ge
nt mac
h
i
n
e visio
n
,
target tracking,
al
gorit
hm,
Ad
abo
ost
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Vision i
s
the
most impo
rt
ant way for
human to p
e
r
ce
pt the foreign obj
ect
s
to ge
t
informatio
n. Giving machi
ne vision fun
c
tions can ma
ke life and p
r
odu
ction mo
re conve
n
ient.
Th
e
human
visu
a
l
system
ca
n qui
ckly
capture
an
i
m
age, o
b
je
ct recognitio
n
,
and ta
ke
th
e
corre
s
p
ondin
g
actio
n
s i
n
accordan
ce
with the info
rmation
obtai
ned [1]. The
machine vi
sion
make
s u
s
e o
f
input device
s
su
ch a
s
ca
mera
s in
stea
d of eyes to get image
s. Use a co
mpu
t
er
instea
d of
the
hum
an
brain
for ima
g
e
proce
s
sing
in
o
r
de
r to
a
c
com
p
lish
a
simil
a
r functio
n
of th
e
human visual
system [2]. In rec
ent years, with the improvem
ent
of compute
r
pe
rforma
nce an
d
the falling price of cam
e
ras, intelligent
machin
e vision tech
nolo
g
y is more
widely used
in
military, medi
cal, indu
stri
al and ele
c
tro
n
i
cs m
anuf
a
c
t
u
ring i
ndu
stri
es. That intel
ligent ma
chin
e
vision techn
o
logy repl
a
c
e
s
the hu
man eye to
distingui
sh
has b
e
come
a trend of
the
developm
ent
of prod
uctio
n
, and targ
et tracking al
gor
it
hms i
s
an im
portant p
a
rt i
n
machine vi
sion
techn
o
logy. The tradition
al target tra
cki
n
g
algorit
h
m
mainly inclu
d
e
s
continu
o
u
s
frame differe
nce
method
and
backg
rou
nd
removal meth
od [3]. The m
o
tion region
detecte
d by succe
ssive f
r
a
m
e
differen
c
e
m
e
thod
relate
s to velo
city of
the ta
rge
t.
Whe
n
the
target spee
d i
s
very sl
ow, it
may
not be abl
e to dete
c
t the target m
o
tion;
backg
rou
nd
removal m
e
th
od is
sen
s
itive to ch
ange
s in
illumination a
nd ba
ckgro
u
n
d
, and in the ca
se of inte
rf
eren
ce, it ma
y fail to detect the target [4-5].
This p
ape
r p
r
ese
n
ts a ta
rg
et tracking f
e
ature
sele
ctio
n algo
rithm b
a
se
d on Ad
a
boo
st to
extract the ch
ara
c
teri
stics
of the target and in t
he ne
xt frame to find a pictu
r
e t
hat match
e
s t
he
cha
r
a
c
teri
stics of the targ
et in orde
r to deter
mi
ne
the target lo
cation, and ult
i
mately achie
v
e
target trac
k
i
ng.
