TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 11, Novembe
r
2014, pp. 77
7
2
~ 777
7
DOI: 10.115
9
1
/telkomni
ka.
v
12i11.60
30
7772
Re
cei
v
ed Ma
rch 2
9
, 2014;
Re
vised June
20, 2014; Accepte
d
Jul
y
1
0
, 2014
Robust Multiple Ship Tracking in Inland Waterway
CCTV System
Fei Teng*
1
, Qing Liu
2
1
School of En
e
r
g
y
a
nd Po
w
e
r
Engi
neer
in
g, W
uhan U
n
iver
sit
y
of T
e
chnol
og
y, W
uhan, 4
300
70, Ch
ina
2
W
uhan Un
iver
sit
y
of T
e
chnol
og
y, W
uhan, 4
300
70, Ch
ina
*Corres
p
o
ndi
n
g
author, em
ail
:
change
erh
a
o
_
lov
e
@1
26.co
m
A
b
st
r
a
ct
In recent y
e
a
r
s, singl
e o
b
j
e
ct tracking
h
a
s be
en
exte
nsively
studi
e
d
an
d ac
hiev
ed
muc
h
deve
l
op
ment.
How
e
ver,
mu
lti
p
le
ob
jects tra
cking
is st
i
ll
an
issu
e th
at re
mains
to
be
ad
d
r
essed. Ge
ner
all
y
speak
ing, ex
isting
mu
ltipl
e
ob
jects tracking
meth
ods
e
m
pl
oy a man
ner
of simultan
eo
u
s
ly tracking e
a
c
h
obj
ect respecti
vely. In this p
a
per, w
e
dev
elo
p
a
mu
ltip
l
e
sh
ip tracki
ng a
l
g
o
rith
m b
a
sed
o
n
defor
mabl
e p
a
r
t
mo
de
l to acco
mp
lish
multip
le
ship tracki
ng i
n
inl
a
n
d
w
a
terw
ay CCTV (Closed-
Circu
it Televisi
on) a
u
to
mate
d
surveillance. Our meth
o
d
uti
l
i
z
e
s
HOG feat
ures to c
onstr
uct t
he a
p
p
ear
ance
mod
e
ls
of ships. T
h
en
by
taking ful
l
adv
antag
es of the spat
ia
l cons
trains betw
een
ships, w
e
can successful
ly
explor
e mut
u
al
relati
ons for mu
ltipl
e
shi
p
s, thus acco
mp
lishi
ng
mu
lt
ipl
e
ship tracki
n
g
in its true sense. More
o
v
er,
structured le
ar
nin
g
metho
d
is
used to lear
n how
to
update
the mode
l par
a
m
eters.
Nu
mer
ous exp
e
ri
me
n
t
al
results o
n
ch
alle
ng
ing
inl
a
n
d
w
a
terw
ay CCT
V
video s
e
que
nces d
e
m
onstrate that
our
meth
od c
a
n
effectively an
d
accurate
ly perf
o
rm
robust multiple ship tracking.
Ke
y
w
ords
:
multiple s
h
ip track
i
ng, CCTV system
, De
f
o
rmable part model, m
u
tual relations
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
In recent yea
r
s, CCTV a
u
tomated
surv
eilla
n
c
e sy
ste
m
has m
ade
great
cont
rib
u
tion to
inland
wate
rway man
age
ment. Com
p
ared
with a
r
t
i
ficial monito
ring, CCTV system ma
ke
s it
possibl
e to
keep full 24 hours’
surv
eillance and
cruise forensi
c
s proc
ess records. I
n
the
appli
c
ation of
electro
n
ic
cruise in inla
nd
waterway
, CCTV system
plays a
s
an e
y
e’s role in
cruise
check, thus guaranteeing ship tr
ajectory
tracking, illegal disposal
, safety forewarning and so on.
Obje
ct tracki
ng is a fund
amental chall
enge in
com
puter visio
n
with appli
c
ati
ons in a
wide
rang
e o
f
domain
s
su
ch a
s
hum
an
-co
m
pute
r
int
e
ra
ction, mov
i
ng re
cog
n
itio
n, video index
and
so
on.
