I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
1
4
,
No.
5
,
Oc
tober
20
2
4
,
pp
.
5681
~
5690
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/
ij
e
c
e
.
v
1
4
i
5
.
pp
5
681
-
5690
5681
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
A
c
ost
-
e
f
f
e
c
t
iv
e
,
r
e
li
ab
le
a
n
d
ac
c
u
r
at
e
f
r
am
e
w
or
k
f
o
r
m
u
ltip
l
e
-
t
a
r
ge
t
t
r
a
c
k
i
n
g b
y
d
e
t
e
c
t
io
n
a
p
p
r
oac
h
u
si
n
g d
e
e
p
n
e
u
r
al
n
e
t
w
o
r
k
Divyap
r
ab
h
a
1
,
Gu
r
u
p
r
as
ad
S
e
e
b
aiah
2
1
D
e
pa
r
tm
e
nt
of
E
le
c
tr
oni
c
s
a
nd C
omm
uni
c
a
ti
on E
ngi
ne
e
r
in
g, S
r
i
S
id
dha
r
th
a
I
ns
ti
tu
te
of
T
e
c
hnol
ogy, T
umkur
, I
ndi
a
2
D
e
pa
r
tm
e
nt
of
B
io
me
di
c
a
l
E
ngi
ne
e
r
in
g, S
r
i
S
id
dha
r
th
a
I
ns
ti
tu
t
e
of
T
e
c
hnol
ogy, T
umkur
,
I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
F
e
b
27,
2024
R
e
vis
e
d
J
ul
5,
2024
Ac
c
e
pted
J
ul
9,
2024
O
v
er
t
h
e
y
ears
t
h
e
ar
ea
o
f
o
b
j
ect
t
rack
i
n
g
an
d
d
et
ec
t
i
o
n
h
as
emer
g
ed
a
n
d
b
eco
me
u
b
i
q
u
i
t
o
u
s
o
w
i
n
g
t
o
i
t
s
p
o
t
en
t
i
a
l
co
n
t
ri
b
u
t
i
o
n
t
o
w
ar
d
s
v
i
d
e
o
s
u
r
v
ei
l
l
a
n
ce
ap
p
l
i
ca
t
i
o
n
s
.
Mu
l
t
i
p
l
e
o
b
j
ect
t
rac
k
i
n
g
(
MO
T
)
es
t
i
ma
t
es
t
h
e
t
raj
ec
t
o
r
y
o
f
s
ev
era
l
o
b
j
ec
t
s
o
f
i
n
t
ere
s
t
s
i
m
u
l
t
an
e
o
u
s
l
y
o
v
er
t
i
me
i
n
a
s
eri
e
s
o
f
v
i
d
eo
frame
s
.
E
v
en
t
h
o
u
g
h
v
ari
o
u
s
res
earc
h
p
ro
p
o
s
al
s
h
a
v
e
en
c
o
u
ra
g
e
d
t
h
e
u
s
e
o
f
mach
i
n
e
l
ear
n
i
n
g
t
ech
n
i
q
u
e
s
i
n
d
e
s
i
g
n
i
n
g
mu
l
t
i
-
o
b
j
ect
t
rack
er
s
,
t
h
e
ex
i
s
t
i
n
g
s
o
l
u
t
i
o
n
s
n
eed
t
o
b
e
mo
re
p
ract
i
cab
l
e
fo
r
o
n
l
i
n
e
t
rack
i
n
g
d
u
e
t
o
mo
re
co
mp
l
i
ca
t
ed
a
l
g
o
ri
t
h
m
s
,
T
h
e
s
t
u
d
y
,
t
h
erefo
re,
i
n
t
r
o
d
u
ces
a
co
s
t
-
effect
i
v
e
t
rack
i
n
g
s
o
l
u
t
i
o
n
fo
r
mu
l
t
i
p
l
e
–
t
ar
g
et
t
rack
i
n
g
b
y
d
e
t
ect
i
o
n
w
h
ere
i
t
i
n
c
o
rp
o
rat
e
s
t
h
e
y
o
u
o
n
l
y
l
o
o
k
o
n
ce
v
er
s
i
o
n
4
(Y
O
L
O
v
4
)
an
d
p
ers
o
n
re
-
i
d
e
n
t
i
fi
ca
t
i
o
n
n
et
w
o
r
k
,
w
h
i
ch
are
fu
r
t
h
er
i
n
t
eg
ra
t
ed
w
i
t
h
t
h
e
p
ro
p
o
s
e
d
t
rack
i
n
g
mo
d
e
l
,
w
h
i
ch
co
n
s
i
d
er
s
b
o
t
h
b
o
u
n
d
i
n
g
b
o
x
an
d
ap
p
earan
ce
fea
t
u
re
s
t
o
h
an
d
l
e
t
h
e
mo
t
i
o
n
p
red
i
c
t
i
o
n
an
d
d
at
a
as
s
o
ci
a
t
i
o
n
res
p
ect
i
v
e
l
y
.
T
h
e
n
o
v
el
t
y
o
f
t
h
i
s
ap
p
ro
ac
h
l
i
e
s
i
n
co
n
s
i
d
eri
n
g
ap
p
earan
ce
feat
u
res
,
w
h
i
ch
n
o
t
o
n
l
y
h
e
l
p
p
red
i
ct
t
rac
k
s
t
h
r
o
u
g
h
a
l
l
o
cat
i
o
n
s
p
r
o
b
l
em
s
o
l
v
i
n
g
b
u
t
a
l
s
o
h
an
d
l
e
t
h
e
co
s
t
o
f
co
m
p
u
t
at
i
o
n
p
ro
b
l
em
s
.
H
ere,
t
h
e
s
y
s
t
em
u
t
i
l
i
zes
a
p
re
-
t
ra
i
n
e
d
as
s
o
ci
a
t
i
o
n
met
ri
c
w
i
t
h
w
h
i
ch
t
h
e
o
cc
l
u
s
i
o
n
ch
a
l
l
e
n
g
es
are
al
s
o
h
an
d
l
e
d
,
w
h
erea
s
t
h
e
t
arg
e
t
t
rack
i
n
g
h
as
t
ak
e
n
p
l
ace
ev
e
n
i
n
mo
re
ex
t
en
d
e
d
p
eri
o
d
s
o
f
o
ccl
u
s
i
o
n
,
mak
i
n
g
i
t
s
u
i
t
ab
l
e
w
i
t
h
t
h
e
ex
i
s
t
i
n
g
eff
i
ci
e
n
t
t
rack
i
n
g
al
g
o
r
i
t
h
ms
.
K
e
y
w
o
r
d
s
:
C
omput
e
r
vis
ion
C
onvolut
ional
ne
ur
a
l
ne
twor
ks
Da
ta
a
s
s
oc
iation
De
e
p
lea
r
ning
M
ult
ipl
e
tar
ge
t
tr
a
c
king
P
e
r
s
onR
e
-
I
DN
e
twor
k
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Divya
pr
a
bha
De
pa
r
tm
e
nt
of
E
lec
tr
onics
a
nd
C
omm
unica
ti
on
E
n
ginee
r
ing,
S
r
i
S
iddhar
tha
I
ns
ti
tut
e
of
T
e
c
hnology
839V+
8G9,
T
umku
r
-
Kuniga
l
R
d,
S
a
r
a
s
wa
thi
pur
a
m,
T
umkur
,
I
ndia
E
mail:
divyas
y74@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
Ove
r
the
ye
a
r
s
,
the
unde
r
lyi
ng
c
onc
e
pt
of
obj
e
c
t
de
tec
ti
on
a
nd
tr
a
c
king
ha
s
e
mer
ge
d
a
nd
i
s
c
ons
ider
e
d
one
of
the
mos
t
e
s
s
e
nti
a
l
a
nd
c
ha
ll
e
nging
br
a
nc
he
s
of
c
omput
e
r
vis
ion.
How
e
ve
r
,
it
s
id
e
a
s
ha
ve
be
e
n
wide
ly
a
ppli
e
d
in
va
r
ious
f
ields
,
including
he
a
lt
hc
a
r
e
moni
tor
ing,
a
utonom
ous
dr
ivi
ng
,
a
nomaly
de
tec
ti
on,
ve
hicle
de
tec
ti
on
a
nd
pe
de
s
tr
ian
tr
a
c
ki
ng
[
1]
.
I
t
c
a
n
a
ls
o
be
s
e
e
n
that
a
va
r
iety
o
f
mo
da
li
ti
e
s
,
including
r
a
da
r
,
l
ight
de
tec
ti
on
a
nd
r
a
nging
(
L
I
DA
R
)
,
a
nd
c
omput
e
r
vis
ion
(
C
V)
,
ha
ve
be
c
ome
a
va
il
a
ble
f
or
the
pur
pos
e
of
objec
t
de
tec
ti
on
a
nd
t
r
a
c
king.
I
n
r
e
c
e
nt
ye
a
r
s
,
s
igni
f
ica
nt
p
r
ogr
e
s
s
c
ould
be
s
e
e
n
e
ve
n
in
the
f
ield
o
f
im
a
ging
tec
hnology
a
s
c
a
mer
a
s
a
r
e
now
c
he
a
pe
r
,
s
maller
a
nd
of
higher
qua
li
ty
than
e
ve
r
be
f
or
e
.
On
the
other
ha
nd,
c
omput
ing
platf
o
r
ms
ha
ve
be
e
n
opti
mi
z
e
d
towa
r
d
pa
r
a
ll
e
li
z
a
ti
on,
s
uc
h
a
s
m
ult
i
-
c
or
e
pr
oc
e
s
s
ing
a
nd
gr
a
phica
l
p
r
oc
e
s
s
ing
unit
s
(
GPUs
)
.
S
uc
h
ve
r
s
ions
a
ll
ow
C
V
towa
r
ds
e
f
f
icie
nt
objec
t
de
tec
ti
on
a
nd
tr
a
c
king
to
pur
s
ue
r
e
a
l
-
ti
me
im
p
leme
ntation.
T
he
r
a
pid
de
ve
lopm
e
nt
o
f
de
e
p
lea
r
ning
(
DL
)
ne
twor
ks
a
long
with
GPU
’
s
c
omput
ing
powe
r
ha
s
s
igni
f
ic
a
ntl
y
c
ontr
ibut
e
d
towa
r
ds
im
p
r
oving
the
pe
r
f
or
m
a
nc
e
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5681
-
5690
5682
objec
t
de
tec
tor
s
a
nd
tr
a
c
ke
r
s
.
T
he
c
onc
e
pts
be
hind
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
a
n
d
GPU’
s
e
nha
nc
e
d
c
omput
ing
powe
r
a
r
e
the
pr
im
e
r
e
a
s
ons
be
hind
f
a
s
t
e
volut
ion
of
C
V
-
ba
s
e
d
objec
t
de
tec
ti
on
a
nd
tr
a
c
king
[
2]
.
I
n
thi
s
c
ontext
it
ha
s
to
be
mentio
ne
d
that
de
e
p
lea
r
ning
idea
s
ha
ve
e
volved
f
r
om
mac
hine
lea
r
ning
(
M
L
)
a
nd
a
ls
o
ther
e
c
omes
c
ha
r
a
c
ter
is
ti
c
s
dif
f
e
r
e
nc
e
s
.
B
a
s
ica
ll
y,
M
L
is
c
ons
ider
e
d
a
s
a
br
a
nc
h
of
a
r
ti
f
icia
l
int
e
ll
igenc
e
(
AI
)
whe
r
e
it
ba
s
ica
ll
y
lea
r
ns
pa
tt
e
r
ns
f
r
om
e
xa
mpl
e
s
or
s
a
mpl
e
da
ta.
He
r
e
the
mac
hine
is
de
s
igned
in
s
uc
h
a
wa
y
that
it
ha
s
the
a
bil
it
y
t
o
a
c
c
e
s
s
a
nd
lea
r
n
f
r
om
the
da
ta.
He
r
e
the
da
ta
c
ould
be
labe
led,
unlabe
led
o
r
thei
r
c
ombi
na
ti
on
.
T
he
lea
r
ning
c
ould
be
e
it
he
r
s
upe
r
vis
e
d,
un
-
s
upe
r
vis
e
d
or
s
e
mi
-
s
upe
r
vis
e
d.
Ar
ti
f
icia
l
ne
ur
a
l
ne
twor
ks
(
AN
Ns
)
a
r
e
s
ubjec
ted
to
ha
ve
the
a
bil
it
y
to
lea
r
n
the
r
e
lation
be
twe
e
n
input
a
nd
output
f
r
om
e
xa
mpl
e
s
a
nd
a
r
e
ba
s
ica
ll
y
c
ons
ider
e
d
a
s
potential
s
olut
ion
s
of
M
L
.
I
n
the
e
a
r
ly
2000s
,
c
e
r
tain
br
e
a
kthr
oughs
in
e
nginee
r
ing
r
e
s
e
a
r
c
h
f
o
r
mul
ti
-
laye
r
e
d
ne
ur
a
l
ne
twor
ks
(
M
L
P
)
ha
d
a
id
in
a
r
r
ival
o
f
de
e
p
lea
r
ning.
DL
r
e
f
e
r
s
to
a
lea
r
ning
in
de
pth
pa
r
a
digm
whic
h
invol
ve
s
dif
f
e
r
e
nt
laye
r
s
a
nd
s
tage
s
[
3]
.
Unlike
c
onve
nti
ona
l
s
ha
ll
ow
lea
r
ning
models
,
D
L
is
de
pe
nde
nt
ove
r
hundr
e
ds
or
thous
a
nds
of
f
r
a
mes
a
s
a
r
e
s
ult
it
ha
s
be
c
ome
c
omput
a
ti
ona
ll
y
int
e
ns
ive
a
nd
dif
f
icult
to
e
nginee
r
whic
h
is
c
ons
ider
e
d
a
s
a
pr
omi
ne
nt
r
e
s
e
a
r
c
h
c
ha
ll
e
nge
by
s
tudi
e
s
[
4]
a
nd
[
5]
.
I
t
r
e
quir
e
s
a
ve
r
y
high
-
pe
r
f
or
manc
e
GPU
to
pr
ovide
f
a
s
t
a
nd
a
c
c
ur
a
te
objec
t
r
e
c
ognit
ion
a
nd
mot
ion
de
tec
ti
on.
