I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
2025
,
pp.
3343
~
3353
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
33
43
-
3353
3343
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
T
r
af
f
ic
f
lo
w
p
r
e
d
i
c
t
io
n
u
si
n
g
l
on
g sh
or
t
-
t
e
r
m
m
e
m
or
y
-
Kom
od
o
M
li
p
ir
al
gor
it
h
m
:
m
e
t
ah
e
u
r
is
t
ic
op
t
i
m
iz
at
io
n
t
o m
u
lti
-
t
a
r
g
e
t
ve
h
ic
l
e
d
e
t
e
c
t
io
n
I
m
am
Ahm
ad
As
h
ar
i
1
,4
,
Wah
yu
l
Am
ien
S
yaf
e
i
2
,
Adi
Wib
owo
3
1
D
oc
to
r
a
l
P
r
ogr
a
m of
I
nf
or
ma
ti
on
S
ys
te
ms
, U
ni
ve
r
s
it
a
s
D
ip
one
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
2
F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g, U
ni
ve
r
s
it
a
s
D
ip
one
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
3
F
a
c
ul
ty
of
S
c
ie
nc
e
a
nd
M
a
th
e
ma
ti
c
s
,
U
ni
ve
r
s
it
a
s
D
ip
on
e
gor
o, S
e
ma
r
a
ng, I
ndone
s
ia
4
F
a
c
ul
ty
of
S
c
ie
nc
e
a
nd
T
e
c
hnol
ogy,
U
ni
ve
r
s
it
a
s
H
a
r
a
pa
n
B
a
ng
s
a
, P
ur
w
oke
r
to
, I
ndone
s
ia
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
De
c
12,
2024
R
e
vis
e
d
J
un
24,
2025
Ac
c
e
pted
J
ul
13,
2025
Mu
l
t
i
-
t
arg
e
t
v
eh
i
cl
e
d
e
t
ect
i
o
n
i
n
u
r
b
an
t
raff
i
c
faces
ch
a
l
l
e
n
g
e
s
s
u
c
h
as
p
o
o
r
l
i
g
h
t
i
n
g
,
s
ma
l
l
o
b
j
ec
t
s
i
zes
,
a
n
d
d
i
v
ers
e
v
eh
i
cl
e
t
y
p
e
s
,
i
mp
act
i
n
g
t
raff
i
c
fl
o
w
p
red
i
ct
i
o
n
accu
racy
.
T
h
i
s
s
t
u
d
y
i
n
t
ro
d
u
ce
s
an
o
p
t
i
mi
zed
l
o
n
g
s
h
o
rt
-
t
erm
memo
ry
(L
ST
M)
mo
d
el
u
s
i
n
g
t
h
e
K
o
mo
d
o
M
l
i
p
i
r
a
l
g
o
ri
t
h
m
(K
M
A
)
t
o
en
h
a
n
ce
p
re
d
i
c
t
i
o
n
a
ccu
rac
y
.
T
raff
i
c
v
i
d
e
o
d
a
t
a
are
p
r
o
ces
s
ed
w
i
t
h
Y
O
L
O
fo
r
v
eh
i
cl
e
c
l
as
s
i
f
i
cat
i
o
n
an
d
o
b
j
ect
co
u
n
t
i
n
g
.
T
h
e
L
ST
M
mo
d
el
,
t
ra
i
n
e
d
t
o
cap
t
u
re
t
raffi
c
p
a
t
t
er
n
s
,
emp
l
o
y
s
p
aramet
er
s
o
p
t
i
m
i
zed
b
y
K
M
A
,
i
n
c
l
u
d
i
n
g
l
earn
i
n
g
rat
e,
n
eu
ro
n
co
u
n
t
,
an
d
ep
o
ch
s
.
K
MA
i
n
t
e
g
rat
e
s
mu
t
at
i
o
n
an
d
cro
s
s
o
v
er
s
t
rat
e
g
i
e
s
t
o
en
a
b
l
e
ad
a
p
t
i
v
e
s
el
ec
t
i
o
n
i
n
g
l
o
b
al
an
d
l
o
cal
s
earch
es
.
T
h
e
mo
d
e
l
's
p
erf
o
rman
ce
w
as
ev
a
l
u
a
t
ed
o
n
an
u
rb
an
t
raffi
c
d
at
as
e
t
w
i
t
h
u
n
i
fo
rm
co
n
fi
g
u
ra
t
i
o
n
s
fo
r
p
o
p
u
l
at
i
o
n
s
i
z
e
an
d
k
e
y
L
ST
M
p
aramet
ers
,
en
s
u
ri
n
g
co
n
s
i
s
t
e
n
t
ev
a
l
u
a
t
i
o
n
.
Res
u
l
t
s
s
h
o
w
ed
L
ST
M
-
K
MA
ach
i
e
v
ed
a
ro
o
t
mean
s
q
u
are
erro
r
(RMSE
)
o
f
1
4
.
5
3
1
9
,
o
u
t
p
erfo
rm
i
n
g
L
ST
M
(1
6
.
6
8
2
7
),
L
ST
M
-
i
mp
r
o
v
e
d
d
u
n
g
b
eet
l
e
o
p
t
i
m
i
zat
i
o
n
(
ID
BO
)
(1
5
.
0
9
4
6
),
an
d
L
ST
M
-
p
art
i
c
l
e
s
w
arm
o
p
t
i
m
i
zat
i
o
n
(
PSO
)
(1
5
.
0
3
6
8
).
It
s
mean
ab
s
o
l
u
t
e
erro
r
(MA
E
),
at
8
.
7
0
4
1
,
al
s
o
s
u
r
p
a
s
s
e
d
L
ST
M
(9
.
9
9
0
3
),
L
ST
M
-
ID
B
O
(9
.
0
3
2
8
),
an
d
L
ST
M
-
PS
O
(9
.
0
0
1
5
).
L
ST
M
-
K
MA
effect
i
v
e
l
y
t
ack
l
es
mu
l
t
i
-
t
ar
g
et
d
et
ec
t
i
o
n
c
h
al
l
en
g
es
,
i
mp
r
o
v
i
n
g
p
re
d
i
ct
i
o
n
accu
rac
y
an
d
t
ran
s
p
o
rt
a
t
i
o
n
s
y
s
t
em
effi
c
i
en
c
y
.
T
h
i
s
re
l
i
a
b
l
e
s
o
l
u
t
i
o
n
s
u
p
p
o
rt
s
rea
l
-
t
i
me
u
rb
a
n
t
raff
i
c
man
ag
eme
n
t
,
ad
d
res
s
i
n
g
t
h
e
d
ema
n
d
s
o
f
d
y
n
am
i
c
u
r
b
an
en
v
i
ro
n
men
t
s
.
K
e
y
w
o
r
d
s
:
Komodo
M
li
pir
a
lgor
it
hm
L
ong
s
hor
t
-
ter
m
memor
y
M
e
tahe
ur
is
ti
c
opti
mi
z
a
ti
on
M
ult
i
-
tar
ge
t
ve
hicl
e
de
tec
ti
on
T
r
a
f
f
ic
f
low
pr
e
diction
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
:
I
mam
Ahma
d
As
ha
r
i
Doc
tor
a
l
P
r
ogr
a
m
of
I
nf
o
r
mation
S
ys
tems
,
Unive
r
s
it
a
s
Dipone
gor
o
S
e
mar
a
ng,
I
ndone
s
ia
E
mail:
im
a
mahma
d@s
tudents
.
undip.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
W
it
h
a
dva
nc
e
ments
in
c
omm
unica
ti
on
tec
hnolog
y
a
nd
c
omput
e
r
s
c
ienc
e
,
int
e
ll
igent
tr
a
ns
por
tation
s
ys
tems
(
I
T
S
)
ha
ve
a
s
s
umed
a
n
incr
e
a
s
ingl
y
s
igni
f
ica
nt
r
ole
in
da
il
y
li
f
e
[
1
]
.
S
mar
t
tr
a
ns
por
tation
ha
s
be
c
ome
a
c
or
ne
r
s
tone
in
the
de
ve
lopm
e
nt
of
tec
hnology
-
ba
s
e
d
I
T
S
to
mee
t
the
e
volvi
ng
ne
e
ds
of
ur
ba
n
s
oc
ieties
[
2]
.
I
t
r
e
f
e
r
s
to
a
n
a
pp
r
oa
c
h
that
int
e
gr
a
tes
moder
n
tec
hnology
int
o
tr
a
ns
por
tation
s
ys
tems
to
e
nha
nc
e
ur
ba
n
mobi
li
ty
e
f
f
icie
nc
y
[
3
]
.