By now, obj
e
c
t tra
cki
ng h
a
s a
c
hi
ev
ed
rema
rkable
p
r
og
re
ss. In
g
eneral, tra
cki
ng
method
s fall i
n
to two
categ
o
rie
s
: ge
nerative trac
kin
g
method and
discrimi
native
trackin
g
met
hod
[1]. Generative tracking
method m
o
d
e
ls the
obje
c
t of intere
st by just de
scribing the
obj
ect
appe
ara
n
ce [
2
-6].
Discri
minant tracke
r
model
s b
o
th t
he o
b
je
ct of i
n
tere
st a
nd t
he b
a
ckg
r
ou
n
d
. It
focu
se
s on finding a d
e
ci
sion bou
nda
ry to sepa
rate the obje
c
t fro
m
the backg
round [7
-10]. Du
e
to the co
nsi
d
ered
ba
ckgro
und info
rmati
on, discri
min
a
tive tracking
method al
ways outp
e
rfo
r
ms
the gene
rative one
s.
Discri
minativ
e tra
cki
ng m
e
thod a
c
compl
i
she
s
tr
ackin
g
by dete
c
tion.
In the field
of obje
c
t
detectio
n
, many prog
re
sses an
d achie
v
ements
hav
e been d
e
m
onstrated. SVM classifier with
the HOG feat
ure is a typical detecto
r for detectin
g
specifi
c
huma
n
[11], face [12] and so o
n
. In
recent yea
r
s,
anothe
r det
ector
ba
sed
on the defo
r
mable p
a
rt
model [13] a
l
so d
r
a
w
s
pu
blic
attention du
e
to its pe
rformable
dete
c
tion [14
]. It is kno
w
n
as
one of th
e b
e
st dete
c
to
rs.
Re
cently, Lu
Zhang
et al. [15] pro
p
o
s
e
a structu
r
e p
r
ese
r
ving o
b
je
ct tra
cki
ng m
e
thod to tracking
arbitrary obj
e
c
ts b
a
se
d on
a singl
e (b
o
undin
g
box
)
annotatio
n of object of int
e
re
st in the first
frame. It demonst
r
ate
s
favorabl
e perf
o
rma
n
ce by
combi
n
ing th
e su
ccess of
the Dalal-Tri
ggle
detecto
r a
nd
the defo
r
mab
l
e pa
rt mod
e
l
.
Genqu
an
Duan et
al. [16
]
introdu
ce
a
mutual relatio
n
model to grou
p multiple obj
ects tog
e
ther.
Acco
rdi
ng to
the online
algorith
m
the
y
pr
opo
se
d, they can track multiple
unseen
obje
c
ts. Moti
vated by the
method
in [
16], we
prop
ose
a m
u
ltipl
e
ship tracki
ng meth
od t
o
be
applie
d into CCTV su
rveilla
nce vide
o se
quen
ce
s in inl
and waterwa
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Multip
le Ship Tra
cki
ng in Inland
Wate
rwa
y
CCTV System
(Fei Te
ng)
7773
The re
maind
e
r of this pa
per i
s
org
ani
zed a
s
follo
ws: In Sectio
n 2, we de
scrib
e
ou
r
tracke
r in
det
ails. In Se
ctio
n 3, we pe
rfo
r
m expe
rime
nts to
sho
w
how
ou
r multi
p
le ship tracker
works
in inland waterway CCTV videos
.
Fina
lly, the concl
u
si
on is g
i
ven in Sectio
n 4.
2. Proposed
Track
er
Gene
rally, a
tracking
syst
em
comp
ri
se
s th
ree
mai
n
co
mpo
nent
s [17]: an
ap
p
eara
n
ce
model, lo
cati
on mo
del a
n
d
a
sea
r
ch st
rategy. In the
pro
p
o
s
ed m
u
ltiple ship tracker,
we
m
odel
the appe
ara
n
c
e of ou
r shi
p
s of interest
with the Dal
a
l-Tri
g
g
s
det
ector [1
1]. Other tha
n
tracking
multiple
ship
respe
c
tively, we
utilize
the
mutual
rel
a
tion m
odel
s [1
6] ba
sed
on
deform
able
p
a
rt
model to d
e
s
cribe
a con
f
iguration
of ship
s’
st
ate. Then, an
onl
ine structu
r
e
d
SVM [18] is
adopte
d
to learn
and i
d
e
n
tify the conf
iguratio
ns
of ship
s. As fo
r se
arch
stra
tegy, a slidin
g-
wind
ow exha
ustive se
arch
is an advisab
le choi
ce.