E
ve
n
though
DL
models
a
r
e
mos
tl
y
c
ompl
e
x
in
de
s
ign
a
nd
ha
ving
de
pe
nde
nc
y
ove
r
huge
a
mount
of
tr
a
ini
ng
da
ta
a
nd
GPU’
s
c
omput
ing
powe
r
but
thes
e
f
a
c
tor
s
c
a
n
be
be
tt
e
r
ha
ndled
with
DL
models
a
s
c
laimed
by
[
5
]
a
nd
[
6]
.
I
t
ha
s
be
e
n
f
ound
that
ther
e
e
xis
t
va
r
ious
be
nc
hmar
k
da
tas
e
ts
f
o
r
objec
t
de
tec
ti
on
s
uc
h
a
s
C
a
li
f
or
nia
I
ns
ti
tut
e
of
T
e
c
hnology
(
C
a
lt
e
c
h)
[
6]
,
K
a
r
ls
r
uhe
I
ns
ti
tut
e
of
T
e
c
hnology
a
nd
T
oyota
T
e
c
hn
ologi
c
a
l
I
ns
ti
tut
e
(
KI
T
T
I
)
[
7]
,
I
mage
Ne
t
[
8
]
,
P
ASC
AL
vis
ua
l
objec
t
c
las
s
e
s
(
P
ASC
AL
VO
C
)
[
9]
,
M
icr
os
of
t
C
omm
on
Obje
c
ts
in
C
ontext
(
M
S
C
OC
O)
[
10]
a
nd
T
he
Op
e
n
I
mage
s
Da
tas
e
t
V5
[
11
]
.
Als
o
due
to
the
a
va
il
a
bil
it
y
of
huge
a
mount
s
of
da
tas
e
ts
a
nd
the
r
a
pid
de
ve
lopm
e
nt
of
GPUs
,
DL
models
a
r
e
wide
ly
a
dopted
f
or
objec
t
de
tec
ti
on
by
r
e
s
e
a
r
c
he
r
s
a
nd
the
a
dva
nc
e
ment
is
s
ti
ll
in
p
r
ogr
e
s
s
.
B
oc
hkovs
kiy
e
t
al.
[
12]
a
ddr
e
s
s
e
s
the
de
s
ign
c
ha
ll
e
nge
s
of
two
s
tage
de
tec
tor
s
s
uc
h
a
s
r
e
gion
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
R
C
NN
)
[
13
]
a
nd
f
a
s
t
R
C
NN
[
13]
whic
h
is
the
ne
xt
a
dva
nc
e
d
v
e
r
s
ion
of
R
C
NN
.
R
C
NN
is
c
ons
ider
e
d
to
be
the
f
i
r
s
t
mod
e
l
of
two
s
tage
de
tec
ti
on
whic
h
s
hows
that
de
e
p
-
C
NN
is
be
tt
e
r
than
c
onvolut
ional
method
f
or
objec
t
de
tec
ti
on.
On
the
other
ha
nd,
R
NN
is
f
ound
to
be
inv
olved
in
mul
ti
-
s
tage
tr
a
ini
ng
pr
oc
e
s
s
wh
e
r
e
a
s
f
a
s
t
R
C
NN
us
e
only
one
s
tage
e
nd
-
to
-
e
nd
tr
a
ini
ng
pr
oc
e
s
s
.
Hus
s
a
in
e
t
al.
[
14
]
in
thei
r
s
tudy
highl
ight
s
va
r
ious
de
e
p
lea
r
ning
-
ba
s
e
d
s
olut
ions
f
or
de
tec
ti
ng
objec
ts
in
im
a
ge
f
r
a
mes
or
videos
us
ing
s
e
ve
r
a
l
a
lgor
it
hms
whic
h
include
you
only
look
onc
e
(
YO
L
O)
,
s
pa
ti
a
l
pyr
a
mi
d
pooli
ng
(
S
P
P
-
ne
t)
,
f
a
s
t
R
-
C
NN
,
C
NN
ba
s
e
d
on
r
e
gions
(
R
-
C
NN
)
,
or
ient
gr
a
dient
his
togr
a
ms
(
HO
G)
,
s
ingl
e
s
hot
de
tec
tor
(
S
S
D
)
,
r
e
gion
-
ba
s
e
d
f
ull
y
c
onvolut
i
ona
l
ne
twor
k
(
R
-
F
C
N)
,
howe
ve
r
the
s
tudy
made
a
c
laim
that
e
ve
n
though
objec
t
de
tec
ti
on
a
lgor
it
hms
a
r
e
f
ound
to
be
a
c
c
ur
a
te
f
r
om
the
de
tec
ti
on
a
c
c
ur
a
c
y
point
of
view
but
the
s
pe
e
d
o
f
in
f
e
r
e
nc
ing
is
a
lwa
ys
be
i
ng
a
p
r
oblem.
YO
L
O
ha
s
be
c
ome
ve
r
y
popu
lar
in
objec
t
de
tec
ti
on
a
nd
c
ons
ider
e
d
to
be
s
tate
-
of
-
the
-
a
r
t
tec
h
nology
in
objec
t
de
tec
ti
on
a
lgor
it
hm.
YO
L
O
is
f
ou
nd
to
be
mor
e
a
c
c
ur
a
te
a
nd
f
a
s
ter
whe
n
c
ompar
e
d
with
R
C
NN
a
nd
f
a
s
t
R
C
NN
a
nd
a
ls
o
f
ound
s
uit
a
ble
f
o
r
dyna
mi
c
mul
ti
-
objec
t
tr
a
c
king.
I
n
the
c
ontext
of
YO
L
O,
obj
e
c
t
de
tec
ti
on
is
c
ons
ider
e
d
to
be
the
r
e
g
r
e
s
s
ion
pr
oblem.
I
t
ha
s
to
be
noted
that
thi
s
a
pp
r
oa
c
h
is
bui
lt
on
the
t
op
of
C
NN
f
ound
s
uit
a
ble
f
or
r
e
a
l
-
ti
me
objec
t
de
t
e
c
ti
on.
I
n
the
c
ontext
of
objec
t
de
tec
ti
on
one
s
tage
de
tec
tor
s
ha
ve
be
c
ome
popular
whic
h
include
YO
L
O
[
15]
,
YO
L
Ov2
[
16]
,
YO
L
Ov3
[
17]
.
He
r
e
YO
L
O
s
tands
f
or
you
only
look
onc
e
.
How
e
ve
r
,
it
ha
s
be
e
n
obs
e
r
ve
d
in
the
e
xis
ti
ng
s
ys
tem
that
e
ve
n
though
c
ompar
e
d
to
R
C
NN
,
YO
L
O
r
e
duc
e
s
the
ba
c
kgr
ound
f
a
ls
e
po
s
it
ives
by
thr
e
e
ti
mes
,
but
it
ha
s
li
mi
tations
f
a
c
tor
s
s
uc
h
a
s
it
im
pos
e
s
s
tr
ong
s
p
a
ti
a
l
c
ons
tr
a
int
s
on
bounding
box
pr
e
dictions
.
A
s
igni
f
ica
nt
im
pr
ove
ment
ove
r
YO
L
O
ha
s
take
n
plac
e
in
it
s
s
e
c
ond
it
e
r
a
ti
on
whic
h
is
na
mely
YO
L
Ov2
whic
h
ha
s
im
pr
ove
d
the
s
pe
e
d
a
nd
the
de
tec
ti
on
p
r
e
c
is
ion
to
s
ome
e
xtent
.
B
ut
it
s
l
im
it
a
ti
on
c
omes
with
high
r
e
s
olut
ion
de
tec
ti
on
a
nd
s
ingl
e
c
las
s
objec
t
s
whic
h
is
f
ur
ther
a
ddr
e
s
s
e
s
in
YO
L
Ov3
whic
h
is
the
a
dva
nc
e
d
ve
r
s
ion
of
YO
L
Ov2,
a
nd
it
inco
r
por
a
tes
de
e
p
C
NN
Da
r
k
N
et
-
53
f
or
the
f
e
a
tur
e
ge
ne
r
a
ti
on
model
ing.
I
t
is
ba
s
ica
ll
y
c
a
pa
ble
of
mul
t
i
-
c
las
s
c
las
s
if
ica
ti
on
a
nd
is
us
e
d
f
o
r
mos
tl
y
s
mall
objec
t
c
las
s
if
ica
ti
on
pr
oblem,
howe
ve
r
,
YO
L
Ov3
model
lac
ks
e
f
f
e
c
ti
ve
ne
s
s
f
or
the
de
tec
ti
on
of
medium
a
nd
la
r
ge
s
ize
d
objec
ts
f
or
whic
h
li
mi
ts
it
s
s
c
ope
of
a
ppli
c
a
bil
it
y
in
r
e
a
l
-
ti
me
t
r
a
c
king
ope
r
a
ti
ons
.
How
e
ve
r
,
W
a
ng
e
t
al
.
[
18]
e
xhibi
ts
the
a
dva
ntage
of
YO
L
Ov4
in
mul
ti
objec
t
de
tec
ti
on
a
nd
pr
opos
e
s
a
ne
w
va
r
iant
of
YO
L
Ov4
thr
ough
whic
h
it
tar
ge
ted
to
e
nha
nc
e
the
mul
ti
-
objec
t
tr
a
c
king
pe
r
f
o
r
manc
e
a
s
we
ll
.
How
e
ve
r
,
the
ba
s
e
li
ne
a
ppr
oa
c
h
of
L
iu
e
t
al
.
[
19]
i
s
a
ls
o
c
ons
ider
e
d
in
the
pr
opos
e
d
s
tudy
whe
r
e
it
is
f
ound
that
the
da
ta
a
s
s
oc
iation
a
nd
t
r
a
c
k
mana
ge
ment
poli
c
ies
in
r
e
c
e
nt
a
lgor
i
thm
s
f
a
c
e
s
dif
f
i
c
ult
ies
in
de
a
li
ng
with
mo
ti
on
p
r
e
diction
a
nd
oc
c
lus
ion
ha
n
dli
ng
a
nd
a
ls
o
objec
t
pe
r
s
on
r
e
-
identif
ica
ti
on
e
nti
f
ica
ti
on.
M
a
ny
r
e
c
e
nt
a
lgor
it
hms
a
r
e
f
ound
to
e
mpl
oy
mot
i
on
a
nd
a
ppe
a
r
a
nc
e
indi
c
a
ti
ons
to
de
a
l
with
s
uc
h
c
ha
ll
e
nge
s
but
e
nd
up
incr
e
a
s
ing
the
c
omput
a
ti
on
c
os
t
a
nd
f
or
thi
s
r
e
a
s
on
the
s
pe
e
d
of
the
a
lgor
it
hm
a
ls
o
de
c
r
e
a
s
e
s
whic
h
make
s
them
uns
uit
a
ble
f
or
onli
ne
tr
a
c
king
a
ppli
c
a
ti
ons
.
Als
o
ther
e
e
xis
t
va
r
ious
a
lgor
it
hms
whic
h
only
us
e
s
mot
ion
c
ue
s
in
c
ontr
a
s
t
to
incr
e
a
s
e
the
s
pe
e
d
of
a
lgor
it
hm
but
c
a
nnot
p
r
ope
r
ly
ha
ndle
oc
c
lus
ions
a
nd
pe
r
s
on
r
e
-
identif
ica
ti
on
identif
y
the
los
t
objec
ts
a
s
s
hown
in
the
s
tudi
e
s
by
[
20
]
–
[
22]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
c
os
t
-
e
ff
e
c
ti
v
e
,
r
e
li
able
and
ac
c
ur
ate
fr
ame
w
or
k
for
multi
ple
-
tar
ge
t
tr
ac
k
ing
by
…
(
Div
y
apr
abha
)
5683
Dis
c
us
s
ion
on
r
e
late
d
r
e
s
e
a
r
c
h
e
xhibi
ts
that
s
igni
f
ica
nt
pr
ogr
e
s
s
ha
s
be
e
n
made
in
the
li
ne
o
f
r
e
s
e
a
r
c
h
f
or
objec
t
de
tec
ti
on
a
nd
tr
a
c
king
-
by
-
de
te
c
ti
on
a
ppr
oa
c
he
s
,
whic
h
ha
ve
a
ls
o
s
igni
f
ica
ntl
y
c
o
ntr
ibut
e
d
towa
r
ds
mul
ti
ple
objec
t
t
r
a
c
king.
How
e
ve
r
,
the
c
h
a
ll
e
nge
a
r
is
e
s
to
a
ppr
opr
iate
ly
de
a
l
with
objec
t
tr
a
j
e
c
tor
ies
,
mos
tl
y
c
ons
ider
e
d
global
opti
mi
z
a
ti
on
pr
oblem
s
.
R
e
late
d
f
r
a
mew
or
ks
s
uc
h
a
s
the
idea
of
global
da
ta
a
s
s
oc
iation
f
or
mul
ti
objec
t
t
r
a
c
king
(
M
OT
)
whe
r
e
f
low
ne
two
r
k
f
o
r
mul
a
ti
ons
c
ould
be
f
ound
,
a
ls
o
globally
opti
mal
gr
e
e
dy
a
lgor
it
hms
a
r
e
us
e
d
f
o
r
t
r
a
c
king
va
r
iable
number
o
f
objec
ts
.
T
he
us
e
of
pr
oba
bil
is
ti
c
gr
a
phica
l
models
mainly
pr
oc
e
s
s
e
s
the
video
ba
tche
s
s
im
ult
a
ne
ous
ly
whe
r
e
non
-
li
ne
a
r
mot
ion
pa
tt
e
r
ns
a
long
with
r
obus
t
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
a
r
e
e
xplor
e
d
a
nd
e
va
luate
d.
How
e
ve
r
,
mos
t
of
thes
e
a
ppr
oa
c
he
s
of
M
OT
a
r
e
c
omput
a
ti
ona
ll
y
e
xpe
ns
ive
a
nd
do
not
pr
ovide
s
c
ope
of
a
ppli
c
a
bil
it
y
in
on
li
ne
s
c
e
na
r
ios
whe
r
e
tar
ge
t
identit
ies
c
ould
be
a
va
il
a
ble
in
e
a
c
h
ti
me
s
tep.