I
n
the
c
ontext
of
s
mar
t
c
it
i
e
s
,
c
utt
ing
-
e
dge
t
e
c
hnologi
e
s
s
uc
h
a
s
the
int
e
r
ne
t
of
thi
ngs
(
I
oT
)
,
da
ta
a
na
lyt
ics
,
a
nd
a
r
ti
f
icia
l
int
e
ll
igenc
e
(
A
I
)
s
e
r
ve
a
s
f
ounda
ti
ona
l
pi
ll
a
r
s
f
or
c
r
e
a
ti
ng
int
e
ll
i
ge
nt
a
nd
int
e
r
c
onne
c
ted
tr
a
ns
por
tation
e
c
os
ys
tems
[
4]
.
S
m
a
r
t
mobi
li
ty
ha
s
be
c
ome
a
n
int
e
gr
a
l
pa
r
t
of
da
il
y
li
f
e
,
with
40%
of
the
global
population
tr
a
ve
li
ng
f
o
r
a
t
lea
s
t
one
hour
e
a
c
h
da
y
[
5
]
.
B
y
int
e
gr
a
ti
ng
tec
hnologi
e
s
s
uc
h
a
s
c
omput
e
r
vis
ion,
AI
,
a
nd
I
T
S
,
c
it
ies
c
a
n
mo
r
e
a
c
c
ur
a
tely
de
tec
t
tr
a
f
f
ic
c
ondit
ions
,
identi
f
y
ve
hicle
t
ype
s
,
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
334
3
-
3353
3344
pr
e
dict
c
onge
s
ti
on.
T
his
int
e
gr
a
ti
on
he
lps
a
ddr
e
s
s
ur
ba
niza
ti
on
c
ha
ll
e
nge
s
,
s
uc
h
a
s
poll
uti
on
,
tr
a
f
f
ic
a
c
c
idents
,
a
nd
e
xc
e
s
s
ive
r
e
s
our
c
e
c
ons
umpt
ion
[
6]
.
T
r
a
f
f
ic
f
low
pr
e
diction
is
a
c
r
it
ica
l
e
leme
nt
in
I
T
S
a
s
it
pr
ovides
va
luable
ins
ight
s
f
o
r
tr
a
f
f
ic
c
ont
r
ol,
r
oute
planning,
a
nd
ope
r
a
ti
ona
l
mana
ge
ment
[
7
]
.
T
r
a
dit
ional
tr
a
f
f
ic
f
low
pr
e
diction
models
of
te
n
f
a
il
to
a
de
qua
tely
a
c
c
ount
f
or
the
c
ompl
e
x
a
nd
dyna
mi
c
c
ha
r
a
c
ter
is
ti
c
s
of
ur
ba
n
tr
a
f
f
ic
ne
two
r
ks
[
8
]
.
W
it
h
the
a
c
c
e
ler
a
ti
on
of
ur
ba
niza
ti
on
a
nd
a
dva
nc
e
ments
in
I
T
S
,
s
hor
t
-
ter
m
tr
a
f
f
ic
f
low
p
r
e
diction
ha
s
e
mer
g
e
d
a
s
a
n
incr
e
a
s
ingl
y
s
igni
f
ica
nt
a
r
e
a
of
r
e
s
e
a
r
c
h
[
9]
.
A
c
c
ur
a
te
pr
e
dictions
of
f
e
r
s
ubs
tantial
be
ne
f
it
s
,
i
nc
ludi
ng
opti
mi
z
e
d
tr
a
f
f
ic
planning,
im
p
r
ove
d
r
oa
d
uti
li
z
a
ti
on,
r
e
duc
e
d
c
onge
s
ti
on,
f
e
we
r
tr
a
f
f
ic
a
c
c
idents
,
a
nd
de
c
r
e
a
s
e
d
e
nvir
onmenta
l
poll
uti
on
[
10
]
.
Ac
c
ur
a
te
tr
a
f
f
ic
f
low
p
r
e
diction
r
e
quir
e
s
the
e
f
f
i
c
ient
e
xtr
a
c
ti
on
a
nd
a
na
lys
is
of
lar
ge
-
s
c
a
le
ur
ba
n
tr
a
f
f
ic
da
ta,
including
the
a
ppr
opr
iate
s
e
lec
ti
on
of
da
ta
s
a
mpl
e
s
ize
s
.
T
e
c
hnologi
c
a
l
a
dva
nc
e
ments
,
s
uc
h
a
s
r
oa
ds
ide
c
los
e
d
-
c
ir
c
uit
tele
vis
ion
(
C
C
T
V
)
c
a
mer
a
s
a
nd
unmanne
d
a
e
r
ial
ve
hicle
s
(
UA
Vs
)
,
pr
ovide
ne
w
video
da
ta
that
e
na
ble
mor
e
c
omp
r
e
he
ns
ive
tr
a
f
f
ic
in
f
or
mation
c
oll
e
c
ti
on
thr
ough
c
omput
e
r
vis
ion
tec
hniq
ue
s
[
11]
.
T
he
s
e
a
dva
nc
e
ments
s
uppor
t
a
c
c
ident
-
ba
s
e
d
s
a
f
e
ty
a
na
lys
is
a
nd
f
a
c
il
it
a
te
r
e
a
l
-
ti
me
tr
a
f
f
ic
c
ontr
ol,
r
oute
guidanc
e
,
poli
c
y
f
or
mul
a
t
ion,
a
nd
mo
r
e
e
f
f
e
c
ti
ve
tr
a
f
f
ic
a
ll
oc
a
ti
on
.
T
oge
the
r
,
thes
e
e
f
f
o
r
ts
e
nha
nc
e
tr
a
f
f
ic
e
f
f
icie
nc
y
a
nd
im
p
r
ove
the
qua
li
ty
o
f
u
r
ba
n
li
f
e
[
1
2]
.
I
n
p
r
a
c
ti
c
e
,
c
omput
e
r
vis
ion
models
s
uc
h
a
s
Y
OL
O
a
nd
it
s
a
dva
nc
e
ments
a
r
e
wide
ly
a
ppli
e
d
to
de
tec
t
a
nd
a
na
lyze
ur
ba
n
tr
a
f
f
ic
c
ondit
ions
[
13
]
–
[
18]
.
I
n
I
T
S
,
tr
a
dit
ional
ob
jec
t
de
tec
ti
on
a
lgor
it
hms
f
a
c
e
va
r
ious
c
ha
ll
e
nge
s
,
pa
r
ti
c
u
lar
ly
in
de
a
li
ng
with
c
ompl
e
x
e
nvir
onments
a
nd
va
r
ying
li
gh
ti
ng
c
ondit
ions
.
T
he
s
e
c
ha
ll
e
nge
s
be
c
ome
mor
e
s
igni
f
ica
nt
whe
n
de
tec
ti
ng
s
mall
objec
ts
or
a
na
lyzing
mul
ti
modal
da
ta
[
16]
.
T
o
a
ddr
e
s
s
thes
e
li
mi
tations
,
e
nha
nc
ing
da
ta
qua
li
ty
a
nd
diver
s
it
y
thr
ough
a
ugmenta
ti
on
tec
hniques
is
a
c
omm
on
a
pp
r
oa
c
h
[
19
]
.
P
r
e
vious
r
e
s
e
a
r
c
h
ha
s
de
mons
tr
a
ted
that
c
ombi
ning
objec
t
de
tec
ti
on
with
long
s
hor
t
-
ter
m
memor
y
(
L
S
T
M
)
a
lgo
r
it
hms
c
a
n
e
f
f
e
c
ti
ve
ly
pr
e
dict
tr
a
f
f
ic
volum
e
[
20
]
.
Addit
ionally,
s
tudi
e
s
ha
ve
pr
opos
e
d
the
de
ve
lopm
e
nt
of
ne
w
models
leve
r
a
ging
a
nd
opti
mi
z
ing
L
S
T
M
,
whic
h
ha
s
pr
ov
e
n
e
f
f
e
c
ti
ve
in
ha
ndli
ng
ti
me
-
s
e
r
ies
da
ta
a
nd
im
pr
oving
the
a
c
c
ur
a
c
y
of
ur
ba
n
tr
a
f
f
ic
de
ns
it
y
pr
e
dictions
[
20]
–
[
25]
.
R
e
c
e
nt
tr
e
nds
s
ugge
s
t
a
n
incr
e
a
s
ing
f
oc
us
on
opti
mi
z
ing
L
S
T
M
pa
r
a
mete
r
s
th
r
ough
meta
he
ur
is
ti
c
a
ppr
oa
c
he
s
to
im
pr
ove
tr
a
f
f
ic
pr
e
diction
pe
r
f
or
manc
e
[
22]
,
[
26]
–
[
28]
.
S
uc
h
a
n
a
ppr
oa
c
h
is
a
nti
c
ipate
d
to
tac
kle
the
c
ha
ll
e
nge
s
of
c
r
e
a
ti
ng
mor
e
r
e
li
a
ble
a
nd
e
f
f
icie
nt
pr
e
dictive
models
f
or
va
r
ious
t
r
a
f
f
ic
c
ond
it
ions
.