2.1. Appea
r
a
n
ce Model
Feature represe
n
tation i
s
critical to th
e
perfo
r
m
an
ce o
f
a
tr
ack
e
r
.
O
u
r tr
ac
ke
r
u
s
es
th
e
HOG
feature
s
to
rep
r
e
s
e
n
t ea
ch
ship
of inte
rest a
nd SVM to o
b
tain thei
r ap
pearan
ce m
o
del
r
e
spec
tively.
HOG featu
r
e
s
mea
s
ure the magnitud
e
and the
orie
nt of the image
gradi
ent. Actually, we
first assum
e
to divide the image pat
ch i
n
to many
8*8
pixel cells wi
thout overlap
p
ing. Then
we
cal
c
ulate th
e
gradi
ent in e
a
c
h
cell u
s
in
g the sim
p
le
[-1, 0, 1] masks for ho
ri
zontal
dire
ction a
nd
[-
1, 0, 1]
’
for
vertical
direct
ion. Each pix
e
l calculate
s
a
weig
hted
vote for a
n
edge
ori
entat
ion
histog
ram
ch
annel ba
se
d on the orie
ntation of
the gradi
ent ele
m
ent cente
r
e
d
on it, and the
votes a
r
e a
c
cumul
a
ted i
n
to (u
nsi
gne
d) orie
ntati
on b
i
ns ove
r
cell
s. Thirdly, ea
ch two
adja
c
e
n
t
cell
s
comp
rise a bl
ock. Th
e hi
stogram
of gra
d
ient
of those blo
c
ks wo
uld b
e
n
o
rmalize
d
with
the
L2-n
o
rm
re
sp
ectively. Finally, by simply jointing
all hist
ogra
m
s of blo
c
ks, we
can o
b
tain the HO
G
feature
s
of the obje
c
t.
Dalal
-
Tri
g
g
s
detecto
r sh
o
w
s
optim
al
p
e
rform
a
n
c
e
in the
practi
cal ap
plication
s
by
many
schola
r
s. In
a
ddition,
HOG
feature
s
al
so
have ma
ny th
eoreti
c
al
adv
antage
s. Fi
rst
l
y, it count
s f
o
r
not only the hori
z
ontal
an
d vertical di
re
ction
s
but
al
so many othe
r dire
ction
s
so
as to de
scri
be
the ship
s mo
re pre
c
isely. Seco
ndly, it is illumi
nation
-
in
variant due to
the normali
zation and that
i
t
works o
n
the
relatively sm
all pat
che
s
th
at call
ed
cell
s. Both of the
m
are h
e
lpful
for u
s
to
mo
del
those
ship
s a
ppea
ran
c
e eff
e
ctively.
Next, we ta
ke advanta
ge
of the typical
linear SVM to model
ea
ch shi
p
. Line
a
r
SVM is
t
he simple
st
sup
port
v
e
ct
or ma
chin
e.
I
t
classi
fie
s
l
a
rge n
u
mb
er of sample
s
into two types,
positive or n
e
gative that labeled wi
th 1
or 0, with a linear di
scrimi
native functio
n
. Acco
rdin
g to
maximum m
a
rgin
prin
cipl
e, we can fi
nally
obtain the
co
rrespo
nding
weig
ht output as
our
appe
ara
n
ce model
s by a certai
n amo
u
n
t of training sampl
e
s
with kno
w
n la
bels.
2.2. Mutual Relatio
n
Mo
del Base
d o
n
Deformabl
e
Part Model
The multiple
ship tra
c
king
framework we ad
opt is the mutual rel
a
tion model [
16]. We
denote
all th
e
shi
p
s of inte
rest a
s
V. E
a
ch ship
patch i
V
is re
pre
s
ented as
O
i
= {
x
i
, w
i
, h
i
} w
i
th
c
e
n
t
er
loc
a
tion
x
i
=
(
x
i
, y
i
), width
w
i
a
n
d
height
h
i
. T
hen th
e mo
d
e
l can b
e
int
u
itively described
usin
g (1
).