T
he
p
r
opos
e
d
a
ddr
e
s
s
e
s
the
s
e
c
ha
ll
e
nge
s
a
nd
f
ur
ther
int
r
oduc
e
s
a
c
os
t
-
e
f
f
e
c
ti
ve
,
r
e
li
a
ble
a
nd
a
c
c
ur
a
te
M
OT
de
s
ign
f
r
a
mew
or
k
c
ons
ider
ing
YO
L
Ov4
a
nd
a
s
s
oc
iation
metr
ic
f
or
the
pur
pos
e
of
tr
a
c
king
a
nd
dis
c
r
im
inating
pe
de
s
tr
ian
e
nti
ti
e
s
moveme
nt
a
c
r
os
s
a
s
tr
e
e
t.
I
n
thi
s
joi
nt
a
ppr
oa
c
h
of
M
OT
,
YO
L
Ov4
c
ontr
ibut
e
s
towa
r
ds
obtaining
the
de
tec
ti
ons
o
f
o
bjec
ts
in
indi
vidual
f
r
a
mes
.
And
a
ls
o,
f
ur
ther
thr
ough
c
onn
otation
metr
ic
the
p
r
opos
e
d
s
tudy
model
e
xe
c
utes
tr
a
c
k
of
a
s
s
oc
iation
a
nd
mana
ge
ment
a
c
r
os
s
the
f
r
a
mes
to
de
ter
mi
ne
the
tr
a
c
king
a
nd
dis
c
r
im
inating
objec
t
o
f
int
e
r
e
s
t
moveme
nt
whic
h
is
pe
de
s
tr
ian
in
the
pr
opos
e
d
s
tu
dy
c
ontext.
2.
M
E
T
HO
D
T
he
pr
im
e
objec
ti
ve
of
th
is
s
tudy
is
to
int
r
oduc
e
a
joi
nt
f
r
a
mew
or
k
o
f
M
OT
whic
h
e
xplor
e
s
the
s
tr
e
ngth
f
a
c
tor
s
of
a
s
s
oc
iation
metr
ic
by
mea
ns
of
int
r
oduc
ing
s
ingl
e
hypothes
is
tr
a
c
king
meth
od
with
Ka
lm
a
n
f
il
ter
ing
(
KF
)
a
pp
r
oa
c
h.
He
r
e
the
de
s
ig
n
a
nd
de
ve
lopm
e
nt
o
f
the
f
r
a
mew
or
k
f
or
mul
ti
-
objec
t
of
int
e
r
e
s
t
tr
a
c
king
a
im
s
to
o
f
f
e
r
c
os
t
-
e
f
f
e
c
ti
ve
ne
s
s
a
long
with
a
c
c
ur
a
c
y
in
d
is
c
r
im
inating
the
pe
de
s
tr
ian
moveme
nt
pa
tt
e
r
ns
.
He
r
e
the
de
s
ign
a
nd
modeli
ng
of
the
t
r
a
c
ke
r
will
plan
to
uti
li
z
e
the
de
e
p
a
p
pe
a
r
a
nc
e
f
e
a
tur
e
s
f
r
om
the
de
e
p
ne
ur
a
l
ne
two
r
k
(
DN
N)
-
ba
s
e
d
model
to
im
pr
ove
the
tr
a
c
king
a
c
c
ur
a
c
y
of
iden
ti
f
ica
ti
on
of
pe
de
s
tr
ians
’
e
nti
ti
e
s
.
T
he
pr
im
e
objec
ti
ve
of
t
his
pr
opos
e
d
s
tudy
model
will
be
to
e
s
ti
mate
the
c
ount
of
objec
t
of
int
e
r
e
s
t
in
a
given
video
s
c
e
ne
.
I
t
wil
l
a
ls
o
a
im
to
e
s
ti
mate
thei
r
pos
it
ion
p
r
e
c
is
e
ly
while
maintaining
their
unique
f
or
m
o
f
identi
ti
e
s
.
T
he
s
tudy
model
he
r
e
a
im
s
to
pe
r
f
or
m
mul
t
i
-
objec
t
tr
a
c
king
e
ve
n
in
the
pr
e
s
e
nc
e
of
ove
r
oc
c
lus
ions
.
He
r
e
the
wor
k
-
f
low
s
tr
a
tegic
de
s
ign
of
mul
ti
-
objec
t
tr
a
c
king
will
a
ls
o
c
ons
ider
e
f
f
icie
nt
a
nd
s
im
pli
f
ied
e
xe
c
uti
on
s
teps
s
o
that
c
os
t
of
ope
r
a
ti
ons
c
a
n
be
c
ontr
oll
e
d
with
a
ppr
op
r
i
a
te
da
ta
a
s
s
oc
iation
metr
ics
.
T
he
s
tudy
a
ls
o
a
im
s
to
im
pleme
nt
the
DN
N
model
f
or
the
pur
pos
e
of
a
ll
oc
a
ti
ons
ope
r
a
ti
ons
a
nd
a
ls
o
pr
e
dict
the
t
r
a
c
ks
us
ing
KF
methods
.
T
he
s
tudy
a
ls
o
he
r
e
will
in
tr
oduc
e
a
n
ove
l
c
os
t
model
to
opti
mi
z
e
the
c
os
t
of
ove
r
a
ll
a
ll
oc
a
ti
ons
o
pe
r
a
ti
ons
f
or
both
de
tec
ti
ons
a
nd
t
r
a
c
ks
.
T
he
s
tud
y
f
ur
ther
e
labor
a
tes
the
methodology
a
dopted
in
the
pr
opos
e
d
s
ys
tem
f
or
M
OT
.
2.
1.
YOL
O
ob
j
e
c
t
d
e
t
e
c
t
or
YO
L
O
is
c
ons
id
e
r
e
d
to
be
a
n
e
f
f
icie
nt
objec
t
d
e
tec
tor
with
one
s
tage
whic
h
wa
s
de
s
igned
a
f
ter
F
a
s
ter
R
C
NN
.
T
he
a
dva
nc
e
d
va
r
iant
of
YO
L
Ov3
c
ons
ider
s
mul
ti
-
leve
l
c
las
s
if
ica
ti
on
a
nd
a
dop
ts
mor
e
c
ompl
e
x
da
tas
e
ts
whic
h
mi
ght
c
ontain
ove
r
lappi
ng
labe
ls
.
I
t
uti
li
z
e
s
th
r
e
e
di
f
f
e
r
e
nt
s
c
a
le
f
e
a
tur
e
maps
to
pr
e
dict
the
bound
ing
box
.
I
n
YO
L
Ov3
the
las
t
c
onvolut
ional
laye
r
int
e
gr
a
tes
3D
tens
or
e
nc
od
ing
c
las
s
pr
e
dictions
,
objec
tnes
s
a
nd
bounding
box.
How
e
ve
r
,
it
is
obs
e
r
ve
d
that
YO
L
Ov3
int
r
oduc
e
s
mor
e
e
f
f
e
c
ti
ve
a
nd
r
obus
t
f
e
a
tur
e
e
xtr
a
c
tor
whe
r
e
inclus
ion
of
Da
r
k
N
et
-
53
c
a
n
be
s
e
e
n
in
[
23]
,
[
24
]
.
T
he
pr
opos
e
d
s
tudy
f
or
de
s
igni
ng
c
os
t
-
e
f
f
e
c
ti
ve
,
r
e
li
a
ble
a
nd
a
c
c
ur
a
te
f
r
a
mew
or
k
of
M
OT
c
ons
ider
s
the
po
tential
a
s
pe
c
ts
of
dif
f
e
r
e
nt
va
r
iants
of
YO
L
O
a
s
ba
s
e
li
ne
models
a
nd
f
inds
th
e
s
tr
e
ngth
of
YO
L
Ov4
a
s
a
s
tate
-
of
-
a
r
t
a
lgor
it
hm
whic
h
is
12%
f
a
s
ter
c
ompar
e
d
to
the
pr
e
vious
ve
r
s
ion
o
f
YO
L
Ov3.
I
t
a
ls
o
r
e
a
li
z
e
s
that
if
a
n
e
f
f
e
c
ti
ve
tr
a
c
king
f
r
a
mew
or
k
model
is
de
s
igned
e
f
f
e
c
ti
ve
ly
c
ons
ider
ing
YO
L
Ov4
de
tec
ti
on
then
it
c
a
n
s
igni
f
ica
ntl
y
c
ontr
ibut
e
towa
r
ds
M
OT
whic
h
c
ould
be
c
ompar
a
ble
with
R
C
NN
a
nd
F
a
s
ter
R
C
NN
.
T
he
s
tudy
c
ons
ider
s
YO
L
O
one
-
s
tage
de
tec
tor
a
s
ba
s
e
li
ne
r
e
f
e
r
e
nc
e
a
nd
e
xplo
r
e
s
it
s
a
dva
ntage
f
a
c
tor
s
to
bui
ld
the
pr
opos
e
d
t
r
a
c
king
s
olut
ion.
T
he
c
onve
nti
ona
l
YO
L
O
ne
twor
k
ba
s
ica
ll
y
c
ons
is
ts
of
24
c
onvolut
ion
laye
r
s
a
nd
2
f
ull
y
c
onne
c
ted
laye
r
s
.
I
n
it
s
a
dva
nc
e
ve
r
s
ion
o
f
YO
L
Ov2,
the
im
p
r
ove
ment
c
a
n
be
s
e
e
n
ove
r
s
pe
e
d
a
nd
de
tec
ti
on
a
c
c
ur
a
c
y
whe
r
e
maj
or
tas
ks
include
s
ix
major
c
omput
a
ti
ona
l
s
teps
viz.
i)
ba
tch
nor
maliza
ti
on,
ii
)
h
igh
r
e
s
olut
ion
c
las
s
if
ier
,
ii
i)
c
onvolut
ion
with
a
nc
hor
boxe
s
,
iv)
s
ize
a
nd
a
s
p
e
c
t
r
a
ti
o
pr
e
diction
of
a
nc
hor
box
,
v)
a
na
lys
is
of
f
ine
-
gr
a
ined
f
e
a
tur
e
s
a
nd
vi
)
mul
ti
-
s
c
a
le
tr
a
ini
ng
pha
s
e
[
23
]
.
2.
2.
YOL
Ov4
T
he
ba
s
e
li
ne
de
s
ign
idea
o
f
YO
L
Ov4
is
a
one
-
s
tage
de
tec
ti
on
ne
twor
k
modeling
ba
s
ica
ll
y
c
ons
is
ts
of
thr
e
e
c
or
e
s
e
gments
whic
h
include
ba
c
kbone
,
ne
c
k
a
nd
he
a
d.
I
n
the
de
s
ign
of
the
p
r
opos
e
d
s
tr
a
tegy
of
M
OT
,
the
ba
c
kbone
of
YO
L
Ov4
c
a
n
be
a
pr
e
-
tr
a
ined
C
NN
s
uc
h
a
s
VG
G16
or
C
S
P
Da
r
kNe
t53
whic
h
tr
a
ined
on
the
r
e
s
pe
c
ti
ve
da
tas
e
ts
.
He
r
e
in
the
pr
opos
e
d
s
tudy
c
ontext
the
ba
c
kbone
pa
r
t
o
f
YO
L
Ov4
model
e
na
bles
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5681
-
5690
5684
f
e
a
tur
e
e
xtr
a
c
ti
on
ne
twor
k
whic
h
is
s
uit
a
ble
f
or
o
bjec
t
de
tec
ti
on.
I
n
f
e
a
tur
e
e
xt
r
a
c
ti
on
s
tage
the
s
tudy
model
c
omput
e
s
f
e
a
tur
e
maps
(
ƒ
)
whic
h
is
obtaine
d
f
r
om
the
input
im
a
ge
f
r
a
mes
of
pe
de
s
tr
ian
da
tas
e
ts
.
T
he
YO
L
Ov4
ne
twor
k
incor
po
r
a
tes
C
S
P
Da
r
kNe
t53
modul
e
a
s
the
ba
c
kbone
whic
h
pe
r
f
or
ms
f
e
a
tur
e
e
xtr
a
c
ti
on
f
r
om
the
input
im
a
ge
f
r
a
mes
.
T
he
ba
c
kbone
ba
s
ica
ll
y
c
onne
c
ts
with
f
ive
r
e
s
idual
lay
e
r
s
.
T
he
pr
opos
e
d
M
OT
f
r
a
mew
or
k
a
ls
o
int
e
gr
a
tes
YO
L
Ov4
s
pa
ti
a
l
pyr
a
mi
d
pooli
ng
(
S
P
P
)
laye
r
[
25]
a
nd
pa
th
a
ggr
e
ga
ti
on
ne
twor
k
(
P
AN
)
[
26]
.
He
r
e
S
P
P
in
n
e
c
k
pa
r
t
ba
s
ica
ll
y
e
na
bles
c
onc
a
tena
ti
on
o
f
max
-
pooli
ng
output
s
.
He
r
e
the
max
-
pooli
ng
output
s
a
r
e
c
ons
tr
uc
ted
in
the
f
or
m
of
low
-
r
e
s
olut
ion
f
e
a
tur
e
map
to
not
only
e
xtr
a
c
t
the
mos
t
r
e
p
r
e
s
e
ntative
f
e
a
tur
e
s
,
bu
t
i
t
a
ls
o
he
lps
in
f
e
a
tur
e
dim
e
ns
ionalit
y
r
e
duc
ti
on
f
r
om
c
o
mput
ing
point
of
view
.
T
he
ba
c
kbone
C
S
P
Da
r
kNe
t53
is
d
e
s
igned
ba
s
e
d
on
c
or
e
c
r
os
s
-
s
t
a
ge
pa
r
ti
a
l
ne
twor
k
f
e
a
tur
e
s
.
He
r
e
the
output
o
f
m
-
laye
r
C
NN
be
e
xp
r
e
s
s
e
d
a
s
(
1)
:
y
=
F
(
x
0
)
=
x
k
=
H
k
(
x
k
−
1
)
,
H
k
−
1
(
x
k
−
2
)
,
H
k
−
2
(
x
k
−
3
)
,
…
…
H
1
(
x
0
)
,
x
0
)
(
1)
He
r
e
in
(
1)
(
∙
)
de
notes
the
mapping
f
unc
ti
on
whe
r
e
a
s
0
de
notes
the
input
to
tar
ge
t
.