T
he
Kom
odo
m
li
p
ir
op
ti
mi
z
a
t
ion
a
l
gor
it
h
m
(
KM
A)
dr
a
ws
i
ns
pi
r
a
ti
on
f
r
om
two
un
ique
phe
nome
na
:
the
be
ha
vio
r
o
f
K
omod
o
d
r
a
go
ns
na
t
ive
to
E
a
s
t
N
us
a
T
e
ngga
r
a
,
I
ndon
e
s
ia,
a
nd
the
tr
a
dit
iona
l
J
a
va
ne
s
e
wa
lki
n
g
s
tyl
e
k
nown
a
s
m
l
ipi
r
[
29
]
.
I
n
the
c
ont
e
xt
of
the
t
r
a
ve
l
ing
s
a
les
ma
n
pr
oble
m
(
T
S
P
)
,
K
M
A
ha
s
e
xhib
it
e
d
s
upe
r
i
or
pe
r
f
or
manc
e
c
o
mpar
e
d
to
a
l
g
or
i
thm
s
li
ke
the
d
r
a
go
nf
ly
a
lg
or
it
h
m
(
DK
A
)
,
a
n
t
c
o
lon
y
opti
mi
z
a
t
ion
(
AC
O
)
,
pa
r
ti
c
le
s
wa
r
m
opt
im
iza
ti
on
(
P
S
O)
,
ge
ne
ti
c
a
lgo
r
i
thm
(
GA
)
,
blac
k
hole
(
B
H
)
,
dyna
m
ic
tabu
s
e
a
r
c
h
a
lgo
r
it
hm
(
D
T
S
A)
,
a
nd
d
is
c
r
e
te
jaya
a
l
gor
i
thm
(
DJ
AY
A
)
[
3
0]
.
I
n
o
ur
p
r
op
os
e
d
r
e
s
e
a
r
c
h
,
L
S
T
M
is
c
ombi
ne
d
w
it
h
KM
A
f
or
tr
a
f
f
ic
volu
me
p
r
e
d
icti
on.
T
h
e
L
S
T
M
-
KM
A
mode
l
is
then
c
ompa
r
e
d
with
the
s
tanda
r
d
L
S
T
M
a
nd
o
the
r
s
tate
-
of
-
the
-
a
r
t
c
omb
in
a
ti
ons
,
na
me
ly
L
S
T
M
-
im
p
r
ove
d
d
ung
be
e
tl
e
op
ti
mi
z
a
t
ion
(
I
DB
O
)
a
nd
L
S
T
M
-
P
S
O
.
P
r
e
vious
s
tud
ies
ha
ve
s
hown
that
L
S
T
M
-
I
DB
O
out
pe
r
f
o
r
ms
met
hods
s
uc
h
a
s
g
r
a
y
wolf
op
ti
m
iza
t
ion
(
GW
O
)
,
s
pa
r
r
ow
o
pti
mi
z
a
ti
on
a
l
gor
i
t
hm
(
S
S
A
)
,
w
ha
le
opt
im
iz
a
ti
on
a
lgo
r
it
hm
(
W
OA
)
,
a
n
d
night
ha
wk
op
ti
m
iza
t
ion
(
NG
O
)
[
26
]
.
S
i
mi
la
r
l
y,
L
S
T
M
-
P
S
O
ha
s
p
r
ove
n
s
u
pe
r
ior
t
o
met
hods
li
ke
s
tand
a
r
d
LS
T
M
,
r
a
nd
om
f
or
e
s
t
r
e
g
r
e
s
s
ion
(
R
F
R
)
,
k
-
ne
a
r
e
s
t
r
e
gr
e
s
s
ion
(
KN
R
)
,
a
n
d
de
c
is
io
n
tr
e
e
r
e
g
r
e
s
s
ion
(
D
T
R
)
[
28
]
.
T
he
main
pr
oblem
a
dd
r
e
s
s
e
d
in
thi
s
s
tudy
is
the
low
a
c
c
ur
a
c
y
in
pr
e
dicting
c
ompl
e
x
a
nd
dyna
mi
c
tr
a
f
f
ic
vo
lum
e
s
,
pa
r
ti
c
ula
r
ly
unde
r
r
e
a
l
-
wor
ld
c
on
dit
ions
that
of
ten
invol
ve
c
ha
ll
e
nge
s
s
uc
h
a
s
poor
li
ghti
ng,
oc
c
lus
ions
,
a
nd
diver
s
e
ve
hicle
types
.
T
o
a
ddr
e
s
s
thi
s
is
s
ue
,
the
s
tudy
a
im
s
to
de
ve
lop
a
tr
a
f
f
ic
pr
e
diction
model
that
in
tegr
a
tes
the
L
S
T
M
a
lgo
r
it
hm
with
th
e
KM
A
a
s
a
n
opti
mi
z
a
ti
on
method,
s
uppor
ted
by
r
e
a
l
-
ti
me
ve
hicle
de
tec
ti
on
da
ta
us
ing
YO
L
O.
T
his
r
e
s
e
a
r
c
h
s
pe
c
if
ica
ll
y
f
oc
us
e
s
on
how
the
int
e
gr
a
ti
on
of
KM
A
c
a
n
im
pr
ove
the
pr
e
dictive
a
c
c
ur
a
c
y
of
L
S
T
M
in
mod
e
li
ng
dyna
mi
c
tr
a
f
f
ic
volum
e
s
,
a
nd
e
va
luate
s
the
potential
im
pleme
ntation
of
the
YO
L
O
-
L
S
T
M
-
K
M
A
s
ys
te
m
unde
r
r
e
a
l
t
r
a
f
f
ic
c
ondit
ions
.
T
he
main
c
ontr
ib
uti
on
of
thi
s
s
tudy
is
the
de
ve
lopm
e
nt
of
a
n
int
e
ll
igent
p
r
e
dictive
model
c
a
pa
ble
of
im
pr
oving
t
r
a
f
f
ic
f
low
p
r
e
diction
a
c
c
ur
a
c
y,
of
f
e
r
ing
bo
th
theo
r
e
ti
c
a
l
c
ontr
ibu
ti
ons
i
n
the
f
ield
of
opti
mi
z
a
ti
on
a
nd
ti
me
-
s
e
r
ies
f
or
e
c
a
s
ti
ng,
a
nd
pr
a
c
ti
c
a
l
c
ontr
ibut
ions
i
n
s
uppor
ti
ng
da
ta
-
dr
iven
d
e
c
is
ion
-
making
withi
n
I
T
S
.
2.
M
E
T
HO
D
2.
1.
Ve
h
icle
ob
j
e
c
t
d
e
t
e
c
t
ion
Da
ta
c
oll
e
c
ti
on
wa
s
c
ondu
c
ted
us
ing
YO
L
O
a
s
the
objec
t
de
tec
ti
on
model
f
or
identif
ying
ve
hicle
s
in
tr
a
f
f
ic.
S
pe
c
if
ica
ll
y
,
the
YO
L
Ov8n
model
wa
s
us
e
d
be
c
a
us
e
of
it
s
high
-
pe
r
f
or
manc
e
a
bil
it
y
to
de
tec
t
ve
hicle
s
in
c
ompl
e
x
tr
a
f
f
ic
c
ondit
ions
.
T
o
im
pr
ove
de
tec
t
i
on
a
c
c
ur
a
c
y,
mul
ti
-
a
ugmenta
ti
on
tec
hniques
we
r
e
a
ppli
e
d,
c
ombi
ning
s
c
a
li
ng,
z
oom
-
in,
br
ight
ne
s
s
a
djus
tm
e
nt,
c
olor
ji
tt
e
r
,
a
nd
nois
e
inj
e
c
ti
on
.
T
a
ble
1
p
r
e
s
e
nts
the
s
pe
c
if
ic
va
lues
f
or
e
a
c
h
a
ugmenta
ti
on
tec
hnique
us
e
d
in
the
s
tudy.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
r
aff
ic
fl
ow
pr
e
diction
us
ing
long
s
hor
t
-
ter
m
me
mor
y
-
K
omodo
M
li
pir
A
lgor
it
hm:
…
(
I
mam
A
hmad
A
s
har
i)
3345
T
a
ble
1
.
Augme
ntation
v
a
lues
No
A
ug
V
a
lu
e
A
ugme
nt
a
ti
on
f
a
c
to
r
(
im
a
ge
)
R
e
f
e
r
e
nc
e
s
1
2
3
1
B
r
ig
ht
ne
s
s
a
dj
us
tm
e
nt
B
r
ig
ht
ne
s
s
f
a
c
to
r
-
0.8
1.2
[
31]
2
C
ol
or
j
it
te
r
(
B
r
ig
ht
ne
s
s
, c
ont
r
a
s
t,
s
a
tu
r
a
ti
on)
a
nd hue
-
R
a
nd (
0.6,1.4)
a
nd
R
a
nd
(
-
0.1,0.1)
R
a
nd (
0.6,1.4)
a
nd
R
a
nd (
-
0.1,0.1)
[
32]
3
N
oi
s
e
i
nj
e
c
ti
on
G
a
us
s
ia
n nois
e
-
R
a
nd (
0, 0.1)
R
a
nd (
0, 0.1)
[
33]
4
S
c
a
li
ng
S
c
a
le
i
ma
ge
-
R
a
nd (
0.8, 1.2)
R
a
nd (
0.8, 1.2)
[
34]
5
Z
oom i
n
Z
oom i
n
-
1.2
1.5
[
35]
I
n
the
im
a
ge
a
ugmenta
ti
on
pr
oc
e
s
s
s
umm
a
r
ize
d
in
T
a
ble
1,
br
ight
ne
s
s
a
djus
tm
e
nt
wa
s
pe
r
f
or
med
with
a
br
ight
ne
s
s
f
a
c
tor
o
f
0
.