2. The O
v
erall Design
The targ
et tracki
ng selectio
n algorithm fr
amework pro
posed ba
sed
on Adabo
ost
can b
e
divided into the followi
ng three p
a
rt
s:
(1) Indi
cate
s
the target ob
ject usi
ng
Ha
rr featu
r
e
s
, u
s
e
s
and Inte
gral ma
p me
thod to
accele
rate th
e spe
ed of op
eration;
(2) Usin
g Ad
aboo
st al
gori
t
hm to
sele
ct
the
m
o
st
re
pre
s
entative
of the ta
rget
obje
c
t
recta
ngle fe
a
t
ures
(wea
k
cla
ssifie
r
s), b
a
se
d on t
he
detectio
n
rat
e
of the wei
ght given to
the
corre
s
p
ondin
g
value of the wea
k
cla
s
sifier
s a
r
e
con
s
tructed a
s
a st
rong cl
assifier;
(3) The
nu
mb
er of
se
rie
s
to
form
a
stron
g
cl
as
sif
i
er
c
a
sc
ade
st
r
u
ct
u
r
e
ca
sc
ade
cl
as
sif
i
er
ca
scade stru
cture can
effe
ctively
improv
e
the detectio
n
rate of cla
s
sificatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 734 – 7
4
0
735
Adaboo
st al
gorithm i
s
d
i
vided into t
he follo
wing
part
s
. First, cal
c
ulate
th
e targ
et
cha
r
a
c
teri
stic para
m
eters,
and the
n
cal
c
ulate
ea
ch
classifier th
re
shold p
a
ra
met
e
r a
c
cording
to
the param
ete
r
s, and
conv
ert the target
feature in
to the corre
s
po
nding wea
k
cla
ssifie
r
s, train
sampl
e
set to
elect th
e b
e
st wea
k
cla
ssif
i
er a
nd
cal
c
ul
ate its
weig
ht. After T time
s to g
ene
rate
a
stron
g
cla
s
sifier ca
scad
e to achieve th
e pur
p
o
se of tracki
ng. Ca
scade me
ntio
ned above i
s
to
detect the im
age wi
ndo
w i
t
eration
step
by step thr
o
u
gh a wea
k
cl
assifier, de
ci
de targ
et are
a
,
filter the targ
et area to test the wea
k
target im
a
ge
as the targ
et, then the output is 1; if it
is
determi
ned t
o
be n
on-ta
rget, the outp
u
t is 0. The
ov
erall d
e
si
g
n
of the alg
o
r
ithm flow
ch
art is
s
h
ow
n
in
F
i
gu
r
e
1
.
Figure 1. Overall De
sig
n
3. Target Tr
acking Fe
atu
r
e Selection
Algorithm Based on
Ada
boost
3.1. Adaboo
st Algori
t
hm
Principle
Adaboo
st alg
o
rithm'
s ba
si
c pri
n
ci
ple i
s
that
integrat
e a wea
k
lea
r
ning
algo
rith
m in the
colle
ction
by several iterations
i
n
to a
strong le
arni
ng
algorith
m
. Ad
aboo
st alg
o
rit
h
m is
agai
nst
a
set to
train
weak cl
assifie
r
s, an
d
cla
s
sify the re
sultin
g casca
d
e
of
wea
k
to
gethe
r to f
o
rm
a
strong
cla
ssifie
r
. Initial wei
ghts of
the sa
mple
s
are th
e
same
. In this
sam
p
le di
strib
u
tio
n
, train
a ba
sic
cla
ssif
i
e
r
h
1
(
x
),
inc
r
ea
sing
h1
(x
) mi
s
c
la
ssif
i
e
d
sampl
e
w
e
ight
s;
re
duc
e t
he
co
rr
ect
cla
ssif
i
cat
i
on
of the sample
weight
s, so i
t
can highli
g
h
t
the
error
sa
mples a
nd ge
t a new sam
p
le distri
butio
n.
Then a
c
cordi
ng to the prin
ciple
s
that the less the e
r
ror of the weig
ht is, the grea
ter the weig
ht is
and the
more
the erro
r is,
the less the
weig
ht is
, ent
itle h1 (x) to
a ne
w weight
. Unde
r the
n
e
w
sampl
e
di
stri
bution, ba
si
c
cla
ssifie
r
is t
r
ained to
ba
si
c cl
assifier
h2
(x) an
d its
weight, and
so
on
,
throug
h the
T
cycl
es, it
get
s a
ba
si
c
cla
s
sificati
o
n
of T
and
its
weig
ht [2]. Finally
add
up th
ese
T
basi
c
cl
assifi
er at a ce
rtain
weight to ob
tain the final stron
g
cla
s
sifier.