He
r
e
is
a
n
ope
r
a
ti
ona
l
f
unc
ti
on
f
o
r
the
th
laye
r
of
the
C
NN
.
How
e
ve
r
,
the
c
or
e
e
mphas
ize
is
to
opti
mi
z
e
the
f
unc
ti
on
a
t
dif
f
e
r
e
nt
laye
r
.
I
t
c
a
n
be
a
c
hieve
d
with
the
e
xpr
e
s
s
ion
(
2)
.
y
=
M
(
[
x
0
′
,
T
(
F
(
x
0
′′
)
]
)
(
2)
He
r
e
f
or
opti
mi
z
a
ti
on
modeling
0
is
r
e
pr
e
s
e
nted
int
o
two
dis
ti
nc
t
pa
r
ti
ti
ons
whic
h
c
a
n
be
s
hown
a
s
0
=
[
0
′
,
0
′′
]
.
He
r
e
is
the
tr
a
ns
it
ion
f
unc
ti
on
to
tr
unc
a
te
the
g
r
a
diant
f
lows
of
whe
r
e
a
s
is
a
nother
tr
a
ns
it
ion
f
unc
ti
on
whic
h
is
us
e
d
to
mi
x
two
s
e
gmente
d
pa
r
ts
.
How
e
ve
r
,
in
the
pr
opos
e
d
s
tudy
c
o
ntext
the
YO
L
Ov4
ba
c
kbone
c
ons
is
ts
of
f
ive
r
e
s
idual
block
s
a
nd
output
f
r
om
the
r
e
s
idual
b
locks
a
r
e
a
ggr
e
ga
t
e
d
a
t
the
ne
c
k
of
YO
L
Ov4
ne
twor
k.
T
he
de
s
ign
of
S
P
P
laye
r
[
1]
in
the
pr
opos
e
d
s
tudy
f
oll
ows
a
c
onf
igur
a
ti
o
n
s
e
tt
ing
whe
r
e
it
incor
por
a
tes
ke
r
ne
ls
of
s
ize
1
×
1,
5
×
5,
9
×
9
,
a
nd
13
×
13
whic
h
a
r
e
s
ubjec
ted
to
pe
r
f
o
r
m
max
-
pooing
ope
r
a
ti
ons
whe
r
e
the
c
r
os
s
-
s
tage
pa
r
ti
a
l
ne
twor
k
a
ls
o
c
ontr
ibut
e
s
towa
r
ds
im
pr
oving
the
lea
r
n
a
b
il
it
y
a
nd
pa
r
a
mete
r
uti
li
z
a
ti
on.
He
r
e
in
the
pr
opos
e
d
s
tudy
c
ontext
the
s
tr
ide
va
lue
is
s
e
t
to
1.
T
he
s
ys
tem
f
r
a
mew
or
k
int
e
gr
a
ti
on
f
or
mul
ti
-
tar
ge
t
tr
a
c
king
a
ls
o
c
onf
i
gur
e
s
the
YO
L
Ov4
ne
twor
k
in
s
uc
h
a
wa
y
whe
r
e
it
c
onc
a
tena
tes
the
(
ƒ
)
c
ons
ider
ing
the
S
P
P
pooli
ng
ope
r
a
ti
ons
.
He
r
e
the
a
dva
ntage
o
f
e
mpl
oying
c
onc
a
tena
ti
on
f
unc
ti
on
is
that
it
inc
r
e
a
s
e
s
the
r
e
c
e
pti
ve
f
ield
of
ba
c
kbone
f
e
a
tur
e
c
omponents
a
nd
a
ls
o
incr
e
a
s
e
s
the
a
c
c
ur
a
c
y
of
YO
L
Ov4
ne
twor
k
towa
r
ds
de
tec
ti
on
of
s
maller
objec
ts
.
T
he
c
onc
a
tena
ted
f
e
a
tur
e
maps
ge
ne
r
a
ted
f
r
om
the
YO
L
Ov4
S
P
P
modu
le
w
hich
is
r
e
pr
e
s
e
nted
with
ƒ
a
r
e
f
u
r
ther
f
us
e
d
with
the
high
-
r
e
s
olut
ion
f
e
a
tur
e
maps
ƒ
us
ing
P
AN
a
s
(
3)
:
ƒ
m
a
p
c
om
b
←
ƒ
m
a
p
S
P
P
⊕
ƒ
m
a
p
HR
(
3)
He
r
e
the
P
AN
c
omponent
ba
s
ica
ll
y
e
mpl
oys
up
s
a
mpl
ing
a
nd
downs
a
mpl
ing
ope
r
a
ti
ons
to
s
e
t
bott
o
m
-
up
a
nd
top
-
down
pa
ths
whic
h
f
ur
ther
c
onc
a
tena
te
low
-
leve
l
a
nd
high
-
leve
l
f
e
a
tur
e
s
.
F
ur
ther
the
P
AN
s
e
gment
e
na
bles
the
output
of
a
ggr
e
ga
ted
ƒ
whic
h
is
f
u
r
ther
us
e
d
f
or
pr
e
diction
.
He
r
e
in
the
p
r
opos
e
d
s
tage
the
YO
L
Ov4
ba
s
ica
ll
y
c
ons
is
ts
of
thr
e
e
dif
f
e
r
e
nt
pr
e
diction
he
a
ds
whic
h
a
r
e
YO
L
Ov3
ne
twor
ks
th
a
t
f
ur
the
r
ge
ne
r
a
te
the
f
inal
p
r
e
dictions
.
He
r
e
the
YO
L
Ov4
ne
twor
k
output
s
the
ƒ
of
s
ize
s
19
×
19,
38
×
38
,
a
nd
76
×
76
f
or
the
pur
pos
e
of
pr
e
dictive
a
na
lyt
ics
.
2.
3
.
M
u
lt
i
-
t
ar
ge
t
t
r
ac
k
in
g
s
t
r
at
e
gy
I
n
the
pr
opos
e
d
c
onc
e
pt
of
mu
lt
i
-
objec
t
or
tar
ge
t
tr
a
c
king
poli
c
y
int
e
gr
a
tes
the
YO
L
Ov4
f
or
the
pur
pos
e
of
pr
e
c
is
e
objec
t
de
tec
ti
on.
F
u
r
ther
it
a
im
s
to
pe
r
f
or
m
mu
lt
i
-
tar
ge
t
t
r
a
c
king
e
ve
n
in
the
pr
e
s
e
nc
e
of
oc
c
lus
ions
a
nd
ye
t
a
c
c
ompl
is
he
s
higher
c
omput
a
ti
ona
l
e
f
f
icie
nc
y
f
r
om
the
pe
r
s
pe
c
ti
ve
of
e
xe
c
uti
on.
T
he
pr
opos
e
d
c
ontext
of
tr
a
c
king
-
by
-
de
tec
ti
on
a
ppr
oa
c
h
ini
ti
a
ll
y
obtains
the
de
tec
ti
ons
of
objec
ts
in
e
a
c
h
f
r
a
me
a
nd
f
ur
ther
it
t
r
a
c
ks
the
a
s
s
oc
i
a
ti
on
a
nd
mana
ge
m
e
nt
a
c
r
os
s
the
video
f
r
a
me
s
e
que
n
c
e
s
.
T
he
s
tudy
a
ddr
e
s
s
e
s
the
c
omput
a
ti
ona
l
c
ompl
e
xit
y
c
ha
ll
e
nge
s
in
e
xis
ti
ng
a
ppr
oa
c
he
s
a
nd
a
ls
o
a
im
s
to
bu
il
d
a
n
e
f
f
icie
nt
tr
a
c
king
model
that
c
a
n
e
a
s
il
y
de
a
l
with
the
oc
c
lus
ions
.
I
n
thi
s
r
e
ga
r
d
a
long
with
YO
L
Ov4,
it
a
ls
o
int
e
gr
a
tes
the
pe
r
s
on
r
e
-
identif
ica
ti
on
ne
twor
ks
towa
r
ds
c
omput
ing
the
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
f
r
om
the
video
f
r
a
mes
.
He
r
e
the
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
(
)
a
r
e
a
ls
o
c
a
ll
e
d
a
s
a
ppe
a
r
a
nc
e
e
mbeddings
(
)
whic
h
ba
s
ica
ll
y
r
e
pr
e
s
e
nts
the
vis
ua
l
a
ppe
a
r
a
nc
e
of
a
n
objec
t
withi
n
the
f
r
a
me.
He
r
e
the
im
por
tanc
e
of
is
that
it
of
f
e
r
s
a
n
a
ddit
ional
mea
s
ur
e
of
s
im
il
a
r
it
y
f
a
c
tor
(
o
r
dis
tanc
e
)
be
twe
e
n
the
de
te
c
ti
ons
a
nd
tr
a
c
k
or
li
ne
of
moveme
nt
(
L
oM
)
.
He
r
e
the
pr
opos
e
d
s
tr
a
tegy
of
mul
ti
-
objec
t
of
in
ter
e
s
t
tr
a
c
king
uti
li
z
e
s
the
a
s
s
oc
iation
metr
ic
c
onc
e
pt
whe
r
e
it
a
ls
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
c
os
t
-
e
ff
e
c
ti
v
e
,
r
e
li
able
and
ac
c
ur
ate
fr
ame
w
or
k
for
multi
ple
-
tar
ge
t
tr
ac
k
ing
by
…
(
Div
y
apr
abha
)
5685
a
dopts
the
s
ingl
e
hypothes
is
tr
a
c
king
methodology
us
ing
KF
a
ppr
oa
c
h
[
27]
.
He
r
e
the
s
tudy
a
ls
o
int
e
gr
a
tes
the
a
ppe
a
r
a
nc
e
inf
or
mation
int
o
the
f
r
a
me
-
by
-
f
r
a
me
da
ta
a
s
s
oc
iation
to
ha
ndle
the
tr
a
c
king
ove
r
oc
c
lus
ion.
T
his
a
ppr
oa
c
h
a
ls
o
mi
nim
ize
s
the
number
of
s
witche
s
in
tr
a
c
k
identit
ies
.
T
he
pr
opos
e
d
f
r
a
mew
or
k
o
f
mu
lt
i
tar
ge
t
tr
a
c
king
a
ppr
oa
c
h
ba
s
ica
ll
y
r
e
li
e
s
on
thr
e
e
d
if
f
e
r
e
nt
types
o
f
dis
tanc
e
s
f
or
a
ll
oc
a
ti
ons
s
tage
y
w
hich
a
r
e
bounding
box
I
oU,
M
a
ha
lanobis
dis
tanc
e
,
a
nd
a
ppe
a
r
a
nc
e
c
os
ine
dis
tan
c
e
.
I
n
the
pr
opos
e
d
mu
lt
i
-
tar
ge
t
tr
a
c
king
model,
the
s
tudy
ba
s
ica
ll
y
f
or
mul
a
tes
two
dis
ti
nc
t
br
a
nc
he
s
whic
h
a
r
e
a
ppe
a
r
a
nc
e
br
a
nc
h
a
nd
mot
ion
br
a
nc
h
r
e
s
pe
c
ti
ve
ly.
He
r
e
the
p
r
opos
e
d
s
tudy
ini
t
ially
f
or
mul
a
tes
a
f
unc
ti
ona
l
block
f
or
tr
a
c
k
ha
nd
li
ng
a
nd
s
tate
a
ppr
oxim
a
ti
on.
T
he
pr
opos
e
d
s
tr
a
tegy
e
na
bles
the
tr
a
c
ke
r
with
di
f
f
e
r
e
nt
f
e
a
tur
e
s
viz.
i)
e
s
ti
mati
on
f
il
ter
,
ii
)
a
s
s
oc
iation
c
os
t
,
ii
i
)
a
s
s
oc
iation
type
,
a
nd
iv)
tr
a
c
k
maintena
nc
e
.
2.
4
.
I
n
c
lu
s
ion
o
f
e
s
t
im
a
t
ion
f
il
t
e
r
a
n
d
t
r
ac
k
as
s
oc
iat
ion
wit
h
m
e
as
u
r
e
s
T
he
pr
opos
e
d
s
tr
a
tegy
ini
ti
a
ll
y
de
f
ines
KF
by
mea
ns
of
int
r
oduc
ing
a
de
tec
ti
on
mea
s
ur
e
ment
whic
h
is
r
e
pr
e
s
e
nted
with
a
2D
bounding
box
.
He
r
e
the
de
tec
ti
on
mea
s
ur
e
ne
t
f
or
KF
c
a
n
be
r
e
p
r
e
s
e
nted
with
the
(
4)
.
=
[
,
,
,
ℎ
]
(
4)
He
r
e
de
notes
the
de
tec
ti
on
mea
s
ur
e
ment
f
or
2D
b
ounding
box.
Als
o,
he
r
e
(
,
)
r
e
pr
e
s
e
nts
the
c
oor
dinate
s
in
the
f
or
m
of
pixels
c
or
r
e
s
pond
to
the
top
-
lef
t
c
or
ne
r
of
bounding
box
whe
r
e
a
s
(
,
ℎ
)
r
e
pr
e
s
e
nts
the
wi
dth
a
nd
he
ight
o
f
the
bounding
box.
How
e
ve
r
,
de
f
in
ing
the
s
tate
e
s
ti
mation
f
i
lt
e
r
f
or
e
ight
-
dim
e
ns
ional
s
pa
c
e
us
ing
a
s
tanda
r
d
KF
f
or
bounding
box
tr
a
c
king
c
a
n
be
r
e
pr
e
s
e
nted
a
s
(
,
,
,
ℎ
,
,
̇
̇
,
̇
,
ℎ
̇
)
.
He
r
e
(
,
)
r
e
pr
e
s
e
nts
the
bounding
box
pos
it
ion
in
c
e
nter
,
r
e
pr
e
s
e
nts
a
s
p
e
t
r
a
ti
o
a
nd
ℎ
r
e
pr
e
s
e
nts
he
ight
a
long
with
their
r
e
s
pe
c
ti
ve
ve
lociti
e
s
in
im
a
ge
c
oor
dinate
s
.
How
e
ve
r
,
in
s
tanda
r
d
KF
e
s
ti
mation
model
the
bounding
c
oor
dinate
s
s
uc
h
a
s
(
,
,
,
ℎ
)
a
r
e
take
n
a
s
a
dir
e
c
t
obs
e
r
va
ti
on
of
the
objec
t
s
tate
.