8
f
or
im
a
ge
2
a
nd
1
.
2
f
or
im
a
ge
3
.
F
or
the
c
olor
ji
tt
e
r
tec
hnique,
the
b
r
i
ghtnes
s
,
c
ontr
a
s
t,
a
nd
s
a
tur
a
ti
on
f
a
c
tor
s
we
r
e
r
a
ndomi
z
e
d
w
it
hin
the
r
a
nge
o
f
0
.
6
to
1.
4
,
while
the
hue
f
a
c
tor
wa
s
r
a
ndomi
z
e
d
be
twe
e
n
-
0.
1
a
nd
0.
1.
Nois
e
inj
e
c
ti
on
uti
li
z
e
d
Ga
us
s
ian
nois
e
with
va
lues
r
a
ndomi
z
e
d
be
twe
e
n
0
a
nd
0
.
1
f
or
both
im
a
ge
s
.
T
he
s
c
a
li
ng
tec
hniqu
e
wa
s
a
ppli
e
d
with
a
f
a
c
tor
r
a
nge
of
0
.
8
to
1
.
2
f
or
bo
th
im
a
ge
2
a
nd
im
a
ge
3,
while
the
z
oom
-
in
tec
hnique
uti
li
z
e
d
a
f
a
c
tor
of
1.
2
f
or
i
mage
2
a
nd
1.
5
f
o
r
i
mage
3.
T
his
c
ombi
na
ti
on
of
va
lues
wa
s
de
s
igned
to
c
r
e
a
te
s
igni
f
ica
nt
i
mage
va
r
iations
,
the
r
e
by
i
mpr
o
ving
the
model's
pe
r
f
or
manc
e
unde
r
diver
s
e
c
on
dit
ions
.
B
a
s
e
d
on
t
h
e
c
o
n
du
c
te
d
e
xp
e
r
i
me
n
ts
,
Y
OL
O
v8
n
o
u
t
p
e
r
f
o
r
me
d
Y
OL
O
v
9
t
,
a
c
h
ie
v
i
n
g
th
e
h
i
gh
e
s
t
m
A
P
5
0
-
9
5
v
a
l
u
e
o
f
0
.
5
36
.
A
de
t
a
i
l
e
d
p
e
r
f
o
r
ma
n
c
e
a
n
a
l
ys
is
is
p
r
e
s
e
n
te
d
i
n
a
ma
n
us
c
r
i
p
t
t
it
l
e
d
"
b
o
os
t
i
ng
r
e
a
l
-
t
i
me
v
e
h
ic
l
e
de
t
e
c
t
i
on
i
n
u
r
ba
n
t
r
a
f
f
i
c
us
i
ng
a
no
v
e
l
m
u
l
t
i
-
a
u
g
m
e
n
t
a
ti
o
n
"
.
T
h
e
e
x
p
e
r
im
e
n
ta
l
r
e
s
u
lt
s
id
e
nt
i
f
i
e
d
t
h
e
be
s
t
-
p
e
r
f
o
r
m
i
n
g
m
o
de
l
,
na
m
e
d
b
e
s
t
.
p
t
,
a
s
t
h
e
f
o
u
nd
a
t
ion
f
o
r
t
he
ve
h
i
c
le
de
t
e
c
t
i
on
p
r
oc
e
s
s
i
n
t
h
is
s
t
u
d
y
.
T
h
e
m
od
e
l
w
o
r
k
f
l
o
w
is
d
e
p
i
c
te
d
i
n
F
i
gu
r
e
1
,
d
e
ta
i
l
in
g
t
h
e
s
t
e
p
s
f
r
o
m
d
a
t
a
p
r
e
p
r
o
c
e
s
s
in
g
t
o
n
u
me
r
i
c
f
e
a
t
u
r
e
e
x
t
r
a
c
t
ion
.
F
igur
e
1.
Nume
r
ica
l
f
e
a
tur
e
e
xtr
a
c
ti
on
pr
oc
e
s
s
f
r
o
m
YO
L
O
model
T
he
model
wor
k
f
low,
a
s
il
lus
tr
a
ted
in
F
igu
r
e
1,
be
gins
with
the
c
oll
e
c
ti
on
of
video
da
ta
f
r
o
m
tr
a
f
f
ic
C
C
T
V
r
e
c
or
dings
.
T
his
video
da
ta
is
pr
oc
e
s
s
e
d
thr
ough
a
pr
e
pr
oc
e
s
s
ing
s
tage
whe
r
e
it
is
c
onve
r
ted
int
o
indi
vidual
f
r
a
mes
f
or
f
ur
ther
a
na
lys
is
.
E
a
c
h
f
r
a
me
is
manua
ll
y
a
nnotate
d
us
ing
the
R
obof
low
a
ppli
c
a
ti
on
to
labe
l
ve
hicle
objec
ts
,
whic
h
include
mot
or
c
yc
les
,
c
a
r
s
,
tr
uc
ks
,
a
nd
bus
e
s
.
T
he
a
nnotation
pr
oc
e
s
s
invol
ve
d
c
r
e
a
ti
ng
f
our
ve
hicle
c
las
s
e
s
a
nd
dr
a
wing
boundi
ng
boxe
s
a
r
ound
e
a
c
h
objec
t
in
e
ve
r
y
f
r
a
me.
I
n
t
otal,
the
da
tas
e
t
c
ontains
720
im
a
ge
s
with
45,
347
a
nno
tations
,
c
ons
is
ti
ng
o
f
31,
481
mot
o
r
c
yc
les
,
12
,
4
02
c
a
r
s
,
1,
184
tr
uc
ks
,
a
nd
280
bus
e
s
.
T
he
da
tas
e
t
is
divi
de
d
int
o
two
pa
r
ts
:
80%
f
or
t
r
a
ini
ng
a
nd
20%
f
o
r
va
li
da
ti
on
[
36]
,
[
37]
.
T
his
80:20
s
pli
t
is
c
omm
only
us
e
d
in
mac
hine
lea
r
ning
e
xpe
r
im
e
nts
to
e
ns
ur
e
that
the
model
ha
s
s
uf
f
icie
nt
da
ta
to
lea
r
n
pa
tt
e
r
ns
dur
ing
t
r
a
ini
ng
while
maintaining
a
n
a
de
qua
te
por
ti
on
of
uns
e
e
n
da
ta
f
or
unbias
e
d
va
li
da
ti
on,
a
ll
owing
f
or
a
c
c
ur
a
te
e
va
lu
a
ti
on
of
the
model’
s
ge
ne
r
a
li
z
a
ti
on
a
bil
it
y
.
T
he
tr
a
ini
ng
s
ubs
e
t
include
s
22,
136
mot
or
c
yc
les
,
8
,
804
c
a
r
s
,
839
t
r
uc
ks
,
a
nd
199
bus
e
s
,
while
the
va
li
da
ti
o
n
s
ubs
e
t
c
ontains
9,
345
mot
or
c
yc
les
,
3,
598
c
a
r
s
,
345
tr
uc
ks
,
a
nd
81
bus
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
334
3
-
3353
3346
T
o
incr
e
a
s
e
da
ta
diver
s
it
y
a
nd
im
p
r
ove
model
ge
n
e
r
a
li
z
a
ti
on,
a
ugmenta
ti
on
tec
hniques
we
r
e
a
ppli
e
d
to
the
tr
a
ini
ng
da
tas
e
t.
T
he
s
e
tec
hniques
include
s
c
a
li
ng,
z
oom
-
in,
br
ight
ne
s
s
a
djus
tm
e
nt,
c
olor
ji
tt
e
r
,
a
nd
nois
e
inj
e
c
ti
on.
E
a
c
h
tec
h
nique
wa
s
a
ppli
e
d
us
ing
two
pa
r
a
mete
r
va
lues
,
r
e
s
ult
ing
in
a
ten
f
old
incr
e
a
s
e
in
the
a
mount
of
tr
a
ini
ng
da
ta.
T
he
or
igi
na
l
tr
a
ini
ng
da
ta
s
e
t
c
ons
is
ts
of
576
im
a
ge
s
without
a
ugmenta
ti
on,
while
the
a
ugmente
d
da
tas
e
t
c
on
s
is
ts
of
5,
760
im
a
ge
s
,
a
s
s
h
own
in
T
a
ble
1.