AdaBoo
st alg
o
rithm is d
e
scrib
ed a
s
follows:
Assu
me that X is the sam
p
le of spa
c
e,
Y repre
s
ent
s a collectio
n of sample
cat
egori
e
s
identified. He
re set Y = {-1, +1}
so tha
t
S = {(xj,
yj) | j = 1,2, ...,
m} is the
sa
mple traini
ng
set,
whe
r
e x
j
∈∈
X,
yj
Y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Targ
et Tra
cki
ng Featu
r
e Selectio
n Algorithm
Based on Adaboo
st (Che
n Yi)
736
A. initializes the weight
s of n sample
s to make Dt uniformly distribute:
n
j
D
t
1
)
(
.
rep
r
e
s
ent
s the weig
ht of the sampl
e
(xj, yj) assi
gne
d in t iteration;
B. Let T
represe
n
t the number of iterat
i
ons. Fo
r each t = 1 ... T,
based on the
sample
distrib
u
tion Dt, generate
sample to form the colle
ct
ion St. Train the cla
s
sifier
ht on the set
St.
Use the
cla
s
sifier ht to cl
assify all sa
mpl
e
s of
the
set S. Get cla
ssifi
er in thi
s
roun
d and
minimu
m
error rate Ej [3];
|
)
(
|
1
j
j
n
j
t
j
y
x
h
E
C. Calculate
the weig
hts:
)
/
1
log(
);
1
/(
j
j
j
j
a
E
E
D. Update th
e weig
hts:
)
)
(
(
)
(
1
T
t
t
t
x
h
sign
x
H
Adaboo
st alg
o
rithm flow
chart is
sho
w
n
in Figure 2.
Figure 2. Adaboo
st Agorith
m
Flowcha
r
t
In
itial weigh
t
n
D
i
t
1
,
t<T?
Select h
(x) t
o
make the sm
allest error rate
E
j
|
)
(
|
1
j
n
j
j
j
j
y
x
h
E
C
a
l
c
ul
at
e wei
ght
s
)
/
1
log(
);
1
/(
j
j
j
j
a
E
E
Adju
st th
e wei
g
h
t
Yes
Out
put
f
unct
i
o
n
)
)
(
(
)
(
1
T
t
t
t
x
h
sign
x
H
No
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 734 – 7
4
0
737
3.2. Target F
eatur
e Extr
a
c
tion
The targ
et tracking
sele
cti
on algo
rithm
bas
ed on Ad
aboo
st pro
p
o
s
ed in the p
a
per is in
the light of gray distributio
n cha
r
a
c
teri
st
ics of the targ
et object to d
e
tect and tra
ck. Here we
use
Harr feature
s
as
sele
ctio
n f
eature. Harr featu
r
e i
s
ba
sed o
n
the integral i
m
age featu
r
es.
Comp
ared
wi
th the traditio
nal pixel
-
ba
sed
com
put
ing
eige
nvalue
s,
be
cau
s
e
gray-scale
ima
ges
posse
ss
the cha
r
a
c
teri
stic
og
a sin
g
le color, small
co
mputation, it i
s
ea
sie
r
to
use integ
r
al im
a
ge
cha
r
a
c
teri
stic to extract pa
ramete
r [6]. Target
tra
c
king
sele
ction alg
o
rithm ba
se
d on Adabo
ost
is
adopte
d
to
d
e
tect a
nd t
r
a
c
k the
target o
b
ject,
who
s
e
high deg
ree
of
cla
s
sificati
on re
sults
im
pact
target dete
c
ti
on and tra
c
ki
ng sig
n
ifica
n
tly. Harr featu
r
es is a
b
le to accurately distingui
sh o
b
j
e
cts
and n
on-obje
c
ts. Fo
r exam
ple, Ha
ar feat
ured
as t
he f
a
ce
edg
e feat
ure
s
. Fa
ce
Haar featu
r
e
s
are
divided i
n
to
three
catego
ries:
linea
r f
eature
s
,
edg
e featu
r
e
s
a
nd
cente
r
su
rro
und
features.
Linea
r ch
ara
c
teri
stic
can
be accu
rately
represent
the linear di
re
ction of the image informati
on;
edge featu
r
e
can a
c
curatel
y
show the e
dge
s of
the image inform
ation; cente
r
surro
und feat
ure
can a
c
curatel
y
show the
central po
rtion
of the
image information [
4
]. Feature te
mplate co
nsi
s
ts
o
f
b
l
ac
k a
nd w
h
ite re
c
t
ang
le
s
.