()
(
)
,
(i
,
j
)
E
(O
)
,
ii
i
i
j
i
j
iV
f
wO
w
O
O
fj
ÎÎ
=·
+
·
åå
(
1
)
Whe
r
e
(
)
i
O
f
is the
HOG fe
ature
vector exa
c
t
ed from im
ag
e
I
corre
s
p
o
n
d
ing to
ship
patch
es.
(
)
,
ij
OO
j
is the
mutua
l
relatio
n
vect
or b
e
twe
en
i
O
and
j
O
that de
scrib
e
s ho
w m
u
ch i
m
pa
ct
one o
b
je
ct h
a
s o
n
an
oth
e
r on
e.
i
w
and
,
ij
w
are the relational
wei
ghts that b
a
l
ance all
intera
ction
s
o
f
object
s
and
the re
spon
ses of ea
ch
o
b
ject.
G
= (
V, E
) rep
r
e
s
ent
s the relation
al
grap
h wh
ose node
s
V
are
obje
c
ts an
d e
dge
s
e
ij
indicate mutual rel
a
ted obje
c
ts.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
72 – 777
7
7774
Actually, we
cho
o
se the
spring
chan
ge
bet
ween
two obj
ect
s
a
s
the m
u
tual
relatio
n
vectors. Th
e
n
we
can
de
fine a
score
of
an asso
ci
ated c
onfig
uration C=
(
1
O
,
2
O
,…
V
O
) a
s
follows
:
()
2
,
(i
,
j
)
E
(C
)
(
)
ii
i
j
i
j
i
j
iV
sO
w
bf
ÎÎ
=·
-
·
-
-
åå
wx
,
x
e
(
2
)
Whe
r
e th
e fi
rst te
rm in
the ri
ght
side
of (2
) i
s
th
e ap
pea
ran
c
e sco
r
e th
at mea
s
u
r
e
s
t
he
compatibility between the
obser
ved im
age patches and the
object
patches. Especi
a
lly,
i
w
ar
e
the HOG
feat
ure
weights.
The
se
co
nd t
e
rm i
s
t
he
de
formabl
e
sco
r
e th
at calcul
ates th
e
sp
ri
n
g
cha
nge
bet
ween th
e o
b
se
rved o
b
je
cts l
o
catio
n
s an
d
the groun
d truth lo
cation
o
f
those
obj
ect
s
.
H
e
re
we
c
ons
id
er
th
e
,
ij
w
as a hyper-pa
r
a
m
eter an
d se
t
i, j
:
,
ij
ww
=
. The
n
all paramet
ers
can b
e
rep
r
e
s
ented by
=(
i
w
,
|V|
w
,
1
,
.
..,
E
ee
).
The tra
cki
ng
output is the config
urati
on
that maximizing
(2) a
nd this ca
n be found
efficiently usi
ng dyna
mic
p
r
og
ramin
g
. O
b
viously, t
he
smalle
r the
spring
chan
ge
is, the
clo
s
e
r
our
track
i
ng output is
to the true loc
a
tion.
2.3. Learnin
g
In our t
r
a
c
ker, learni
ng aim
s
to u
pdate
p
a
ram
e
ters
with pos
i
tive
(i.e. objec
t of
interes
t
sele
cted fro
m
the first frame an
d the
output co
nfiguratio
ns i
n
previou
s
fra
m
es) and
ne
gative
config
uratio
n
s
. We ad
opt
the favora
bl
e structu
r
ed
SVM to upd
ate mod
e
l p
a
ram
e
ters
wi
th
a
gradi
ent de
ce
nt method. Similar to [
17], we defin
e the
stru
cture
d
SVM loss
as
follows
:
(
,
)
m
a
x
(s
(
)
(C
)
(
C
,
))
C
CC
s
C
ib
b
b
=-
+
D
,
,
(
3
)
Whe
r
e
C
is the
grou
nd truth
and
C
is a particular
co
nfig
uration.
(,
)
CC
D
is
the loss
func
tion
that indicate
s the accu
ra
cy
of the predi
ction.