T
he
YO
L
O
model
wa
s
tr
a
ined
us
ing
thi
s
e
nha
nc
e
d
da
tas
e
t,
a
nd
it
s
pe
r
f
or
manc
e
wa
s
e
va
lua
ted
pe
r
iodi
c
a
ll
y
us
ing
the
mAP
50
-
95
metr
ic.
I
f
th
e
model
did
not
mee
t
the
de
s
ir
e
d
a
c
c
ur
a
c
y
th
r
e
s
hold,
t
r
a
i
ning
wa
s
c
onti
nue
d.
Onc
e
the
be
s
t
-
pe
r
f
or
mi
ng
m
ode
l
wa
s
obtai
ne
d,
it
wa
s
us
e
d
to
de
tec
t
ve
hicle
s
in
e
a
c
h
f
r
a
me
a
nd
pr
e
dict
their
c
las
s
e
s
.
T
he
de
tec
ti
on
r
e
s
ult
s
we
r
e
then
c
onve
r
ted
int
o
nume
r
ica
l
f
e
a
tur
e
s
,
s
uc
h
a
s
ve
hicle
c
ounts
by
type,
whic
h
we
r
e
f
u
r
ther
p
r
oc
e
s
s
e
d
int
o
tr
a
f
f
ic
f
low
da
ta
f
or
s
ubs
e
que
nt
a
na
lys
is
.
2.
2
.
L
on
g
s
h
or
t
-
t
e
r
m
m
e
m
or
y
-
Ko
m
od
o
M
li
p
ir
algorit
h
m
T
he
int
e
gr
a
ti
on
of
L
S
T
M
a
nd
KM
A
leve
r
a
ge
s
th
e
s
tr
e
ngths
of
e
a
c
h
method
in
da
ta
a
na
lys
is
a
nd
opti
mi
z
a
ti
on.
L
S
T
M
is
highl
y
e
f
f
e
c
ti
ve
a
t
c
a
ptur
i
ng
tempor
a
l
pa
tt
e
r
ns
in
ti
me
-
s
e
r
ies
da
ta,
making
i
t
s
uit
a
ble
f
or
both
s
hor
t
-
ter
m
a
nd
long
-
ter
m
pr
e
diction
tas
ks
[
38]
.
P
r
e
vious
s
tudi
e
s
ha
ve
s
hown
that
hype
r
pa
r
a
mete
r
opti
mi
z
a
ti
on
us
ing
meta
he
ur
is
ti
c
a
ppr
oa
c
he
s
o
f
te
n
yields
be
tt
e
r
r
e
s
ult
s
c
ompar
e
d
to
c
onve
nti
ona
l
methods
,
f
ur
ther
r
e
in
f
or
c
ing
the
a
dva
ntage
of
c
ombi
n
ing
thes
e
tec
hniques
to
im
pr
ove
model
pe
r
f
o
r
man
c
e
[
39]
.
T
he
incor
por
a
t
ion
of
KM
A
in
thi
s
a
ppr
oa
c
h
is
a
nt
icipa
ted
to
s
ur
pa
s
s
the
pe
r
f
o
r
manc
e
of
other
meta
he
ur
is
ti
c
a
lgor
it
hms
.
T
he
int
e
gr
a
ti
on
of
L
S
T
M
a
nd
KM
A
not
only
a
c
c
e
ler
a
tes
the
opti
mi
z
a
ti
on
pr
oc
e
s
s
but
a
ls
o
e
nha
nc
e
s
the
li
ke
li
hood
of
identif
ying
op
ti
m
a
l
hype
r
pa
r
a
mete
r
c
onf
igur
a
ti
ons
,
ther
e
by
s
ig
nif
ica
ntl
y
im
pr
oving
the
pe
r
f
or
manc
e
o
f
the
L
S
T
M
model
in
t
r
a
f
f
ic
f
low
p
r
e
diction
a
ppli
c
a
ti
ons
.
T
his
p
r
opos
e
d
a
ppr
oa
c
h
is
de
picte
d
in
F
igu
r
e
2
.
F
igur
e
2.
P
r
opos
e
d
m
ode
l
F
igur
e
2
il
lus
tr
a
tes
the
L
S
T
M
-
KM
A
c
omput
a
ti
on
pr
oc
e
s
s
,
be
ginni
ng
with
da
ta
r
e
a
ding
a
nd
pr
e
pr
oc
e
s
s
ing,
f
oll
owe
d
by
divi
ding
the
da
tas
e
t
int
o
tr
a
ini
ng
a
nd
tes
ti
ng
s
e
ts
,
a
ll
oc
a
ti
ng
80%
f
o
r
tr
a
i
ning
a
nd
20%
f
o
r
tes
ti
ng
[
40
]
,
[
41
]
.
T
he
KM
A
is
then
ini
ti
a
li
z
e
d
with
s
pe
c
if
ic
pa
r
a
mete
r
s
.
T
his
s
tep
invol
ve
s
ini
ti
a
li
z
ing
a
population
o
f
c
a
ndidate
s
olut
ions
a
nd
a
pplyi
ng
c
r
os
s
ove
r
a
nd
mut
a
ti
on
ope
r
a
ti
ons
[
42]
.
T
he
f
it
ne
s
s
of
e
a
c
h
c
a
ndidate
s
olut
ion
is
e
va
luate
d
to
de
ter
mi
ne
the
s
uit
a
bil
it
y
o
f
the
pa
r
a
mete
r
s
f
or
th
e
L
S
T
M
mo
de
l
us
ing
the
mea
n
a
bs
olut
e
e
r
r
o
r
(
M
AE
)
me
tr
ic.
T
he
L
S
T
M
pa
r
a
mete
r
s
be
ing
opti
mi
z
e
d
incl
ude
the
number
of
ne
ur
ons
,
lea
r
ning
r
a
te,
a
nd
e
poc
hs
[
26
]
.
KM
A
it
e
r
a
ti
ve
ly
upda
tes
the
c
a
ndidate
s
olut
ions
thr
ough
a
n
opti
mi
z
a
ti
on
loop
unti
l
the
opti
mal
pa
r
a
mete
r
s
a
r
e
identif
ied.
T
he
opti
mi
z
e
d
pa
r
a
mete
r
s
a
r
e
then
a
ppli
e
d
t
o
tr
a
in
the
f
inal
L
S
T
M
model.
T
he
t
r
a
ined
model
i
s
s
ubs
e
que
ntl
y
tes
ted
u
s
ing
the
tes
t
da
ta,
with
r
o
ot
mea
n
s
qua
r
e
e
r
r
or
(
R
M
S
E
)
a
nd
M
AE
c
a
lcula
ted
a
s
a
c
c
u
r
a
c
y
mea
s
ur
e
s
f
o
r
the
p
r
e
dictions
.
T
he
c
onc
e
pt
of
KM
A
in
L
S
T
M
pa
r
a
mete
r
opti
mi
z
a
ti
on
is
il
lus
tr
a
ted
th
r
oug
h
the
ps
e
udoc
ode
pr
e
s
e
nted
in
Algor
it
h
m
1
.
Algor
it
hm
1:
Komodo
M
li
pir
f
o
r
o
pti
mi
z
ing
L
S
T
M
p
a
r
a
mete
r
s
Input:
Maximum number of iterations
(
)
, population size
(
)
.
Range of LSTM parameters to be optimized (neurons, learning rate, and epochs).
Step 1: Initialization
Initialize
a
population
of
individuals
(komodo)
with
random
combinations
of
L
STM
parameters.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
r
aff
ic
fl
ow
pr
e
diction
us
ing
long
s
hor
t
-
ter
m
me
mor
y
-
K
omodo
M
li
pir
A
lgor
it
hm:
…
(
I
mam
A
hmad
A
s
har
i)
3347
Each
individual
in
the
population
is
represented
as
=
[
,
,
]
,
where
,
,
and
respectively denote neurons, learning rate, and epochs.
Step 2: Fitness Evaluation
Evaluate
the
initial
fitness
of
each
c
andidate
solution
by
measuring
the
LSTM
’s
performance on the validation dataset.
Use the objective function: Minimize F=MAE.
Sort
the
individuals
based
on
their
fit
ness
scores
and
categorize
them
into
th
ree
groups:
-
Large males (elite, top performers)
-
Females (moderate performance)
-
Small males (low performers)
Step 3: Main Loop
ℎ
(
≤
)
:
1.
Reassess each individual's
fitness score.
2.
Update their positions as follows:
-
Large males: Adjust positions using exploitation strategies.
-
Females:
-
Mate with the top
-
performing large male using exploitation method.
-
Reproduce asexually via parthenogenesis using exploration strategies.
-
Small
males:
Explore
the
solution
sp
ace
randomly
using
exploration
strategies.
3.