Acc
o
rd
in
g
to t
he
ch
ara
c
teri
stics
of the im
age
obtain
ed, th
ese
feature
s
ca
n be expre
s
sed
as:
)
(
Re
j
T
j
j
r
ctSum
w
feature
Here T is the numbe
r of the rectan
g
u
lar feat
urej, wj is the weight of the recta
ngle,
Re
ctSum (rj) is the
sum
of the pixel v
a
lue
s
surrou
nded
by a
re
ctangl
e rj.
Howeve
r, u
s
e
the
above al
go
rithm ba
se
d o
n
pixel recta
ngle featu
r
e
too mu
ch
wil
l
gre
a
tly red
u
ce
the train
i
ng
spe
ed. Thi
s
p
aper u
s
e
s
the
integral im
ag
e metho
d
, wh
ose
main
ide
a
is to avoi
d
cal
c
ulate
d
p
r
i
o
r
to the
pixel re
gion
befo
r
e
d
ouble
countin
g an
d
save
pi
xels in
the
re
ctang
ular regi
on
startin
g
fro
m
the imag
e p
o
i
n
t to the
cu
rrent j a
s
el
em
ents
of the
a
r
ray. Wh
en
a region
of pixel
s
i
s
to
cal
c
ula
t
e,
it can search
dire
ctly
elem
ents sto
r
e
d
in
the array to spe
ed
up
the comp
utation. Specific
fo
rm
ula
is
as
follows
:
y
y
x
x
y
x
i
y
x
ii
'
,
'
)
'
,
'
(
)
,
(
Here i i
s
th
e
origin
al ima
g
e
, (x, y) i
s
the
image
pixel,
ii (x, y) i i
s
th
e sum of
all t
he pixel
s
in the uppe
r left point.
3.3. Design
of the
Wea
k
Classifie
r
s
Acco
rdi
ng to
Adaboo
st alg
o
rithm the
ab
ove m
ention
e
d
, again
s
t the
origin
al weig
hts with
the sa
me
sa
mple, it train
s
a
se
rie
s
of
wea
k
cl
a
s
sifiers
thro
ugh
cha
ngin
g
its sampl
e
weig
hts.
Wea
k
cla
ssifi
er is i
n
itially prop
osed
wh
en in 19
84
Valian lea
r
ne
d theory in
PAC. Dete
ction
accuracy is b
e
tter than ra
ndom gu
essi
ng, but not obvious wea
k
is calle
d we
a
k
learning [7]. If
detectio
n
a
ccura
cy was si
gnifica
ntly be
tter than
r
a
n
dom g
u
e
ssi
n
g
, we
call
it strong l
earnin
g
. If
the data i
s
sufficient, we
a
k
cl
assifie
r
cascad
e
form
s a
stro
ng
cl
assifier th
rou
gh trai
ning
a
nd
learni
ng. Be
cause of the
di
versity of targ
et dete
c
tion
chara
c
te
risti
c
s, we
can
dete
c
t catego
rie
s
to
determi
ne the
requi
red b
a
si
c cla
s
sifier h (x
). Here i
s
cal
c
ulate
d
as foll
ows:
otherwise
x
f
x
h
j
j
j
0
)
(
1
)
(
Here hj
rep
r
e
s
ent
s the
ch
ara
c
teri
stic value
of cl
assi
fier. The
out
put is 1
if th
e target
detectio
n
is
correct, othe
rwise, the out
put 0; j is
the
threshold val
ue of the ch
a
r
acte
ri
stic val
u
e
and fj (x) indi
cate
s eige
nvalue fun
c
tion.