(,
)
CC
D
=0 if
CC
=
, or
(,
)
CC
D
>0. Equation
(3)
attempts to le
arn
a set of weight pa
ram
e
ters
so th
at the configu
r
ati
on sco
r
e
of the ground
trut
h
part lo
catio
n
s is
greater th
an the
sco
r
e
of any oth
e
r
possibl
e
confi
guratio
ns of p
a
rt lo
cation
b
y
at
least
(,
)
CC
D
, which
enco
d
e
s
the penalty of
predi
cting configuration
s
C
. Obvious
ly, the
stru
ctured S
V
M loss
is
convex in
pa
ramete
r
, b
e
ca
use it is t
he maximum
of a set of a
ffine
function
s.
The g
r
adi
ent
of the structu
r
ed SVM lo
ss in
(3)
with
respe
c
t to m
odel p
a
ra
met
e
rs
is
s
h
ow
n
as
(
4
)
.
*
(,
)
(
C
)
(
C
)
Cs
s
bb
b
ib
b
b
Ñ=
Ñ
-
Ñ
,
,
(
4
)
Whe
r
e the ne
gative config
uration
*
C
is given by:
*
ar
g
m
ax
(
(
)
(
C
,
)
)
C
CC
C
sb
=+
D
,
(5)
Ho
wever, ta
ki
ng the tra
d
e
-
off betwee
n
the
ap
pea
ran
c
e score an
d the defo
r
matio
n
score
into accou
n
t, we re
define t
he neg
ative appea
ran
c
e
score a
s
follows:
()
(C
)
ii
iV
O
s
bf
Î
=·
å
w
,
(6)
As a re
sult, the sea
r
ching d
i
rectio
n
p
for l
earni
ng is d
e
fined a
s
:
*
(C
)
(
C
)
s
ps
bb
b
b
=Ñ
-
Ñ
,
,
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Multip
le Ship Tra
cki
ng in Inland
Wate
rwa
y
CCTV System
(Fei Te
ng)
7775
And a pa
ssin
g-ag
gressive
algorith
m
[19
]
is
used to p
e
rform
the p
a
r
amete
r
s upd
ate as
sho
w
n in (
8
).
2
(,
)
1
||
|
|
2
C
p
p
K
ib
bb
¬-
+
(8)
Whe
r
e K
∈
(0 , +
∞
) is a
hyper-pa
r
am
eter that cont
rols
the “agg
ressivene
ss”
of the param
eter
update.
3. Experimental Re
sults
To verify our multiple shi
p
trackin
g
m
e
thod, we te
st on 3 CCTV videos in
inland
waterway. Fi
gure 1
sho
w
s the experime
n
tal results.
In CCTV 1, t
h
ree
shi
p
s a
r
e sele
cted
wi
th the colo
rful
boundi
ng b
o
x
es in the first frame
.
All the three ship
s are wit
h
different a
ppea
ran
c
e
s
and moving i
n
the sam
e
dire
ction
s
. In the
scene
that i
s
simple
an
d with
out
shi
p
o
ccl
us
i
on, in-pla
ne rotat
i
on
a
nd so
on,
ou
r
tracker
demon
strates excellent mul
t
iple ship tra
c
king p
e
rfo
r
ma
nce.
In additio
n
, in
CCTV 2,
we
sho
w
t
w
o
shi
p
s trac
kin
g
in
the
ca
se
of p
a
rtial o
c
clu
s
io
n. The
two target
shi
p
s move in th
e oppo
site direction
s
. Du
e to the appli
c
at
ion of spatial
con
s
trai
nts, we
can effe
ctively identify th
e ship b
e
ing
partial o
ccl
uded u
s
in
g the shi
p
s tha
t
are not bei
ng
occlud
ed.
CCTV
3
s
u
ffers
f
r
om the in-plane rotation. G
enerally
speaki
ng, shi
p
s in inland
waterwa
y
are al
ways
sailing in a re
latively slow spe
ed. As a
result, even
suffering fro
m
the in-pla
ne
rotation, the
appe
ara
n
ce
of ship i
s
alm
o
st with
little cha
nge. On t
he othe
r han
d, with the m
odel
para
m
eters l
earni
ng
and
updatin
g, ou
r tra
c
ker can
effectively ad
apt to th
e a
p
pearan
ce
ch
a
nge,
thus to pe
rform robu
st mult
iple shi
p
tracking.