Apply selection process:
-
Retain the best
-
performing individual (elitism).
-
Improve weaker individuals using update strategy in equation.
4.
Increment the iteration count
(
=
+
1
)
.
End
While
Step 4: Output the Best Solution
Output the best LSTM parameters (
_
) and the best fitness value (
_
).
Output:
Optimal LSTM parameters
Algor
it
hm
1
is
the
ps
e
udoc
ode
of
the
KM
A
us
e
d
to
opti
mi
z
e
the
pa
r
a
mete
r
s
of
the
L
S
T
M
model.
T
his
a
lgor
it
hm
a
im
s
to
f
ind
the
be
s
t
c
ombi
na
ti
on
of
ne
ur
ons
,
lea
r
ning
r
a
te,
a
nd
number
of
e
p
oc
hs
by
mi
nim
izing
the
M
AE
.
T
he
s
tep
-
by
-
s
tep
pr
oc
e
dur
e
is
outl
ined
in
the
ps
e
udoc
ode
a
bove
.
T
o
f
a
c
il
it
a
te
unde
r
s
tanding,
the
wor
k
f
low
o
f
thi
s
a
lgo
r
it
hm
is
a
ls
o
il
lus
tr
a
ted
in
the
f
lowc
ha
r
t
,
a
s
s
hown
in
F
igur
e
3.
F
igur
e
3.
Optim
iza
ti
on
w
or
kf
low
of
the
L
S
T
M
mo
de
l
us
ing
KM
A
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
334
3
-
3353
3348
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
3.
1.
Dat
a
an
d
e
n
vironm
e
n
t
I
n
t
his
s
t
ud
y
,
t
he
da
ta
us
e
d
f
o
r
t
r
a
f
f
i
c
f
l
ow
p
r
e
d
ic
t
io
n
wa
s
c
o
ll
e
c
t
e
d
f
r
om
C
C
T
V
c
a
m
e
r
a
s
ins
ta
ll
e
d
in
F
a
tm
a
wa
ti
,
S
e
ma
r
a
ng
C
it
y
.
T
he
da
ta
c
o
l
lec
t
io
n
pe
r
io
d
s
p
a
n
ne
d
f
r
o
m
De
c
e
mbe
r
19
,
2
02
3
,
t
o
F
e
b
r
u
a
r
y
15
,
20
24
.
Da
ta
wa
s
ga
ther
e
d
by
e
xtr
a
c
ti
ng
i
mage
s
f
r
om
r
e
c
or
de
d
videos
a
t
5
-
mi
nute
int
e
r
va
ls
.
T
he
f
r
a
me
int
e
r
va
l
wa
s
de
ter
mi
ne
d
by
mul
t
ipl
ying
the
f
r
a
mes
pe
r
s
e
c
ond
(
F
P
S
)
by
60
a
nd
the
s
pe
c
if
ied
number
o
f
m
inut
e
s
.
W
it
h
a
n
F
P
S
of
25
,
the
r
e
s
ult
ing
f
r
a
me
int
e
r
va
l
wa
s
25×
60×
5=
7,
500
f
r
a
mes
.
T
his
mea
ns
the
pr
ogr
a
m
e
xtr
a
c
ted
one
im
a
ge
f
or
e
ve
r
y
7
,
500
f
r
a
mes
.
F
r
om
thi
s
e
xtr
a
c
ti
o
n
pr
oc
e
s
s
,
a
tot
a
l
of
720
im
a
ge
s
we
r
e
obtaine
d.
A
tot
a
l
of
45,
347
a
nnotations
we
r
e
ge
ne
r
a
ted
f
r
om
thes
e
im
a
ge
s
,
c
ompr
is
ing
31,
481
mot
o
r
c
yc
les
,
12,
4
02
c
a
r
s
,
1,
184
tr
uc
ks
,
a
nd
280
bus
e
s
.
I
n
a
ddit
ion
to
ve
hicl
e
types
,
da
te
a
nd
ti
me
inf
or
mation
wa
s
a
ls
o
e
xtr
a
c
ted
f
r
om
the
da
tas
e
t.
T
he
da
tas
e
t
wa
s
then
divi
de
d
int
o
t
wo
s
ubs
e
ts
:
a
tr
a
ini
ng
s
ubs
e
t
a
nd
a
va
li
da
ti
on
s
ubs
e
t,
to
f
a
c
il
it
a
te
model
tr
a
ini
ng
a
nd
e
va
luation.
T
he
d
is
tr
ibut
ion
of
ve
hicle
a
nnotations
in
the
t
r
a
ini
ng
s
e
t
include
s
22,
136
mot
o
r
c
yc
les
,
8
,
804
c
a
r
s
,
839
tr
uc
ks
,
a
nd
1
99
bus
e
s
.
T
he
va
li
da
ti
on
s
e
t
c
ons
is
ts
of
9
,
345
mot
or
c
yc
les
,
3,
598
c
a
r
s
,
345
tr
uc
ks
,
a
nd
81
bus
e
s
.
T
his
s
tr
uc
t
ur
e
d
da
ta
c
oll
e
c
ti
on
a
nd
pr
e
pr
oc
e
s
s
ing
pr
oc
e
s
s
pr
ovides
a
s
oli
d
f
ounda
ti
on
f
o
r
de
ve
lopi
ng
tr
a
f
f
ic
f
low
pr
e
dic
ti
on
models
,
e
ns
ur
ing
that
the
da
tas
e
t
is
r
e
pr
e
s
e
ntative
a
nd
we
ll
-
a
nnotate
d
f
or
e
f
f
e
c
ti
ve
model
tr
a
ini
ng
a
nd
e
v
a
luation.
T
his
s
tudy
uti
li
z
e
d
Google
C
olab
P
r
o
f
o
r
e
xpe
r
i
menta
l
c
onf
igur
a
ti
on
.
Google
C
olab
of
f
e
r
s
c
loud
-
ba
s
e
d
a
nd
ope
n
-
s
our
c
e
c
omput
ing
s
e
r
vice
s
to
ha
nd
le
the
e
xtens
ive
pr
oc
e
s
s
ing
r
e
quir
e
ments
ne
e
de
d
f
o
r
model
tr
a
ini
ng
[
43]
.
T
he
r
unti
me
e
nvir
onment
include
d
P
ython
3
a
nd
a
n
NV
I
DI
A
T
4
GPU.
T
he
pr
og
r
a
mm
ing
langua
ge
uti
li
z
e
d
wa
s
P
ython
3.
10.
12,
a
nd
the
P
yT
or
c
h
f
r
a
mew
or
k
ve
r
s
ion
2.
3
.
0
wa
s
im
pleme
nted
with
C
UD
A
ve
r
s
ion
12.
1
s
uppor
t
.
3.
2.
P
ar
a
m
e
t
e
r
s
e
t
t
in
gs
an
d
m
od
e
l
op
t
im
izat
io
n
I
n
thi
s
s
tudy,
the
pa
r
a
mete
r
s
f
o
r
the
meta
he
ur
is
ti
c
method
we
r
e
s
tanda
r
dize
d
by
s
e
tt
ing
the
population
s
ize
to
30,
a
s
r
e
f
e
r
e
nc
e
d
in
p
r
e
vious
s
t
udies
[
26]
.
T
he
pa
r
a
mete
r
r
a
nge
s
opti
mi
z
e
d
f
or
th
e
L
S
T
M
model
include
the
number
o
f
ne
u
r
ons
(
300
-
500)
,
lea
r
ning
r
a
te
(
0.
001
-
0.
01
)
,
a
nd
number
of
e
poc
hs
(
1
-
150)
.
T
he
s
e
r
a
nge
s
we
r
e
ini
ti
a
ll
y
a
dopted
ba
s
e
d
on
p
r
ior
li
ter
a
tu
r
e
a
nd
then
f
u
r
ther
r
e
f
ined
th
r
ough
mul
ti
ple
tr
ial
-
a
nd
-
e
r
r
or
e
xpe
r
im
e
nts
to
obtain
opti
mal
pe
r
f
or
manc
e
.
F
o
r
the
c
onve
nti
ona
l
L
S
T
M
model,
the
a
na
lys
is
wa
s
c
onduc
ted
us
ing
the
highes
t
va
lues
in
e
a
c
h
r
a
nge
-
500
ne
ur
ons
,
a
lea
r
ning
r
a
te
of
0.
01
,
a
nd
150
e
poc
hs
.
De
tailed
c
onf
igur
a
ti
ons
of
other
pa
r
a
mete
r
s
us
e
d
f
or
e
a
c
h
model
c
a
n
be
f
ound
in
T
a
ble
2
.
T
a
ble
2
p
r
e
s
e
nts
the
pa
r
a
mete
r
s
e
tt
ings
us
e
d
f
o
r
v
a
r
ious
a
lgor
it
hms
in
opti
mi
z
ing
the
L
S
T
M
model
.