3.4. Final Strong Clas
sifier Con
s
tru
c
tor
The mo
st important pa
rt of target tracking
featu
r
e
algorith
m
is the target de
tection
accuracy
and
detectio
n
sp
eed
sho
u
ld b
e
fast to m
a
ke it real
-time
in order to a
c
hieve t
r
a
cki
ng
results. T
h
e
we
ak cl
assi
fier d
e
sig
n
menti
one
d a
bove d
o
e
s
n
o
t meet th
e
re
quirement
s of
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Targ
et Tra
cki
ng Featu
r
e Selectio
n Algorithm
Based on Adaboo
st (Che
n Yi)
738
accuracy, fal
s
e
alarm a
n
d
false
dete
c
ti
on's rate i
s
st
ill very hi
gh.
Adaboo
st al
g
o
rithm i
s
use
d
to
ca
scade
up
and o
p
timize
the we
ak
cla
ssifie
r
s to form
a strong
cl
assifier. Th
i
s
stron
g
classif
i
er
can
qui
ckly d
e
tect the
obje
c
t to be te
ste
d
. In Adab
o
o
s
t lea
r
ning
proce
s
s, usuall
y
every po
sitive
detectio
n
thresh
old rate i
s
98.5% [8].
We
can
ap
prop
riately in
cre
a
se the p
o
sitive exam
ples'
threshold,
the
r
eby fal
s
e
det
ection
of n
o
n
-
targ
et will
al
so i
n
crea
se,
but do
not
worry. T
he
stro
ng
cla
ssifie
r
cl
assifier formed
after the co
m
b
inati
on of weak
cla
ssifie
r
s thro
ugh
stu
d
y can b
e
used
as
the firs
t few hierarc
h
ical leve
ls, the
n
more
preci
s
el
y filter non
-ta
r
get by
mean
s of th
e
stron
g
cla
ssifie
r
. As
the first fe
w l
e
vels filter th
e majo
rity of non-ta
rg
et, the non
-target
past the fin
a
l
level
of strong
cla
ssifie
r
will be
greatly redu
ced.
As a re
sult, the detection accu
ra
cy and detecti
on
spe
ed is g
r
ea
tly increa
sed.
Specific
st
ro
ng cla
s
sifier a
l
gorithm i
s
as follows:
Given the trai
ning sampl
e
s (x1, y1) ... (xt, yt),
where xj denotes the
sampl
e
j. When yj =
1, it is the positive sampl
e
(targ
e
t). If yj
=
0, it indicat
e
s a ne
gative
sample
(no
n
-target).
a. Initialize the weights:
T
j
T
x
w
j
j
...
1
,
1
)
(
;
b. train a wea
k
cla
s
sifier hj
(x) for ea
ch feature;
c
.
Selec
t
hj (x
) to mak
e
the
s
m
alles
t
error rate Ej;
|
)
(
|
1
i
i
T
j
j
j
j
y
x
h
w
E
d. updat
e th
e sample
di
stributio
n a
n
d
ej
=
0
m
ean
s that th
e sample
is co
rrectly
cla
ssifie
d
a
nd ej
=
1 me
an
s t
hat cl
assification i
s
in
corre
c
t,
)
1
/(
j
j
E
E
,
t
e
j
j
t
j
t
Z
x
D
x
D
j
/
)
(
)
(
1
1
,
where Zt is
us
ed to meet the
1
)
(
1
1
i
T
j
t
x
D
;
f. Strong classifier con
s
tru
c
tor [5];
))
(
(
)
(
1
x
h
a
sign
x
H
j
T
j
j
3.5. Strong Classifie
r
Ca
scade Struc
t
ure
Ca
scade in
classifiers refe
rs
that the final strong
cl
assifi
ers are combi
ned by
several
wea
k
cl
assifi
ers a
nd opti
m
ized
com
p
onent
s. In t
he target feat
ure dete
c
tion
, target feature
waiting fo
r d
e
tecting p
a
ss sequ
entially
throug
h ea
ch cla
ssifie
r
, so that the first few level
s
of
wea
k
cla
ssifi
ers ca
n
filter most
of
the non-ta
rg
et
ch
ara
c
teri
stics [
10]. Ultimatel
y
, the detecti
on
area
whi
c
h can pa
ss all a
nd export i
s
the target
a
r
e
a
zon
e
. Stron
g
cla
ssifie
r
ca
scade
schem
atic
diagram is
sh
own in Fig
u
re
3.