How
e
ve
r
,
in
the
p
r
opos
e
d
s
tr
a
tegy
the
e
s
ti
mation
of
s
tate
of
e
s
ti
mate
d
bound
ing
box
c
a
n
be
mea
s
ur
e
d
with
the
f
oll
owing
math
e
matica
l
e
xpr
e
s
s
ion.
X
=
[
u
,
v
,
s
,
r
,
u
,
̇
v
̇
,
s
̇
,
r
̇
]
(
5)
He
r
e
r
e
pr
e
s
e
nts
the
s
c
a
le
o
f
bounding
box
a
nd
i
mpl
ies
width
to
he
ight
r
a
t
io
a
s
s
oc
iate
d
with
the
b
ouding
box.
How
e
ve
r
,
unli
ke
the
e
xis
ti
ng
KF
ba
s
e
d
s
tate
e
s
ti
mation
s
tr
a
tegy,
he
r
e
the
s
tate
e
s
ti
mation
mea
s
ur
e
ment
in
c
onve
r
s
ion
c
a
n
be
r
e
p
r
e
s
e
nted
with
the
ti
me
r
a
t
e
of
c
ha
nge
a
s
s
oc
iate
d
with
the
a
s
pe
c
t
r
a
ti
o.
T
he
f
oll
owing
mathe
matica
l
e
xpr
e
s
s
ions
a
r
e
f
or
mul
a
ted
to
pe
r
f
or
m
c
onve
r
s
ion
of
mea
s
ur
e
ment
to
s
tate
whic
h
a
r
e
a
s
(
6)
-
(
9)
:
u
=
∑
x
,
w
2
⁄
(
6)
v
=
∑
y
,
h
2
⁄
(
7)
s
=
w
∗
h
(
8)
r
=
w
h
⁄
(
9)
He
r
e
it
ha
s
to
be
noted
that
thes
e
e
qua
ti
ons
a
r
e
no
n
-
li
ne
a
r
,
s
o
the
c
onve
r
s
ion
is
pe
r
f
or
med
a
s
pr
e
-
pr
oc
e
s
s
ing
s
tage
in
the
pr
opos
e
d
method.
He
r
e
the
KF
mea
s
u
r
e
ment
f
unc
ti
on
is
a
ls
o
us
e
d
a
s
e
xpli
c
it
f
unc
ti
on
t
o
ha
ndle
the
non
-
li
ne
a
r
it
y.
He
r
e
in
thi
s
c
omput
a
ti
ona
l
a
ppr
oa
c
h
the
s
tr
a
tegy
c
ons
ider
s
pe
de
s
tr
ian
tr
a
c
king
da
tas
e
t
a
nd
f
ur
ther
ge
ne
r
a
tes
a
matr
ix
f
or
m
of
s
tor
a
ge
c
o
r
r
e
s
pond
to
de
tec
ti
on
r
e
c
or
ds
[
]
.
He
r
e
the
mat
r
ix
f
or
m
of
[
]
a
r
e
ge
r
e
nr
a
ted
us
ing
YO
L
Ov4
objec
t
de
tec
tor
whi
c
h
is
a
l
r
e
a
dy
de
f
ined
in
s
ub
s
e
c
ti
on
2
.
2.
He
r
e
r
e
f
e
r
s
to
the
de
tec
ti
on
indexe
s
s
tor
e
d
in
[
]
.
He
r
e
the
pr
opos
e
d
s
tr
a
tegy
f
ur
ther
pe
r
f
o
r
ms
c
onve
r
s
ion
of
[
]
c
ons
ider
ing
the
←
[
]
of
bounding
box
c
oo
r
dinate
s
to
[
,
,
,
ℎ
]
.
I
n
thi
s
ope
r
a
ti
on
,
the
s
ys
tem
incor
por
a
tes
a
n
e
xpli
c
it
f
unc
ti
on
of
bounding
box
de
tec
ti
ons
(
∙
)
:
pa
s
s
in:
[
]
,
.
He
r
e
r
e
f
e
r
s
to
the
c
ova
r
ianc
e
e
s
ti
mation
f
or
s
tanda
r
d
de
viation
of
1
.
T
he
ope
r
a
ti
on
o
f
bounding
box
c
onve
r
s
ion
c
on
s
ider
ing
YO
L
Ov4
objec
t
de
tec
ti
ons
c
a
n
be
r
e
pr
e
s
e
nted
a
s
(
10)
.
[
,
,
,
ℎ
]
←
(
[
]
,
)
(
10)
T
he
s
tate
tr
a
ns
it
ion
pr
inciple
in
the
pr
opos
e
d
s
tr
a
t
e
gy
is
buil
t
on
the
top
of
c
ons
tant
ve
locity
model.
How
e
ve
r
,
the
pr
opos
e
d
s
tudy
a
ls
o
talks
a
bout
t
r
a
c
k
ha
ndli
ng
s
c
he
ma
c
ons
ider
ing
li
ne
a
r
c
ons
tant
-
ve
locity
KF
f
or
bounding
box
t
r
a
c
king.
T
he
idea
o
f
KF
is
us
e
d
in
thi
s
pr
opos
e
d
a
ppr
oa
c
h
in
s
uc
h
a
wa
y
whe
r
e
it
t
r
a
c
ks
a
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5681
-
5690
5686
objec
t
mos
tl
y
pe
de
s
tr
ian
e
nti
ty
us
ing
a
s
e
que
nc
e
of
de
tec
ti
ons
or
mea
s
ur
e
ments
to
e
s
ti
mate
the
s
ta
te
of
the
objec
t
ba
s
e
d
on
the
mot
ion
model
a
s
s
oc
iate
d
with
t
ha
t
objec
t.
I
n
a
mot
ion
model
s
tate
r
e
pr
e
s
e
nts
c
oll
e
c
ti
on
of
va
r
ious
qua
nti
ti
e
s
by
whic
h
s
tatus
of
a
n
objec
t
c
a
n
be
r
e
pr
e
s
e
nted.
He
r
e
the
s
tate
c
ould
be
it
s
pos
it
ion,
ve
locity
or
a
c
c
e
ler
a
ti
on.
T
he
s
tate
tr
a
ns
it
ion
f
r
om
ti
mes
to
+
1
=
+
c
ons
ider
s
a
c
ons
tant
ve
locit
y
model
s
uc
h
a
s
(
11)
:
+
1
=
w
h
e
r
e
∈
(
11)
How
e
ve
r
,
in
the
e
xis
ti
ng
s
ys
tems
the
c
ons
tant
ve
l
oc
it
y
model
lac
ks
e
f
f
e
c
ti
ve
ne
s
s
in
de
s
c
r
ibi
ng
the
a
c
tual
mot
ion
of
the
objec
t
o
f
int
e
r
e
s
t
(
i
.
e
.
,
pe
de
s
tr
ians
)
in
video.
Als
o,
it
c
a
nnot
pr
ope
r
ly
ha
ndle
the
va
r
iations
a
mong
a
s
pe
c
t
r
a
ti
o
a
nd
a
r
e
a
s
tate
mea
s
ur
e
s
.
How
e
ve
r
,
the
p
r
opos
e
d
a
ppr
oa
c
h
of
M
OT
ha
ndles
thes
e
c
ons
tr
a
int
s
by
c
ons
tr
uc
ti
ng
a
n
e
nha
nc
e
d
f
unc
ti
on
of
(
∙
)
.
He
r
e
the
c
us
tom
ize
d
f
unc
ti
on
ini
ti
a
ll
y
e
na
bles
the
c
ons
tant
ve
locity
Ka
lm
a
n
f
il
ter
f
o
r
bounding
box
t
r
a
c
king
a
nd
a
ls
o
mea
s
ur
e
s
the
de
tec
ti
ons
in
a
s
pe
c
if
ic
f
o
r
mat.
F
ur
the
r
it
ini
ti
a
li
z
e
s
a
null
ve
locity
a
nd
a
ls
o
e
s
ti
mate
s
the
ve
locity
s
tate
s
by
mea
ns
of
c
ova
r
ianc
e
mea
s
ur
e
s
.
How
e
ve
r
,
the
s
ys
tem
a
ls
o
models
c
a
ptur
ing
unknown
a
c
c
e
ler
a
ti
on
thr
ough
whic
h
it
c
ons
tr
icts
the
KF
f
o
r
tr
a
c
king
c
ons
ider
in
g
both
mot
ion
model
a
nd
s
tate
tr
a
ns
it
ion
model
a
l
ong
with
mea
s
ur
e
ment
model.
T
he
pr
opos
e
d
s
tr
a
tegy
a
ls
o
f
ur
ther
e
va
luate
s
the
mea
s
ur
e
ment
a
s
s
oc
iation
(
)
f
or
e
a
c
h
tr
a
c
k
(
T
i
)
.
He
r
e
the
s
tr
a
tegy
a
s
s
oc
iate
s
the
t
r
a
c
k
with
a
p
pr
opr
iate
mea
s
ur
e
ments
.
He
r
e
e
a
c
h
de
tec
ti
on
s
hould
be
a
s
s
oc
iate
d
with
the
e
xis
ti
ng
tr
a
c
ks
whic
h
a
r
e
ma
int
a
ined
by
the
mul
ti
-
objec
t
tr
a
c
ke
r
.
T
he
t
r
a
c
king
model
ba
s
ica
ll
y
hypothes
ize
s
a
n
objec
t
of
int
e
r
e
s
t
with
a
s
pe
c
if
ic
tr
a
c
k
va
lue
in
the
c
ur
r
e
nt
f
r
a
me.
F
u
r
ther
the
s
ys
tem
is
us
ing
e
s
ti
mate
s
it
s
bounding
box.
All
the
de
tec
ti
ons
a
nd
it
s
a
s
s
oc
i
a
ti
ve
tr
a
c
ks
a
r
e
f
ur
ther
loa
de
d
in
a
f
il
e
s
tr
uc
tur
e
c
a
ll
e
d
a
s
(
)
.
I
t
ha
s
to
be
a
ls
o
noted
th
a
t
dur
ing
Ka
lm
a
n
F
il
ter
pr
e
diction
,
the
c
ounter
o
f
tr
a
c
k
va
lue
is
incr
e
mente
d
a
nd
onc
e
the
t
r
a
c
k
is
a
s
s
oc
iate
d
with
the
mea
s
ur
e
ment
then
it
is
r
e
s
e
t
to
0.
He
r
e
the
s
ys
tem
a
ls
o
c
ons
ider
s
e
s
ti
mation
of
pr
e
de
f
in
e
d
maximum
a
ge
of
a
t
r
a
c
k
(
m
a
x
)
.
He
r
e
the
s
ys
tem
is
modele
d
in
s
uc
h
a
wa
y
whe
r
e
it
s
e
e
s
if
t
r
a
c
ks
ha
ve
e
xc
e
e
de
d
the
pr
e
de
f
ined
li
mi
t
of
m
a
x
or
not
if
it
e
xc
e
e
ds
then
it
c
ons
ider
s
that
the
tr
a
c
ks
mi
ght
ha
ve
lef
t
t
he
s
c
e
ne
a
nd
de
lete
d
f
r
om
the
tr
a
c
k
s
e
t
.
T
his
is
how
the
s
ys
tem
a
ls
o
e
ns
ur
e
s
memor
y
uti
li
z
a
ti
on
e
f
f
icie
n
c
y
dur
ing
tr
a
c
k
mana
ge
ment.
F
u
r
ther
the
KF
model
is
ini
ti
a
ted
f
or
ne
w
t
r
a
c
k
hypothes
e
s
whic
h
c
or
r
e
s
pond
to
mul
ti
ple
pe
de
s
tr
ian
de
tec
ti
on
whic
h
a
r
e
not
a
s
s
oc
iate
d
with
the
e
xis
ti
ng
t
r
a
c
ks
.
2.
5
.
F
or
m
u
lat
io
n
of
all
oc
at
io
n
s
m
e
t
r
ics
T
he
s
tudy
in
thi
s
pha
s
e
of
r
e
s
e
a
r
c
h
f
or
mul
a
tes
a
l
loca
ti
ons
pr
oblem
whe
r
e
it
a
ppli
e
s
the
poli
c
y
of
Hunga
r
ian
a
lgor
it
hm
to
buil
d
a
s
s
oc
iation
be
twe
e
n
pr
e
dicte
d
Ka
lm
a
n
s
tate
s
a
nd
ne
wly
a
r
r
ived
mea
s
u
r
e
s
.
T
he
pr
opos
e
d
s
tr
a
tegy
ba
s
ica
ll
y
ut
il
ize
s
thr
e
e
dif
f
e
r
e
nt
types
o
f
dis
tanc
e
s
f
or
a
ll
oc
a
ti
ons
s
tr
a
tegy
w
hich
a
r
e
:
bounding
box
int
e
r
s
e
c
ti
on
ove
r
union
(
B
B
I
OU
)
,
M
a
ha
lanobis
dis
tanc
e
mea
s
ur
e
,
a
nd
a
ppe
a
r
a
nc
e
c
os
ine
dis
tanc
e
.
He
r
e
in
the
pr
opos
e
d
a
ppr
oa
c
h
of
M
OT
,
the
dis
tanc
e
of
B
B
I
OU
r
e
f
e
r
s
to
the
mea
s
ur
e
be
twe
e
n
a
tr
a
c
k
a
nd
de
tec
ti
on
ba
s
e
d
on
the
ove
r
lap
r
a
ti
o
of
t
wo
bounding
boxe
s
whic
h
c
ould
be
de
f
ined
as
(
12)
.
=
1
−
⁄
(
12)
On
the
other
ha
nd,
the
mea
s
ur
e
of
M
a
ha
lanobis
dis
tanc
e
is
a
nother
e
va
luation
c
r
it
e
r
ia
thr
ough
whic
h
the
dis
tanc
e
be
twe
e
n
de
tec
ti
on
a
nd
t
r
a
c
ks
a
r
e
e
v
a
luate
d
c
ons
ider
ing
a
s
tatis
ti
c
a
l
e
va
luation
o
f
p
r
oba
bil
it
y
de
ns
it
y
f
unc
ti
ons
.