F
or
L
S
T
M
-
KM
A,
the
s
ize
of
the
population
invol
v
e
d
in
the
s
e
lec
ti
on
pr
oc
e
s
s
is
s
e
t
to
10
.
I
n
the
L
S
T
M
-
I
DB
O
a
lgor
i
thm
,
the
c
oe
f
f
icie
nt
o
f
va
r
iation
is
s
e
t
to
0
.
1,
a
nd
the
s
c
a
li
ng
pa
r
a
mete
r
f
or
ba
lanc
ing
e
xplor
a
ti
on
a
nd
e
xploi
tation
is
s
e
t
to
0.
5.
T
he
L
S
T
M
-
P
S
O
a
lgor
it
hm
us
e
s
a
s
e
lf
-
le
a
r
ning
f
a
c
tor
of
1
.
5
a
nd
a
g
r
oup
lea
r
ning
f
a
c
tor
of
2.
T
a
ble
2.
P
a
r
a
mete
r
s
e
tt
ing
of
the
va
r
ious
a
lgo
r
it
hm
s
A
lg
or
it
hm
P
a
r
a
me
te
r
s
S
e
tt
in
gs
R
e
f
e
r
e
nc
e
L
S
T
M
-
K
M
A
S
iz
e
of
popula
ti
on i
nvol
ve
d i
n s
e
le
c
ti
on
10
[
42]
I
D
B
O
-
L
S
T
M
C
oe
f
f
ic
ie
nt
of
va
r
ia
ti
on
S
c
a
l
e
or
pa
r
a
me
te
r
f
or
s
e
tt
in
g e
xpl
or
a
ti
on a
nd e
xpl
oi
ta
ti
on
0.1
0.5
[
26]
L
S
T
M
-
PSO
S
e
lf
-
le
a
r
ni
ng f
a
c
to
r
G
r
oup le
a
r
ni
ng
f
a
c
to
r
1.5
2
[
28]
3.
3.
E
valu
at
io
n
c
r
it
e
r
ia
T
he
a
ppr
opr
iate
pe
r
f
o
r
manc
e
e
va
luation
met
r
ics
f
or
c
onti
nuous
da
ta
obtaine
d
in
r
e
a
l
-
ti
me
a
r
e
r
e
gr
e
s
s
ion
los
s
f
unc
ti
ons
[
44]
.
T
he
r
e
f
or
e
,
the
pe
r
f
or
manc
e
e
va
luation
metr
ics
us
e
d
in
thi
s
s
tudy
a
r
e
R
M
S
E
a
nd
M
AE
.
R
M
S
E
r
e
f
lec
ts
the
de
g
r
e
e
of
de
viatio
n
of
p
r
e
dicte
d
va
lues
f
r
om
a
c
tual
va
lues
.
T
he
f
or
mul
a
f
o
r
R
M
S
E
is
pr
ovided
in
(
1)
[
45
]
.
M
AE
r
e
p
r
e
s
e
nts
the
mea
n
of
a
bs
olut
e
e
r
r
or
s
,
whe
r
e
a
bs
olut
e
e
r
r
or
is
the
dif
f
e
r
e
nc
e
be
twe
e
n
pr
e
dicte
d
a
nd
a
c
tual
va
lues
.
A
low
M
AE
va
lue
indi
c
a
tes
that
the
model
pr
e
di
c
ts
va
lues
c
los
e
to
the
a
c
tual
va
lues
.
T
he
f
o
r
mul
a
f
or
M
AE
is
pr
ovided
in
(
2
)
[
26
]
.
=
√
1
∑
|
−
̂
|
=
1
2
(
1)
=
1
∑
|
−
̂
|
=
1
(
2)
R
M
S
E
a
nd
M
AE
a
r
e
two
e
va
luation
metr
ics
us
e
d
to
mea
s
ur
e
the
pr
e
diction
e
r
r
or
s
of
a
model.
I
n
the
R
M
S
E
f
or
mu
la,
the
di
f
f
e
r
e
nc
e
be
twe
e
n
the
a
c
tual
va
lue
(
)
a
nd
the
pr
e
dicte
d
va
lue
(
̂
)
is
s
qua
r
e
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
r
aff
ic
fl
ow
pr
e
diction
us
ing
long
s
hor
t
-
ter
m
me
mor
y
-
K
omodo
M
li
pir
A
lgor
it
hm:
…
(
I
mam
A
hmad
A
s
har
i)
3349
c
a
lcula
te
(
−
̂
)
2
,
givi
ng
g
r
e
a
ter
we
ight
to
lar
ge
r
e
r
r
or
s
,
a
nd
then
the
s
qua
r
e
r
oot
is
take
n.
R
M
S
E
pr
ovides
a
ddit
ional
ins
ight
s
by
r
e
f
lec
ti
ng
the
de
gr
e
e
of
de
viation
be
twe
e
n
pr
e
dicte
d
va
lues
a
nd
a
c
tual
va
lues
,
be
ing
mor
e
s
e
ns
it
ive
to
lar
ge
e
r
r
o
r
s
.
M
AE
ha
s
a
s
im
il
a
r
f
or
mul
a
bu
t
with
a
di
f
f
e
r
e
nt
a
pp
r
oa
c
h.
I
n
thi
s
f
or
mul
a
,
r
e
pr
e
s
e
nts
the
tot
a
l
number
of
da
ta
point
s
or
obs
e
r
va
ti
ons
i
n
the
da
tas
e
t,
indi
c
a
ti
ng
the
nu
mber
of
da
ta
point
s
a
na
lyze
d.
is
the
a
c
tual
va
lue
of
the
-
t
h
da
ta
point
,
r
e
pr
e
s
e
nti
ng
the
tr
ue
da
ta
to
be
pr
e
d
icte
d,
s
uc
h
a
s
the
a
c
tual
number
of
ve
hicle
s
in
tr
a
f
f
ic
pr
e
diction.
On
the
o
ther
ha
nd
,
̂
is
the
pr
e
dicte
d
va
lue
ge
ne
r
a
ted
by
the
model
f
or
the
-
th
da
ta
point
,
r
e
f
lec
ti
ng
the
e
s
ti
mate
d
number
of
ve
hicle
s
.
T
he
a
bs
olut
e
dif
f
e
r
e
nc
e
be
twe
e
n
a
c
tual
a
nd
p
r
e
dicte
d
va
lues
is
c
a
lcula
ted
a
s
|
−
̂
|
,
pr
ovidi
ng
a
n
e
r
r
or
mea
s
ur
e
without
r
e
ga
r
d
to
e
r
r
or
dir
e
c
ti
on
.
Al
l
thes
e
a
bs
olut
e
dif
f
e
r
e
nc
e
s
a
r
e
s
umm
e
d
a
nd
divi
de
d
by
the
tot
a
l
numbe
r
o
f
da
ta
point
s
(
)
to
yield
the
M
AE
.
T
he
r
e
f
or
e
,
R
M
S
E
a
nd
M
AE
p
r
ovide
a
n
ove
r
a
ll
m
e
a
s
ur
e
of
how
c
los
e
the
model's
pr
e
dictions
a
r
e
to
the
a
c
tual
va
lues
.
T
he
pr
e
diction
r
e
s
ult
s
of
the
L
S
T
M
,
L
S
T
M
-
KM
A,
L
S
T
M
-
I
DB
O
,
a
nd
L
S
T
M
-
P
S
O
models
a
r
e
c
ompar
e
d
with
the
a
c
tual
da
ta.
T
he
pr
e
dictio
n
outcome
s
o
f
the
uti
li
z
e
d
models
a
r
e
s
hown
in
F
igur
e
4
.
F
igur
e
4
il
lus
tr
a
tes
the
c
ompar
is
on
be
twe
e
n
the
a
c
tual
tr
a
f
f
ic
f
low
da
ta
(
T
R
UE
)
a
nd
the
pr
e
dicte
d
r
e
s
ult
s
f
r
om
s
e
ve
r
a
l
models
,
na
mely
L
S
T
M
,
L
S
T
M
-
KM
A,
L
S
T
M
-
I
DB
O,
a
nd
L
S
T
M
-
P
S
O.
T
he
gr
a
ph
s
hows
how
e
a
c
h
model's
pr
e
dictions
a
li
gn
with
or
de
viate
f
r
o
m
the
a
c
tual
tr
a
f
f
ic
f
low
va
lues
ove
r
t
im
e
,
highl
ig
hti
ng
the
a
c
c
ur
a
c
y
a
nd
pe
r
f
or
manc
e
di
f
f
e
r
e
nc
e
s
a
mong
the
models
.
F
igur
e
4
.
P
r
e
diction
r
e
s
ult
s
of
e
a
c
h
model
3.
4.
Re
s
u
lt
s
an
d
p
e
r
f
or
m
an
c
e
a
n
alys
is
T
he
de
ve
loped
model,
L
S
T
M
-
KM
A,
is
c
ompar
e
d
with
the
ba
s
e
li
ne
L
S
T
M
model.