Figure 3. Schematic Stro
ng Cla
s
sifier
Ca
scade
Thro
ugh thi
s
algorith
m
, ea
ch wea
k
cl
assifier
can eve
n
tually gene
rate a strong
cla
ssifie
r
throug
h T
ite
r
ation
s
.
Cal
c
ulate the
ima
ge xj th
rou
g
h
the
ope
rati
on of
hj
(x) t
o
obtai
n a
hi
gh
accuracy a
n
d
stability. The bigge
r the T
traini
ng time
s is, the mo
re paramete
r
s and judg
me
nt
involved in. The accu
ra
cy and sta
b
ility of h (x)will be i
m
prove
d
[9].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 734 – 7
4
0
739
4. The Exper
i
mental Res
u
lts
In order to tes
t
the effec
t
of target track
i
ng
algo
rith
m, use vi
sual
C
+ +6.0 si
mulation i
n
Wind
ows XP operating sy
stem. Obtain
the followin
g
re
sults,
sh
own in Fi
gu
re 4: observe
the
video se
ction
15, se
ction
25 and
se
cti
on 40 pi
cture
s
, the target
for the imag
e of the hum
an
face, a
nd
we
ca
n dete
c
t t
he p
o
sition
o
f
human
face
s in t
he tracking video
to
achi
eve a ta
rget
tracking. Thi
s
shows that the proposed
algorit
hm has real-time, accuracy and st
ability.
section 15
se
ction 25
se
ction 40
Figure 4. Simulation Results
5. Conclusi
on
In sum
m
ary,
the pap
er
pro
poses
a targ
et tr
ackin
g
fe
ature
re
cog
n
i
t
ion algo
rithm
s
ba
se
d
on Ada
boo
st. By extractin
g
the ta
rget f
eature, th
e in
tegral
gray v
a
lue of th
e resultin
g featu
r
e
image is give
n the appropriate weigh b
a
s
ed o
n
its bei
ng the ca
se o
f
ant detection rate to form
a
wea
k
cla
ssifi
er. After T ti
mes'
a
s
sociat
ion an
d
sub
-
o
p
timal level, it ultimately ge
nerate
s
a
strong
cla
ssifie
r
to a
c
hieve th
e ta
rget tra
c
king
results. Ove
r
all this
algo
rithm can b
e
d
i
vided into th
e
f
o
llowin
g
st
ep
s:
(1) Input
sa
mples,
acco
rding to
the
spe
c
ific
prop
erties of the
target
obje
c
t to dra
w
recta
ngle feat
ure
s
and
cal
c
ulate and g
e
t
a recta
ngul
ar
prototype feat
ure set;
(2) Enter the
feature
to d
e
termin
e the
th
res
hold
value
and
corre
s
p
ond fe
ature
s
to we
ak
cla
ssifie
r
s to
obtain set of wea
k
cl
assifi
ers;
(3) Enter we
ak
cla
s
sifiers. In the t
r
aini
ng rate limit,
sele
ct the
op
timal we
ak al
gorithm
cla
ssifie
r
s a
s
a stron
g
cla
s
sifier inp
u
t by means of Ad
aboo
st;
(4) Ente
r the
strong
cla
ssi
fier and opti
m
ize it ca
sca
de to con
s
titute stron
g
cl
assifier of
the final stag
e.
In additio
n
, throu
gh th
e u
pdate
of the
new we
a
k
cl
assifiers inte
grated
onli
n
e
and
the
existing wea
k
classifie
r
s'
weights, it improves
the a
b
ility that algorithm adapt
s to cha
r
a
c
teri
st
ic
cha
nge of illumination an
d other facto
r
s. In a la
rge
number of e
x
perime
n
ts o
n
real image,
it
comp
ares
with the origi
nal
algorithm
s.