CCTV 1
(Fra
me 1/38/287/
554in
seq
uen
ce)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 77
72 – 777
7
7776
CCTV 2
(1/60
/
127/307 in
seque
nce)
CCTV 3
(Fra
me 1/98/150/
282 in sequ
e
n
ce
)
Figure 1. Experime
n
tal Re
sults
wi
th Mul
t
iple Ship Tracki
ng
4. Conclusio
n
In this pape
r, we pro
p
o
s
e
a robu
st mult
iple shi
p
tracking meth
od
in inland
wat
e
rway.
Benefitting from HOG fe
ature
s
a
nd t
he line
a
r
S
V
M, we
can
effectively model th
e ship
s
appe
ara
n
ces.
More
over, th
e mutual
rela
tions b
a
sed o
n
the defo
r
m
able p
a
rt mo
del ma
ke g
r
e
a
t
contri
bution t
o
utilize the spatial co
nstraints to accom
p
lish multiple
ship tracing. Accordi
ng to th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Multip
le Ship Tra
cki
ng in Inland
Wate
rwa
y
CCTV System
(Fei Te
ng)
7777
theoreti
c
al
di
scussio
n
a
n
d
the exp
e
rim
ental resu
lts,
we
can
co
n
c
lud
e
that
o
u
r m
u
ltiple
ship
tracke
r is
suit
able for the i
n
land
wate
rway and d
e
mo
nstrate
s
outst
andin
g
multip
le shi
p
tra
c
ki
ng
perfo
rman
ce.
In future wo
rk,
we
aim to
explore
the
filter meth
od t
hat can
opti
m
ize
ou
r tra
c
ker
with the rob
u
s
t perfo
rman
ce in seri
ou
s o
ccl
usi
on.
Ackn
o
w
l
e
dg
ements
The autho
rs
woul
d like to
thank the e
d
itor and reviewe
r
s for the
i
r valuable co
mments
and
sugg
esti
ons th
at lead
to an improved man
u
scrip
t.
This wo
rk is supp
orted
by the Nation
al
Scien
c
e Fo
un
dation of Chi
na (NS
F
C 51
2791
52).
Referen
ces
[1]
Julia
A
Lass
e
r
r
e, Christ
oph
e
r
M Bish
op, T
homas
P Mi
nk
a.
Princ
i
pl
ed
Hybrids
of Ge
nerativ
e a
n
d
Discri
m
i
native Mode
ls.
Proc. of CVPR, Ne
w
York. 2006; 8
7
-
94.
[2]
A Balan, M Black.
An ad
ap
tive app
ear
an
ce mode
l ap
p
r
oach for
mo
d
e
l-b
a
sed
articulate
d ob
ject
tracking.
Proc. IEEE Conf. Computer Visi
on
and Patter
n
Re
cogn
ition, N
e
w
York. 2006: 75
8–7
65.
[3]
MJ Black, A
D
Jepso
n
. Ei
gen
T
r
acking: Rob
u
st ma
tchi
ng
a
nd track
i
ng
of
articul
a
ted
ob
j
e
cts usi
n
g
a
vie
w
-b
ase
d
rep
r
esentati
on.
Intl J. Com
p
uter
Vision
. 199
8; 26(1): 63–
84.
[4]
D Ross, J
Lim
,
RS Li
n, MH
Yang. Incr
eme
n
tal l
ear
nin
g
f
o
r rob
u
st visu
al tracki
ng.
Intl J. Com
p
uter
Vision
. 200
8; 77(1): 125
–1
41.
[5]
A Adam, E Rivlin, I Shimshon
i. Robust frag
m
ents-bas
ed trackin
g
usin
g the integr
al histo
g
ram.
CVPR
,
Ne
w
York. 20
0
6
: 798–
80
5.
[6]
SMS Nej
hum,
J Ho, MH Y
a
ng. Visu
al trac
ki
ng
w
i
th histo
g
rams
a
nd arti
culati
ng blocks
.
In
CVP
R,
Anchor
ag
e, AK. 2008; 1-8.
[7]
F
e
i T
eng, Qing Liu. Multi-sca
l
e
ship
tracki
ng
via ran
dom pro
j
ectio
n
s. In
SIV
P
. 2014; 8(3).