How
e
ve
r
,
the
e
va
luation
of
dis
tanc
e
M
a
h
a
lanobis
c
a
n
be
e
s
ti
mate
d
c
ons
ide
r
ing
the
f
oll
owing
e
xpr
e
s
s
ion.
ℎ
=
(
−
)
−
1
(
−
)
(
13)
He
r
e
r
e
pr
e
s
e
nts
the
mea
s
ur
e
s
a
s
s
oc
iate
d
with
bounding
box,
r
e
pr
e
s
e
nts
the
tr
a
c
k
s
tate
a
nd
r
e
f
e
r
s
to
J
a
c
obian
mea
s
ur
e
ment
f
unc
ti
ona
li
ty
f
o
r
dim
e
ns
ional
s
pa
c
e
mea
s
ur
e
.
On
the
other
ha
nd,
r
e
f
e
r
s
t
o
th
e
pr
e
dicte
d
mea
s
ur
e
ment.
He
r
e
de
notes
c
ova
r
ianc
e
matr
ix.
How
e
ve
r
,
the
s
tudy
a
ls
o
e
va
luate
s
the
dis
tanc
e
be
twe
e
n
the
de
tec
ti
on
a
nd
pr
e
dicte
d
tr
a
c
ks
c
on
s
i
de
r
ing
a
ppe
a
r
a
nc
e
f
e
a
tur
e
ve
c
tor
whic
h
is
a
ls
o
r
e
f
e
r
r
e
d
a
s
a
ppe
a
r
a
nc
e
c
os
ine
dis
tanc
e
.
2.
6
.
F
or
m
u
lat
io
n
of
c
os
t
m
e
t
r
ic
T
he
pr
opos
e
d
idea
in
thi
s
r
e
s
e
a
r
c
h
a
ppr
oa
c
h
ba
s
ica
ll
y
c
ombi
ne
s
both
ℎ
a
nd
a
ppe
a
r
a
nc
e
f
e
a
tur
e
c
os
ine
dis
tanc
e
to
e
s
ti
mate
the
s
e
t
of
ne
w
de
tec
ti
ons
to
the
s
e
t
of
c
ur
r
e
nt
li
ne
of
moveme
nt
(
L
oM
)
.
T
he
c
os
t
e
s
ti
mation
in
pr
opos
e
d
s
tudy
is
a
m
ini
mi
z
a
ti
on
pr
oblem
whic
h
is
f
o
r
mul
a
ted
a
s
(
14
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
c
os
t
-
e
ff
e
c
ti
v
e
,
r
e
li
able
and
ac
c
ur
ate
fr
ame
w
or
k
for
multi
ple
-
tar
ge
t
tr
ac
k
ing
by
…
(
Div
y
apr
abha
)
5687
=
∑
×
ℎ
,
(
1
−
)
×
(
)
(
14)
w
he
r
e
0
<
<
1
.
P
r
ope
r
t
r
a
c
k
mana
ge
ment
c
ould
a
ls
o
play
a
c
r
uc
ial
r
ole
in
the
pr
opos
e
d
s
tudy
to
ha
ndle
the
c
os
t
metr
ic.
T
his
a
pp
r
oa
c
h
of
c
os
t
e
va
luation
with
e
f
f
e
c
ti
ve
tr
a
c
k
mana
ge
ment
ba
s
ica
ll
y
ha
ndles
the
oc
c
lus
ion
in
the
pr
opos
e
d
t
r
a
c
king
model
.
I
n
thi
s
r
e
ga
r
d,
th
e
s
tudy
int
r
oduc
e
s
a
matc
hing
e
va
luation
s
olut
ion
that
pr
ior
it
ize
s
the
mo
r
e
f
r
e
que
ntl
y
obs
e
r
ve
d
objec
ts
ov
e
r
a
video
s
c
e
ne
to
a
s
s
e
s
s
the
a
s
s
oc
iation
li
ke
li
hood.
2.
7
.
YOL
Ov4
an
d
p
e
r
s
on
r
e
-
id
e
n
t
if
ica
t
ion
n
e
t
w
or
k
m
od
e
li
n
g
T
he
s
tudy
in
the
pr
opos
e
d
a
ppr
oa
c
h
o
f
M
OT
ba
s
ica
ll
y
c
ons
ider
s
a
da
tas
e
t
of
pe
de
s
tr
ian
tr
a
c
king
whe
r
e
the
de
tec
ti
ons
a
r
e
ge
ne
r
a
ted
f
r
o
m
the
YO
L
Ov4
objec
t
de
tec
tor
model
.
He
r
e
the
s
ys
tem
c
ons
ider
s
both
gr
ound
tr
uth
da
ta
a
nd
the
de
tec
ti
ons
f
r
o
m
the
p
r
e
-
tr
a
ined
YO
L
Ov4.
How
e
ve
r
,
the
p
r
opos
e
d
a
ppr
oa
c
h
a
ls
o
f
ur
ther
pe
r
f
o
r
ms
c
onve
r
s
ion
of
de
tec
ti
ons
to
bounding
box
c
oor
dinate
s
us
ing
(
11)
.
F
ur
ther
the
s
ys
tem
a
ls
o
int
e
gr
a
tes
a
pe
r
s
on
r
e
-
identif
ica
ti
on
ne
twor
k
whic
h
is
a
ls
o
a
pr
e
-
tr
a
ined
ne
twor
k
to
e
va
luate
the
a
p
pe
a
r
a
nc
e
f
e
a
tur
e
f
or
e
a
c
h
de
tec
ti
on.
I
t
ha
s
to
be
noted
that
f
or
e
xtr
a
c
ti
on
of
a
ppe
a
r
e
d
f
e
a
tur
e
ve
c
tor
c
o
r
r
e
s
pond
to
e
a
c
h
de
tec
ti
on,
the
s
ys
tem
c
ons
ider
s
bounding
box
c
oor
dinate
s
a
nd
c
onve
r
t
them
to
c
r
op
f
r
a
me.
F
in
a
ll
y,
the
a
ppe
a
r
a
nc
e
ve
c
tor
o
f
(
)
is
c
ons
tr
uc
ted
c
ons
ider
ing
a
pr
e
diction
s
tr
a
tegy
f
o
r
pe
r
s
on
r
e
-
identif
ica
ti
on
.
He
r
e
the
c
r
oppe
d
f
r
a
me
objec
t
of
int
e
r
e
s
t
im
a
ge
is
us
e
d
to
pr
e
dict
the
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
a
nd
the
f
inal
is
ge
ne
r
a
ted.
F
ur
ther
the
s
ys
tem
of
M
OT
c
ons
ider
s
thes
e
bounding
boxe
s
a
nd
a
nd
it
e
r
a
tes
them
to
e
va
l
ua
te
the
a
lgor
it
hm
1
a
nd
ge
ne
r
a
tes
a
ll
oc
a
ti
ons
whic
h
a
r
e
f
ur
ther
mana
ge
d
unde
r
L
oM
ini
ti
a
ti
on,
de
le
ti
on
a
nd
a
ppe
a
r
a
nc
e
f
e
a
tur
e
upda
te
.
T
he
tr
a
c
ke
r
a
ls
o
e
va
l
ua
tes
KF
to
p
r
e
dict
t
r
a
c
ks
whic
h
a
r
e
f
ur
ther
u
ti
l
ize
d
f
or
ge
ne
r
a
ti
ng
a
ll
oc
a
ti
ons
.
F
inally
with
thes
e
opti
mi
z
e
d
it
e
r
a
ti
on
s
teps
the
s
ys
tem
pe
r
f
or
ms
tr
a
c
king
of
p
e
de
s
tr
ian
moveme
nts
.
F
ur
ther
outcome
o
f
the
p
r
opos
e
d
tr
a
c
king
model
is
e
va
luate
d
c
ons
ider
ing
dif
f
e
r
e
nt
pe
r
f
or
manc
e
pa
r
a
mete
r
s
whic
h
a
r
e
mul
ti
-
objec
t
t
r
a
c
king
a
c
c
ur
a
c
y
(
M
OT
A)
a
nd
mul
ti
-
objec
t
tr
a
c
king
p
r
e
c
is
ion
(
M
OT
P
)
.
Algor
it
hm
1
.
S
t
r
a
tegy
f
o
r
c
os
t
-
e
f
f
e
c
ti
ve
a
ll
oc
a
ti
on
of
tr
a
c
ks
a
nd
t
r
a
c
k
mana
ge
ment
s
c
he
me
Input: Track Indices
(
)
,
Detections (
[
]
)
,
ma
x
Output: Track Management
for minimum cost evaluation
Begin
1.
Video Sequence Exploration Frame Reading
2.
Apply YOLOv4, Person Re
-
Identification
3.
Initialize unallocated detections (
)
4.
Compute:
ℎ
,
(
)
for
and
5.
Evaluate cost using (14)
6.
Apply Linear Allocations of cost
7.
Update all Allocations and unallocated tracks
8.
Update unallocated detection
9.
Initialize
Predicted
10.
Formulate
for last
(
−
1
)
frame instances
11.
For
(
=
1
)
a.
Evaluate (Line
-
4 to Line
-
5)
b.
For
(
=
+
1
)
i.
Compute
cost for remining unallocated detections and tracks
ii.
Linear Allocations
c.
If (
(
≤
ma
x
)
d.
Evaluate (Line
-
11, Line
-
11a)
12.
Update cumulative Allocations with
unallocated detections and tracks
13.
End of For
End
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
he
pr
opos
e
d
s
tudy
c
ons
ider
s
e
va
luation
of
th
e
M
OT
f
r
a
mew
or
k
to
jus
ti
f
y
the
p
r
opos
it
ions
a
ppli
c
a
bil
it
y
in
r
e
a
l
-
ti
me
mul
ti
objec
t
tr
a
c
king.
He
r
e
the
pr
opos
e
d
tr
a
c
king
wor
kf
low
model
c
ons
ider
s
pe
de
s
tr
ian
tr
a
c
king
video
da
ta
f
or
the
pur
pos
e
of
e
va
luation.
T
he
s
tudy
f
ur
the
r
a
ls
o
c
ons
ider
s
objec
t
-
or
iente
d
a
ppr
oa
c
h
to
e
va
luate
the
M
OT
tr
a
c
ke
r
f
r
a
mew
or
k.
He
r
e
the
wor
k
-
f
low
f
o
r
the
e
va
luation
of
the
pr
opos
e
d
M
OT
tr
a
c
ke
r
model
ini
ti
a
ll
y
c
ons
tr
uc
ts
a
L
oM
log
a
nd
f
ur
ther
ini
ti
a
li
z
e
s
a
mi
nim
um
s
c
or
e
f
or
de
t
e
c
ti
on
of
objec
t
of
int
e
r
e
s
t.
F
ur
ther
it
e
na
bles
a
pr
oc
e
s
s
ing
l
ooping
s
tr
uc
tur
e
to
s
igni
f
ica
nt
s
c
or
e
of
de
tec
ti
ons
whic
h
a
r
e
f
ur
ther
us
e
d
to
r
un
the
P
e
r
s
on
re
-
identif
ica
ti
on
ne
twor
k
model.
He
r
e
the
t
r
a
c
ke
r
c
ons
ider
s
high
-
s
c
or
e
de
tec
ti
ons
to
pr
e
dict
the
tr
a
c
king
o
f
L
oM
f
or
the
d
e
tec
ted
pe
de
s
tr
ian
moveme
nt
on
the
s
tr
e
e
t
.
T
he
s
tudy
a
ls
o
maintains
a
r
e
c
or
d
f
or
the
log
o
f
L
oM
whic
h
is
f
u
r
ther
a
ls
o
us
e
d
f
or
e
va
luation
of
the
pe
r
f
or
manc
e
met
r
ic.
How
e
ve
r
,
the
e
va
luation
m
e
a
s
ur
e
s
f
ur
ther
s
how
how
the
tr
a
c
king
outcome
in
the
pr
opos
e
d
s
tudy
model
is
c
ompar
a
ti
ve
ly
be
tt
e
r
than
the
e
xis
ti
ng
tr
a
c
king
models
.
I
n
thi
s
r
e
ga
r
d
the
pe
r
f
or
manc
e
e
va
luation
s
c
or
e
s
a
r
e
c
omput
e
d
in
th
e
mea
s
ur
e
of
M
OT
A
a
nd
M
OT
P
a
long
with
f
a
ls
e
pos
it
ive
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5681
-
5690
5688
(
F
P
)
,
f
a
ls
e
ne
ga
ti
ve
(
F
N
)
,
a
nd
r
e
c
a
ll
s
c
or
e
s
.
T
he
s
tudy
s
hows
that
the
pr
opos
e
d
wor
kf
low
model
f
or
t
he
M
OT
tr
a
c
king
a
c
c
ompl
is
he
s
c
ons
ider
a
ble
M
OT
A
s
c
or
e
whic
h
is
a
ppr
oxim
a
tely
85.
8%
whe
r
e
a
s
M
OT
P
s
c
or
e
is
f
ound
to
be
94.
5
%
.
I
t
a
ls
o
s
hows
that
the
f
a
ls
e
pos
it
ive
s
c
or
e
o
f
30
a
long
with
F
N
va
lue
of
62.
How
e
ve
r
,
the
r
e
c
a
ll
pe
r
f
or
manc
e
is
f
ound
to
be
90
%
whic
h
a
r
e
c
ompar
a
ble
with
the
e
xis
ti
ng
ba
s
e
li
ne
s
.
T
he
c
ompar
a
ti
ve
outcome
a
s
highl
ight
e
d
in
the
F
igu
r
e
1
c
lea
r
ly
s
hows
that
the
pr
opos
e
d
tr
a
c
king
model
outper
f
or
ms
the
ba
s
e
li
ne
a
ppr
oa
c
he
s
of
k
-
de
ns
e
ne
ighbor
s
ba
s
e
d
t
r
a
c
ke
r
(
KDNT
)
[
28]
,
s
im
ple
onli
ne
a
nd
r
e
a
l
-
ti
me
tr
a
c
king
(
S
OR
T
)
[
29]
a
nd
li
f
ted
mul
ti
c
ut
pr
oblem
(
L
M
P
)
[
30]
in
the
mea
s
ur
e
of
both
M
OT
A
a
nd
M
OT
P
.