And the
re
su
lt shows that
this algo
rith
m can n
o
t o
n
ly
respon
d bett
e
r to the ch
a
nge of t
he target feature in
the pre
s
en
ce
of interferin
g backg
rou
n
d
to
st
ably
t
r
a
c
k t
a
rget
s,
but
a
l
so it
d
e
s
c
ri
b
e
s t
h
e
targ
et size mo
re
accurately a
nd si
gnificant
ly
improve the tracking a
c
curacy of the alg
o
rithm.
Referen
ces
[1]
W
ang Sh
up
en
g. Vide
o targ
e
t
tracking a
l
g
o
r
ithm. Xi'
a
n:
X
i
'
an U
n
ivers
i
ty of Electron
ic
Scienc
e a
n
d
T
e
chno
logy.
2
009: 1-8.
[2]
Z
hao Ji
ang,
Xu Lu
’an. T
a
rget
Detectio
n B
a
sed
on a
d
a
b
oost al
gorithm
.
Artificial Intelligence
an
d
Reco
gniti
on T
e
chno
logy
. 2
004
; 30(4): 120-1
2
5
.
[3]
Liu T
L
, Che
n
H
T
. Real time
trackin
g
usi
ng tr
ust regi
on m
e
thods.
IEEE Tr
an s
on P
a
ttern Analysis
and
Machi
ne Intell
i
genc
e
. 200
4; 26(3): 397-
40
2.
[4]
W
ang Xu. T
a
rget detectio
n
system bas
ed o
n
Adab
oost alg
o
rithm. Xi'
a
n:
Xi'
an Univ
ersit
y
of Electroni
c
Scienc
e an
d T
e
chn
o
lo
gy.
201
2; 72: 162-
169.
[5]
Lin
deb
erg T
.
F
eature d
e
tect
i
on
w
i
th
auto
m
atic scal
e
se
lectio
n.
Interna
t
iona
l Jo
urna
l
of Co
mp
uter
Vision
.199
8; 3
0
(2): 79-1
16.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Targ
et Tra
cki
ng Featu
r
e Selectio
n Algorithm
Based on Adaboo
st (Che
n Yi)
740
[6]
Sortomme E, Venkata SS,
Mitra J. Microgrid pr
otecti
on
usin
g commun
i
catio
n
-assiste
d
dig
i
tal rel
a
ys
,
Know
led
ge-B
a
sed Syste
m
s.
201
0; 25(4): 27
89-2
796.
[7]
McArthur SDJ, Davidso
n
EM, Catterson
VM. Multi-age
nt s
y
stem
s for po
w
e
r
engi
neer
in
g
app
licati
ons—
P
art I: concepts, appr
oac
hes
and tech
nica
l
chall
eng
es.
IEEE Transactions on P
o
wer
System
. 20
07; 174
3-17
52.
[8]
Comaniciu D, Ramesh V, M
eer P. Real ti
me T racking of non-rigid objects using mean shift.
IEEE
Confer
ence
on
Computer Vis
i
on an
d Pattern
Recog
n
itio
n.
2
009: 14
2-1
49.
[9]
Hon
g
tao
Liu, H
ongfe
ng Y
un,
Hui C
h
e
n
, Z
h
a
o
y
u L
i
, Yu W
u
. Interest Exc
i
ta
tion Pro
p
a
gatio
n Mod
e
l fo
r
Information P
r
opa
gatio
n on
micro-blo
g
g
i
ng.
T
E
LKOMNIKA Indon
es
ian Jo
urna
l of Electrica
l
Engi
neer
in
g
. 2013; 11(
9): 489
4-49
02.
[10]
Xu, P
e
il
on
g. Stud
y o
n
Mov
i
ng Ob
jects b
y
Vi
de
o Mo
nit
o
rin
g
S
y
stem
of Reco
gn
ition
and T
r
acin
g
Scheme.
T
E
LK
OMNIKA Indon
esia
n Journ
a
l o
f
Electrical Eng
i
ne
erin
g
. 201
3; 11(9): 484
7-4
854.
Evaluation Warning : The document was created with Spire.PDF for Python.