[8]
B Babe
nko, M
H
Yang, S Be
l
ong
ie. Ro
bust
obj
ec
t tracking
w
i
t
h
on
li
ne mul
t
iple i
n
stanc
e l
earn
i
ng.
IEEE
T
r
ans. Pattern Analys
is and M
a
chi
ne Intel
lig
e
n
ce
. 201
1; 33(
8): 1619
–1
632.
[9]
H Grabn
er, C
Leistn
e
r, H Bi
schof. Semi-su
per
vis
ed o
n
-li
n
e bo
ostin
g
for
robust tracki
n
g
. In
ECCV
,
Marseil
l
e-F
r
an
ce. 2008: 2
34–
247.
[10]
Z
Kalal, J
Mata
s, K Mikol
a
jcz
yk. P-n le
arni
ng:
B
ootstrappingbinar
y
classifiers
b
y
structura
l
constrai
nts.
Proc. IEEE Conf. Computer V
i
sion
and P
a
ttern Reco
gniti
on
,
San F
r
ancisco
, CA. 2010: 49
–56.
[11]
N Dal
a
l, B T
r
iggs. Histogr
ams
of
orie
nted gr
adi
ents forhum
an d
e
tectio
n.
Proc. IEEE Conf. Computer
Visio
n
and P
a
ttern Rec
ogn
it
io
n. San Die
go,
CA. June 20
05
; 886–8
93.
[12]
Qiang Z
hu, Ye
h MC, Avidan
S. F
a
st Huma
n De
tectio
n Using a Casc
ad
e
of
Histograms
of Oriented
Gradients.
IEEE Com
p
uter S
o
ciety Conf
erence on Computer Visi
on and Pattern Rec
ognition
, Ne
w
York. 2006; 2;
149
1-14
98.
[13]
P F
e
lzensz
w
a
l
b, D McAllest
er. “A discrimi
nat
ive
l
y
tra
i
n
e
d
, multiscal
e
, deforma
ble
pa
rt model,” In
CVPR
, Anchor
age, AK. June
200
8. pp. 1-8.
[14]
S Brans
on, P
Peron
a
, S Be
l
ong
ie, “Stron
g
superv
i
si
o
n
fro
m
w
eak
a
nnot
ation: Inter
a
cti
v
e trai
nin
g
o
f
deforma
ble p
a
r
t
models,” In
Proc. IEEE Int.
Conf. Com
p
uter Vision
, Barce
l
ona
.
Jun
e
20
1
1
, pp. 183
2–
183
9.
[15]
Lu Z
h
an
g, La
urens
van
d
e
r
Maate
n
.
“Structure
preserv
i
ng
ob
ject trac
king.
IEEE Conferenc
e
on
Co
mp
uter Visi
on an
d Pattern
Recog
n
itio
n
. Portlan
d
, Orego
n. 2013. p
p
. 18
38-1
845.
[16]
Genqu
an D
u
a
n
, Haizh
ou Ai
, Song Cao,
and Sh
ih
ong
Lao. “Group
T
r
acking: “Exp
lorin
g
Mutua
l
Relati
ons for M
u
ltipl
e
Object Tracking. In
ECCV
. F
l
orence.
October 201
2; 129-
143.
[17]
A Yilmaz, O Javed, M Shah. Object tracking
: A surve
y
,”
AC
M Computi
ng
Surveys
. 200
6;
38(4):
[18]
I T
s
ochantarid
is, T
Joachim
s
, T
Hofmann
, Y
Altun.
Lar
ge M
a
rgi
n
M
e
thods for
Stru
ctured
an
d
Interdep
en
dent
Output Variab
l
e
s.
JMLR
. 200
5; 6: 1453-
148
4.
[19]
K Crammer, O
Dek
e
l, J K
e
sh
et, S Sha
l
ev-S
h
w
ar
tz, Y S
i
ng
er. Onli
ne
pas
sive-a
ggress
i
v
e
a
l
gor
ithms
.
Journ
a
l of Mac
h
in
e Le
arn-i
ng
Rese
arch
. 20
0
6
; 7: 551–
585.
Evaluation Warning : The document was created with Spire.PDF for Python.