T
he
pr
im
e
r
e
a
s
on
be
hind
thi
s
is
the
inclus
ion
of
c
os
t
-
e
f
f
i
c
ient
tr
a
c
k
Alloca
ti
ons
a
nd
mana
ge
ment
s
c
he
m
e
s
.
T
he
c
ompar
a
ti
ve
outcome
is
a
ls
o
vis
ua
li
z
e
d
thr
ough
F
i
gur
e
s
1(
a
)
a
nd
1
(
b)
.
T
he
t
r
a
dit
ional
popu
lar
M
OT
methods
by
[
28]
–
[
3
0]
a
ls
o
talk
a
bout
the
s
c
ope
of
mul
ti
ple
hypo
thes
is
(
M
HT
)
tr
a
c
king
a
long
with
pr
oba
bil
is
ti
c
da
ta
a
s
s
oc
iation
pr
oblem.
E
ve
n
though
thes
e
a
ppr
oa
c
he
s
a
r
e
f
ound
to
c
a
ter
the
r
e
quir
e
ments
f
or
tr
a
c
king
a
c
c
ur
a
c
y
whe
n
int
e
gr
a
ted
with
pr
uning
s
c
he
mes
,
but
the
e
f
f
icie
nt
tr
a
c
king
-
by
-
de
te
c
ti
on
c
omes
with
incr
e
a
s
e
d
c
om
putational
a
nd
de
s
ign
c
ompl
e
xit
y
f
or
a
s
s
oc
iatio
n
metr
ic
e
va
luation
unde
r
f
r
e
que
nt
oc
c
lus
ions
.
Unlike
e
xi
s
ti
ng
ba
s
e
li
ne
s
tr
a
tegie
s
of
M
OT
,
the
pr
opos
e
d
model
of
M
OT
c
ons
ider
ing
the
c
os
t
-
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
f
or
opti
mi
z
e
d
tr
a
jec
tor
y
f
or
mu
lation
a
ddr
e
s
s
e
s
thi
s
c
h
a
ll
e
nging
s
c
e
na
r
io
in
c
onve
nti
ona
l
a
ppr
oa
c
he
s
a
nd
im
pr
ovis
e
that
to
r
e
tain
we
ll
ba
lanc
e
be
twe
e
n
both
c
ompl
e
xit
y
a
nd
M
OT
A.
T
he
s
tudy
model
a
ls
o
c
laims
it
s
e
f
f
e
c
ti
v
e
ne
s
s
in
im
pleme
nti
ng
s
im
pli
f
ied
wo
r
k
-
f
low
de
s
i
gn
whic
h
make
s
the
ope
r
a
ti
ons
mor
e
c
os
t
-
e
f
f
e
c
ti
ve
a
nd
r
e
s
our
c
e
f
r
iendly
whe
n
c
ompar
e
d
with
the
ba
s
e
li
ne
s
tudi
e
s
of
M
OT
.
T
he
c
ompar
a
ti
ve
a
na
lys
is
a
ls
o
s
hows
that
the
e
f
f
icie
nt
t
r
a
c
k
mana
ge
ment
s
c
he
me
in
the
pr
opos
e
d
s
ys
tem
not
only
im
p
r
ove
s
it
s
ope
r
a
ti
ona
l
a
nd
de
s
ign
e
f
f
icie
nc
y
f
r
om
the
c
os
t
o
f
c
omput
a
ti
on
point
of
view
but
a
ls
o
e
ns
ur
e
s
higher
r
e
li
a
bil
it
y
in
e
ns
ur
ing
be
t
ter
mul
ti
p
le
tar
ge
t
tr
a
c
king
a
c
c
ur
a
c
y
a
nd
p
r
e
c
is
ion
whic
h
make
s
it
mor
e
s
uit
a
ble
to
be
im
pleme
nted
ove
r
r
e
a
l
-
ti
me
s
c
e
na
r
ios
.
(
a
)
(
b)
F
igur
e
1.
C
ompar
a
ti
ve
a
na
lys
is
(
a
)
M
OT
A
(
%
)
a
n
d
(
b)
M
OT
P
(
%
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
c
os
t
-
e
ff
e
c
ti
v
e
,
r
e
li
able
and
ac
c
ur
ate
fr
ame
w
or
k
for
multi
ple
-
tar
ge
t
tr
ac
k
ing
by
…
(
Div
y
apr
abha
)
5689
4.
CONC
L
USI
ON
T
he
p
r
opos
e
d
s
tudy
int
r
oduc
e
s
a
c
os
t
-
e
f
f
e
c
ti
ve
M
OT
f
r
a
mew
or
k
whe
r
e
it
int
e
gr
a
tes
bo
th
YO
L
Ov4
a
nd
P
e
r
s
on
re
-
identif
ica
ti
on
ne
twor
k
with
a
mul
ti
-
objec
t
tr
a
c
ke
r
de
s
ign
c
ons
ider
ing
the
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
to
im
pr
ove
the
pe
r
f
or
manc
e
of
pe
de
s
tr
ian
t
r
a
c
king.
He
r
e
the
s
ys
tem
f
o
r
mul
a
ti
on
a
ls
o
e
va
luate
s
pr
oblem
of
a
ll
oc
a
ti
ons
dis
tanc
e
s
whe
r
e
im
pleme
nts
a
uniq
ue
a
ppr
oa
c
h
of
a
ll
oc
a
ti
ons
s
tr
a
tegy
with
e
f
f
ici
e
nt
tr
a
c
k
mana
ge
ment
s
c
he
me
wh
e
r
e
it
a
ls
o
e
va
luate
s
the
c
os
t
c
ons
ider
ing
both
M
a
ha
l
a
nobis
c
o
s
t
a
nd
c
os
i
ne
c
os
t.
He
r
e
the
p
r
opos
e
d
objec
t
t
r
a
c
king
a
lgor
it
hm
de
s
ign
is
modele
d
a
s
s
im
ple
a
s
pos
s
ibl
e
f
r
om
the
de
s
ign
point
of
view
a
nd
a
ls
o
pe
r
s
on
r
e
-
id
e
nti
f
ica
ti
on
e
nti
f
ica
ti
on
de
e
p
lea
r
ning
model
he
r
e
c
ontr
ibut
e
s
towa
r
ds
ge
ne
r
a
ti
ng
de
e
p
a
ppe
a
r
a
nc
e
f
e
a
tur
e
s
whic
h
im
pr
ove
the
a
ll
oc
a
ti
ons
ha
ndli
ng
s
c
he
ma
with
tr
a
c
k
mana
ge
me
nt.
He
r
e
int
e
gr
a
ti
on
o
f
a
ppe
a
r
a
nc
e
inf
or
mation
int
o
da
ta
a
s
s
oc
iation
a
ls
o
ha
ndles
the
pr
oblem
of
oc
c
lus
ion
to
a
higher
e
xtent
ove
r
longer
pe
r
iod
o
f
ti
me.
T
he
outcome
o
f
the
s
tudy
s
hows
that
the
pr
opos
e
d
model
outper
f
or
ms
the
e
xis
ti
ng
ba
s
e
li
ne
tr
a
c
king
s
tr
a
tegie
s
with
85.
8%
M
OT
A
a
nd
94.
5%
M
OT
P
s
c
or
e
s
making
it
s
uit
a
ble
f
or
r
e
a
l
-
ti
me
tr
a
c
king
of
objec
t
of
int
e
r
e
s
t.
T
he
s
c
ope
of
the
s
tudy
li
e
s
in
modi
f
ying
the
M
a
ha
lanobis
dis
tanc
e
in
f
utur
e
s
o
that
the
dis
tr
ibut
ion
of
c
ova
r
ianc
e
c
a
n
be
ha
ndled
mor
e
a
ppr
opr
iate
ly
in
the
pr
e
s
e
nc
e
of
oc
c
l
us
ion.
RE
F
E
RE
NC
E
S
[
1]
L
.
J
ia
o
e
t
al
.
,
“
A
s
ur
ve
y
of
d
e
e
p
le
a
r
ni
ng
-
ba
s
e
d
obj
e
c
t
de
te
c
ti
on,”
I
E
E
E
A
c
c
e
s
s
,
vol
.
7,
pp.
128837
–
128868,
2019,
doi
:
10.1109/AC
C
E
S
S
.2019.2939201.
[
2]
S
.
K
.
P
a
l,
A
.
P
r
a
ma
ni
k,
J
.
M
a
it
i,
a
nd
P
.
M
it
r
a
,
“
D
e
e
p
le
a
r
ni
n
g
in
mul
ti
-
obj
e
c
t
de
te
c
ti
on
a
nd
tr
a
c
ki
ng:
s
ta
te
of
th
e
a
r
t,
”
A
ppl
ie
d
I
nt
e
ll
ig
e
nc
e
, vol
. 51, no. 9, pp. 6400
–
6429, S
e
p. 2021, doi:
10.1007/s
10489
-
021
-
02293
-
7.
[
3]
B
o
Y
a
ng
a
nd
R
.
N
e
v
a
ti
a
,
“
M
ul
ti
-
ta
r
ge
t
tr
a
c
ki
ng
by
onl
in
e
le
a
r
ni
ng
of
non
-
li
ne
a
r
mot
io
n
pa
tt
e
r
ns
a
nd
r
obus
t
a
ppe
a
r
a
nc
e
mod
e
ls
,”
in
2012
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
,
J
un.
2012,
pp.
1918
–
1925,
doi
:
10.1109/C
V
P
R
.2012.6247892.
[
4]
S
.
K
.
P
a
l,
D
.
B
houmi
k,
a
nd
D
.
B
hu
ni
a
C
h
a
kr
a
bor
ty
,
“
G
r
a
nul
a
te
d
de
e
p
le
a
r
ni
ng
a
nd
Z
-
nu
mbe
r
s
i
n
mot
io
n
de
t
e
c
t
io
n
a
nd
o
bj
e
c
t
r
e
c
og
ni
ti
o
n,”
N
e
u
r
al
C
om
put
in
g an
d A
ppl
ic
at
io
n
s
, v
ol
.
32,
no.
2
1, pp
. 16
533
–
1654
8, N
ov.
2020
,
do
i:
10
.100
7/
s
0052
1
-
019
-
0420
0
-
1.
[
5]
D
.
B
.
C
ha
kr
a
bor
ty
a
nd
S
.
K
.
P
a
l,
G
r
anul
ar
v
id
e
o
c
om
put
in
g
w
it
h
r
ough
s
e
ts
,
de
e
p
le
ar
ni
ng
and
in
I
oT
.
W
or
ld
S
c
ie
nt
i
f
ic
,
2021.
[
6]
P
.
D
ol
la
r
,
C
.
W
oj
e
k,
B
.
S
c
hi
e
le
,
a
nd
P
.
P
e
r
ona
,
“
P
e
d
e
s
tr
ia
n
de
t
e
c
ti
on:
a
n
e
va
lu
a
ti
on
of
th
e
s
ta
te
of
th
e
a
r
t,
”
I
E
E
E
T
r
ans
ac
ti
ons
on
P
at
te
r
n A
nal
y
s
is
and M
ac
hi
ne
I
nt
e
ll
ig
e
nc
e
, vol
. 34, no. 4, pp. 7
43
–
761, Apr
. 2012,
doi
:
10.1109/T
P
A
M
I
.2011.155.
[
7]
A
.
G
e
ig
e
r
,
P
.
L
e
nz
,
a
nd
R
.
U
r
ta
s
un,
“
A
r
e
w
e
r
e
a
dy
f
or
a
ut
ono
mous
dr
iv
in
g?
th
e
K
I
T
T
I
vi
s
io
n
be
nc
hma
r
k
s
ui
te
,
”
in
2012
I
E
E
E
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
is
io
n and P
at
te
r
n R
e
c
ogni
ti
on
, J
un.
2012, pp. 3354
–
3361, doi:
10.1109/C
V
P
R
.2012.6248074.
[
8]
O
.
R
us
s
a
kov
s
ky
e
t
al
.
,
“
I
ma
ge
N
e
t
la
r
ge
s
c
a
le
vi
s
ua
l
r
e
c
ogni
ti
o
n
c
ha
ll
e
nge
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
C
om
put
e
r
V
is
io
n
,
vol
.
115,
no. 3, pp. 211
–
252, De
c
. 2015, doi:
10.1007/s
11263
-
015
-
0816
-
y.
[
9]
M
.
E
ve
r
in
gha
m,
L
.
V
a
n
G
ool
,
C
.
K
.
I
.
W
il
li
a
ms
,
J
.
W
in
n,
a
nd A
.
Z
is
s
e
r
ma
n,
“
T
h
e
pa
s
c
a
l
vi
s
ua
l
obj
e
c
t
c
la
s
s
e
s
(
V
O
C
)
c
ha
ll
e
n
ge
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
C
om
put
e
r
V
is
io
n
, vol
. 88, no. 2, pp. 30
3
–
338, J
un. 2010, doi:
10.1007/s
11263
-
009
-
0275
-
4.
[
10]
T.
-
Y
.
L
in
e
t
al
.
,
“
M
ic
r
os
of
t
C
O
C
O
:
c
omm
on
obj
e
c
ts
in
c
on
te
xt
,”
I
n:
F
le
e
t,
D
.,
P
a
jd
la
,
T
.,
S
c
hi
e
le
,
B
.,
T
uyt
e
la
a
r
s
,
T
.
(
e
ds
)
C
om
put
e
r
V
is
io
n
–
E
C
C
V
2014.
E
C
C
V
2014.
L
e
c
tu
r
e
N
ot
e
s
in
C
om
put
e
r
S
c
ie
nc
e
,
vol
.
8693.
S
pr
in
ge
r
,
C
ha
m.
2014,
pp.
740
–
7
55
,
doi
:
10.1007/978
-
3
-
319
-
10602
-
1_48
.
[
11]
A
.
K
uz
ne
ts
ova
e
t
al
.
,
“
T
he
ope
n
im
a
ge
s
da
ta
s
e
t
V
4,”
I
nt
e
r
nat
i
onal
J
our
nal
of
C
om
put
e
r
V
is
io
n
,
vol
.
128,
no.
7,
pp.
1956
–
19
81,
J
ul
. 2020, doi:
10.1007/s
11263
-
020
-
01316
-
z.
[
12]
A
.
B
oc
hkovs
ki
y,
C
.
Y
.
W
a
ng,
a
nd
H
.
Y
.
M
.
L
ia
o,
“
Y
O
L
O
v4:
opt
im
a
l
s
pe
e
d
a
nd
a
c
c
ur
a
c
y
of
obj
e
c
t
de
te
c
ti
on,”
a
r
X
iv
pr
e
p
r
in
t
ar
X
iv
:
2004.10934
, 2020.
[
13]
R
.
G
ir
s
hi
c
k,
J
.
D
ona
hue
,
T
.
D
a
r
r
e
ll
,
a
nd
J
.
M
a
li
k,
“
R
ic
h
f
e
a
tu
r
e
hi
e
r
a
r
c
hi
e
s
f
or
a
c
c
ur
a
te
obj
e
c
t
de
te
c
ti
on
a
nd
s
e
m
a
nt
ic
s
e
gme
nt
a
ti
on,”
in
2014
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
,
J
un.
2014,
pp.
580
–
587,
doi
:
10.1109/C
V
P
R
.2014.81.
[
14]
J
.
H
us
s
a
in
,
B
. R
.
P
r
a
th
a
p,
a
nd
A
.
S
ha
r
ma
,
“
A
n
im
pr
ove
d
a
nd
e
f
f
ic
ie
nt
Y
O
L
O
v4
me
th
od
f
or
obj
e
c
t
de
te
c
ti
on
in
vi
de
o
s
tr
e
a
mi
ng,”
I
n:
S
hukl
a
,
S
.,
G
a
o,
X
Z
.,
K
ur
e
e
th
a
r
a
,
J
.V
.,
M
i
s
hr
a
,
D
.
(
e
d
s
)
,
D
at
a
Sc
ie
nc
e
and
Se
c
ur
it
y
.
L
e
c
tu
r
e
N
ot
e
s
in
N
e
tw
or
k
s
and
S
y
s
te
m
s
,
vol
.
462. S
pr
in
ge
r
, S
in
ga
por
e
,
2022, pp. 305
–
316
, doi
:
10.1007
/9
78
-
981
-
19
-
2211
-
4_27
.
[
15]
J
.
R
e
dmon,
S
.
D
iv
va
la
,
R
.
G
ir
s
hi
c
k,
a
nd
A
.
F
a
r
ha
di
,
“
Y
ou
onl
y
lo
ok
onc
e
:
uni
f
ie
d,
r
e
a
l
-
ti
me
obj
e
c
t
de
te
c
ti
on,”
in
2016
I
E
E
E
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
is
io
n and P
at
te
r
n R
e
c
ogni
ti
on (
C
V
P
R
)
, J
un. 2016, pp. 779
–
788,
doi
:
10.1109/C
V
P
R
.2016.91.
[
16]
J
.
R
e
dmon
a
nd
A
.
F
a
r
ha
di
,
“
Y
O
L
O
9000:
be
tt
e
r
,
f
a
s
te
r
,
s
tr
onge
r
,”
in
2017
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on (
C
V
P
R
)
, J
ul
. 2017, pp. 6517
–
6525, doi:
10.1109/C
V
P
R
.2017.690.
[
17]
L
.
A
.
Z
a
de
h,
“
A
not
e
on
Z
-
nu
mbe
r
s
,”
I
nf
or
m
at
io
n
Sc
ie
nc
e
s
,
vol
.
181,
no.
14,
pp.
2923
–
2932,
J
ul
.
2011,
doi
:
10.1016/j
.i
ns
.2011.02.022.
[
18]
C.
-
Y
.
W
a
ng,
A
.
B
oc
hkov
s
ki
y,
a
nd
H
.
-
Y
.
M
.
L
ia
o,
“
S
c
a
le
d
-
Y
O
L
O
v4:
s
c
a
li
ng
c
r
os
s
s
t
a
ge
pa
r
ti
a
l
n
e
twor
k,”
in
2021
I
E
E
E
/
C
V
F
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
(
C
V
P
R
)
,
J
un.
2021,
pp.
13024
–
13033,
doi
:
10.1109/C
V
P
R
46437.2021.01283.
[
19]
K.
-
C
.
L
iu
,
Y
.
-
T
.
S
he
n,
a
nd
L
.
-
G
.
C
he
n,
“
S
im
pl
e
onl
in
e
a
nd
r
e
a
lt
im
e
tr
a
c
ki
ng
w
it
h
s
phe
r
ic
a
l
p
a
nor
a
mi
c
c
a
me
r
a
,”
in
2018
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on C
ons
um
e
r
E
le
c
tr
oni
c
s
(
I
C
C
E
)
, J
a
n
. 2018, pp. 1
–
6, doi:
10.1109/I
C
C
E
.2018.8326132.
[
20]
S
.
M
e
ns
hov,
Y
.
W
a
ng,
A
.
Z
hda
nov,
E
.
V
a
r
la
mov,
a
nd
D
.
Z
hd
a
nov,
“
S
im
pl
e
onl
in
e
a
nd
r
e
a
lt
im
e
tr
a
c
ki
ng
pe
opl
e
w
it
h
ne
w
‘
s
of
t
-
io
u’
me
tr
ic
,”
i
n
A
O
P
C
2019:
A
I
i
n O
pt
ic
s
and P
hot
oni
c
s
, D
e
c
. 2019, doi:
10.1117/12.254792
2.
[
21]
O
.
U
r
ba
nn,
O
.
B
r
e
dt
ma
nn,
M
.
O
tt
e
n,
J
.
P
.
R
ic
ht
e
r
,
T
.
B
a
ue
r
,
a
nd
D
.
Z
ib
r
ic
z
ky,
“
O
nl
in
e
a
nd
r
e
a
l
-
ti
me
tr
a
c
ki
ng
in
a
s
ur
ve
il
la
nc
e
s
c
e
na
r
io
,”
ar
X
iv
p
r
e
pr
in
t
ar
X
iv
:
2106.01153
, 2021.
[
22]
X
.
H
ou,
Y
.
W
a
ng,
a
nd
L
.
-
P
.
C
ha
u,
“
V
e
hi
c
le
tr
a
c
ki
ng
us
in
g
de
e
p
S
O
R
T
w
it
h
lo
w
c
onf
id
e
nc
e
tr
a
c
k
f
il
te
r
in
g,”
in
2019
16t
h
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
A
dv
anc
e
d
V
id
e
o
and
Si
gnal
B
as
e
d
Sur
v
e
il
la
nc
e
(
A
V
SS)
,
S
e
p.
2019,
pp.
1
–
6,
doi
:
10.1109/AVSS
.2019.8909903.
[
23]
S
.
I
of
f
e
a
nd
C
.
S
z
e
g
e
dy,
“
B
a
tc
h
nor
ma
li
z
a
ti
on:
a
c
c
e
le
r
a
ti
n
g
de
e
p
ne
twor
k
tr
a
in
in
g
by
r
e
duc
in
g
in
te
r
na
l
c
ova
r
ia
te
s
hi
f
t,
”
I
nt
e
r
nat
io
nal
c
onf
e
r
e
nc
e
on
m
ac
hi
ne
l
e
a
r
ni
ng
, pp. 448
–
456, 2015.
[
24]
J
. R
e
dmon a
nd A
. F
a
r
ha
di
, “
Y
ol
ov3:
a
n i
n
c
r
e
me
nt
a
l
im
pr
ove
me
nt
,”
ar
X
iv
pr
e
p
r
in
t
ar
X
iv
:
1804.02767
, 2018.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5681
-
5690
5690
[
25]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
S
pa
ti
a
l
pyr
a
mi
d
pool
i
ng
in
de
e
p
c
onvolut
io
na
l
ne
twor
ks
f
or
vi
s
ua
l
r
e
c
ogni
ti
on,”
I
E
E
E
T
r
ans
ac
ti
ons
on
P
at
te
r
n
A
nal
y
s
i
s
and
M
ac
hi
ne
I
nt
e
ll
ig
e
nc
e
,
vol
.
37,
no.
9,
pp.
1904
–
1916,
S
e
p.
2015,
doi
:
10.1109/T
P
A
M
I
.2015.2389824.
[
26]
H
.
Y
u,
X
.
L
i,
Y
.
F
e
ng,
a
nd
S
.
H
a
n,
“
M
ul
ti
pl
e
a
tt
e
nt
io
na
l
pa
th
a
ggr
e
ga
ti
on
ne
twor
k
f
or
ma
r
in
e
ob
je
c
t
de
te
c
ti
on,”
A
ppl
ie
d
I
nt
e
ll
ig
e
nc
e
, vol
. 53, no. 2, pp. 2434
–
2451, J
a
n. 2023, doi:
10.1
007/
s
10489
-
022
-
03622
-
0.
[
27]
P
.
R
.
G
unj
a
l,
B
.
R
.
G
unj
a
l,
H
.
A
.
S
hi
nde
,
S
.
M
.
V
a
na
m,
a
nd
S
.
S
.
A
he
r
,
“
M
ovi
ng
obj
e
c
t
tr
a
c
ki
ng
us
in
g
K
a
lm
a
n
f
il
te
r
,”
in
2018
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
O
n
A
dv
anc
e
s
in
C
om
m
uni
c
at
io
n
and
C
om
put
in
g
T
e
c
hnol
ogy
(
I
C
A
C
C
T
)
,
F
e
b.
2018,
pp.
544
–
547,
doi
:
10.1109/I
C
A
C
C
T
.2018.8529402.
[
28]
F
.
Y
u,
W
.
L
i,
Q
.
L
i,
Y
.
L
iu
,
X
.
S
hi
,
a
nd
J
.
Y
a
n,
“
P
O
I
:
mu
lt
ip
l
e
obj
e
c
t
tr
a
c
ki
ng
w
it
h
hi
gh
pe
r
f
or
ma
nc
e
de
te
c
ti
on
a
nd
a
ppe
a
r
a
nc
e
f
e
a
tu
r
e
,”
I
n:
H
ua
,
G
.,
J
é
gou,
H
.
(
e
ds
)
,
C
o
m
put
e
r
V
is
io
n
–
E
C
C
V
2016
W
or
k
s
hops
.
E
C
C
V
2016.
L
e
c
tu
r
e
N
ot
e
s
in
C
o
m
p
ut
e
r
Sc
ie
nc
e
, vol
.
9914. S
pr
in
ge
r
, C
ha
m.
2016, pp. 36
–
42
, doi
:
10.1
007/
978
-
3
-
319
-
48881
-
3_3
[
29]
A
.
B
e
w
le
y,
Z
.
G
e
,
L
.
O
tt
,
F
.
R
a
mos
,
a
nd
B
.
U
pc
r
of
t,
“
S
im
pl
e
onl
in
e
a
nd
r
e
a
lt
im
e
tr
a
c
ki
ng,”
in
2016
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on I
m
age
P
r
oc
e
s
s
in
g (
I
C
I
P
)
, S
e
p. 2016, pp. 3464
–
3468, doi:
10.1109/I
C
I
P
.2016.7533003.
[
30]
S
.
T
a
ng,
M
.
A
ndr
il
uka
,
B
.
A
ndr
e
s
,
a
nd
B
.
S
c
hi
e
le
,
“
M
ul
ti
pl
e
pe
opl
e
tr
a
c
ki
ng
by
li
f
te
d
mul
ti
c
ut
a
nd
pe
r
s
on
r
e
-
id
e
nt
if
ic
a
ti
on,”
in
2017
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
(
C
V
P
R
)
,
J
ul
.
2017,
pp.
3701
–
3710,
doi
:
10.1109/C
V
P
R
.2017.394.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
D
i
v
y
a
pr
a
bha
i
s
as
s
o
c
i
at
e
p
ro
fe
s
s
o
r
at
Sri
Si
d
d
h
art
h
a
In
s
t
i
t
u
t
e
o
f
T
ec
h
n
o
l
o
g
y
,
T
u
m
k
u
r,
In
d
i
a.
Sh
e
rece
i
v
e
d
h
er
b
ac
h
el
o
r’s
d
e
g
ree
fro
m
Ban
g
a
l
o
re
U
n
i
v
e
rs
i
t
y
an
d
mas
t
er’
s
d
eg
ree
fro
m
BI
T
S,
Pi
l
a
n
i
,
Ra
j
as
t
h
a
n
.
Sh
e
i
s
a
memb
er
o
f
IST
E
.
H
er
f
i
el
d
o
f
i
n
t
eres
t
s
are
i
mag
e
p
ro
c
es
s
i
n
g
,
mach
i
n
e
l
earn
i
n
g
.
Sh
e
h
as
p
u
b
l
i
s
h
e
d
p
ap
ers
i
n
co
n
fere
n
ces
an
d
j
o
u
rn
a
l
s
.
Sh
e
can
b
e
co
n
t
ac
t
ed
a
t
emai
l
:
d
i
v
y
as
y
7
4
@
g
mai
l
.
co
m.
Gurupra
s
a
d
Seeba
i
a
h
i
s
as
s
o
c
i
at
e
p
ro
fes
s
o
r
at
Sri
Si
d
d
h
ar
t
h
a
In
s
t
i
t
u
t
e
o
f
T
ech
n
o
l
o
g
y
,
T
u
m
k
u
r,
In
d
i
a.
H
e
recei
v
ed
h
i
s
b
ach
e
l
o
r’
s
,
mas
t
er’s
an
d
d
o
ct
o
ral
d
e
g
ree
fro
m
V
i
s
v
e
s
v
ara
y
a
T
ech
n
o
l
o
g
i
ca
l
U
n
i
v
ers
i
t
y
(V
T
U
),
Bel
g
au
m,
In
d
i
a.
H
e
i
s
a
memb
er
o
f
IST
E
.
H
i
s
fi
el
d
o
f
i
n
t
eres
t
s
are
i
mag
e
p
r
o
ce
s
s
i
n
g
an
d
b
i
o
me
d
i
c
al
i
n
s
t
r
u
men
t
at
i
o
n
.
H
e
h
as
p
u
b
l
i
s
h
e
d
res
earch
p
a
p
ers
i
n
rep
u
t
ed
co
n
feren
ce
s
an
d
i
n
t
er
n
at
i
o
n
al
j
o
u
r
n
al
s
.
H
e
can
b
e
co
n
t
ac
t
ed
a
t
emai
l
:
g
u
r
u
p
ra
s
ad
s
@
s
s
i
t
.
e
d
u
.
i
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.