I
n
a
ddit
ion
,
two
other
L
S
T
M
models
opti
mi
z
e
d
us
ing
meta
he
ur
is
ti
c
a
lgor
it
hms
,
na
mely
L
S
T
M
-
I
DB
O
a
nd
L
S
T
M
-
P
S
O,
a
r
e
a
ls
o
include
d
in
the
c
ompar
is
on.
T
he
pe
r
f
or
manc
e
of
e
a
c
h
model
ba
s
e
d
on
the
R
M
S
E
is
s
hown
in
F
ig
ur
e
5.
A
s
e
pa
r
a
te
c
ompar
is
on
us
ing
the
M
AE
metr
ic
is
p
r
e
s
e
nted
in
F
igu
r
e
6.
T
his
f
igu
r
e
highl
ight
s
the
a
ve
r
a
ge
pr
e
diction
e
r
r
or
f
o
r
e
a
c
h
model.
T
he
lowe
r
the
M
AE
va
lue,
the
c
los
e
r
the
model's
pr
e
dictions
a
r
e
to
th
e
a
c
tual
da
ta.
T
o
s
uppor
t
the
vis
ua
l
c
ompar
is
on,
both
R
M
S
E
a
nd
M
AE
va
lues
a
r
e
s
umm
a
r
ize
d
in
T
a
ble
3
.
T
his
table
pr
ovides
a
c
lea
r
e
r
view
of
e
a
c
h
model’
s
numer
ica
l
pe
r
f
or
manc
e
.
I
t
c
ompl
e
ments
the
gr
a
phica
l
r
e
s
ult
s
s
hown
in
the
F
igu
r
e
s
5
a
nd
6
.
F
igur
e
5.
R
M
S
E
c
ompar
is
on
of
models
F
igur
e
6.
M
AE
c
ompar
is
on
of
models
0
20
40
60
80
1
0
0
1
2
0
1
4
0
1
6
0
0
7
:
0
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
7
:
0
0
:
0
0
0
6
:
0
5
:
0
0
0
6
:
2
0
:
0
0
0
6
:
3
5
:
0
0
0
6
:
5
0
:
0
0
0
6
:
0
5
:
0
0
0
6
:
2
0
:
0
0
0
6
:
3
5
:
0
0
0
6
:
5
0
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
7
:
0
0
:
0
0
0
6
:
1
0
:
0
0
0
6
:
2
5
:
0
0
0
6
:
4
0
:
0
0
0
6
:
5
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
7
:
0
0
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
0
6
:
0
0
:
0
0
0
6
:
1
5
:
0
0
0
6
:
3
0
:
0
0
0
6
:
4
5
:
0
0
T
r
a
f
f
i
c
F
l
o
w
V
e
h
i
c
l
e
T
i
m
e
/
5
m
i
n
T
R
U
E
L
S
T
M
L
S
T
M
-
K
M
A
L
S
T
M
-
I
D
B
O
L
S
T
M
-
P
S
O
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
334
3
-
3353
3350
T
a
ble
3.
T
he
R
M
S
E
a
nd
M
AE
v
a
lues
of
the
model
s
we
r
e
e
va
luate
d
indi
vidually
M
ode
l
R
M
S
E
M
A
E
L
S
T
M
16.6827
9.9903
L
S
T
M
-
K
M
A
14.5319
8.7041
L
S
T
M
-
I
D
B
O
15.0946
9.0328
L
S
T
M
-
PSO
15.0368
9.0015
T
a
ble
3
s
h
ows
tha
t
t
he
L
S
T
M
-
KM
A
mo
de
l
a
c
hiev
e
s
the
l
owe
s
t
R
M
S
E
va
lue
o
f
14
.
5
319
,
indi
c
a
ti
n
g
the
be
s
t
pe
r
f
o
r
ma
nc
e
c
o
mpa
r
e
d
t
o
th
e
ot
he
r
mode
l
s
.
I
n
c
on
tr
a
s
t,
the
ba
s
e
li
ne
L
S
T
M
mo
de
l
r
e
c
o
r
ds
t
h
e
highes
t
R
M
S
E
va
lue
of
16
.
68
27
,
i
ndica
ti
n
g
th
e
lo
we
s
t
p
r
e
d
ictio
n
a
c
c
u
r
a
c
y
.
T
he
L
S
T
M
-
I
DB
O
a
n
d
L
S
T
M
-
P
S
O
mode
ls
a
c
hieve
R
M
S
E
va
lues
of
15
.
09
46
a
nd
15
.
03
68
,
r
e
s
pe
c
ti
ve
ly
,
de
mons
t
r
a
t
ing
i
mpr
ove
d
pe
r
f
o
r
ma
nc
e
c
ompa
r
e
d
to
t
he
ba
s
e
li
ne
L
S
T
M
bu
t
s
ti
ll
f
a
ll
in
g
s
ho
r
t
o
f
L
S
T
M
-
KM
A.
I
n
te
r
ms
of
M
AE
,
L
S
T
M
-
KM
A
a
ls
o
de
m
ons
tr
a
tes
the
be
s
t
pe
r
f
o
r
manc
e
w
it
h
the
lowe
s
t
va
lue
o
f
8
.
7
0
41,
c
ompa
r
e
d
to
the
ba
s
e
li
ne
L
S
T
M
(
9
.
9
903
)
,
L
S
T
M
-
I
DB
O
(
9
.
032
8)
,
a
nd
L
S
T
M
-
P
S
O
(
9.
0015
)
.
B
a
s
e
d
o
n
thi
s
a
na
lys
is
,
opt
im
iza
ti
o
n
us
ing
t
he
KM
A
ha
s
p
r
ove
n
t
o
be
t
he
mos
t
e
f
f
e
c
ti
ve
meth
od
f
o
r
e
nha
nc
ing
t
he
p
e
r
f
o
r
ma
nc
e
o
f
the
L
S
T
M
m
ode
l
in
te
r
ms
o
f
both
R
M
S
E
a
nd
M
A
E
,
makin
g
it
the
r
e
c
om
mende
d
a
pp
r
oa
c
h
i
n
thi
s
s
tu
dy
.
3.
5.
Chal
lenges
Obje
c
t
de
tec
ti
on
us
ing
YO
L
O
on
thi
s
da
tas
e
t
f
a
c
e
s
s
e
ve
r
a
l
ke
y
c
ha
ll
e
nge
s
,
pr
im
a
r
il
y
due
to
va
r
iations
in
li
ghti
ng
a
nd
tr
a
f
f
ic
de
ns
it
y.
L
ight
r
e
f
lec
ti
ons
,
poor
i
ll
umi
na
ti
on,
a
nd
high
t
r
a
f
f
ic
c
o
nge
s
ti
on
s
igni
f
ica
ntl
y
r
e
duc
e
de
tec
ti
on
a
c
c
ur
a
c
y.
Additi
ona
ll
y,
the
p
r
e
s
e
nc
e
of
mul
ti
ple
objec
t
types
in
a
s
ing
le
f
r
a
me
s
uc
h
a
s
mot
or
c
yc
les
,
c
a
r
s
,
tr
uc
ks
,
a
nd
bus
e
s
make
s
it
dif
f
icult
f
o
r
the
model
to
dis
ti
nguis
h
ove
r
lapping
objec
ts
,
e
s
pe
c
ially
f
or
les
s
domi
na
nt
c
las
s
e
s
li
ke
tr
uc
ks
a
nd
bus
e
s
.
M
e
a
nwhile,
the
us
e
o
f
meta
he
ur
is
ti
c
a
lgor
it
hms
to
opti
mi
z
e
L
S
T
M
pa
r
a
mete
r
s
yields
be
tt
e
r
pr
e
diction
a
c
c
ur
a
c
y.
How
e
ve
r
,
the
dr
a
wba
c
k
li
e
s
in
the
longer
r
unti
me
c
ompar
e
d
to
c
onve
nti
ona
l
methods
,
a
s
the
s
e
a
r
c
h
f
or
opti
mal
pa
r
a
mete
r
s
invol
ve
s
c
ompl
e
x
a
nd
it
e
r
a
ti
ve
p
r
oc
e
s
s
e
s
.
3.
6.
P
r
ac
t
ical
i
m
p
li
c
at
ion
s
f
or
in
t
e
ll
i
ge
n
t
t
r
an
s
p
or
t
at
ion
s
ys
t
e
m
s
d
e
p
loy
m
e
n
t
T
he
pr
opos
e
d
YO
L
O
-
L
S
T
M
-
K
M
A
f
r
a
mew
or
k
de
mons
tr
a
tes
pr
omi
s
ing
potential
f
or
r
e
a
l
-
wor
ld
de
ploym
e
nt
in
I
T
S
.
B
y
int
e
gr
a
ti
ng
r
e
a
l
-
ti
me
objec
t
de
tec
ti
on
with
ti
me
-
s
e
r
ies
tr
a
f
f
ic
pr
e
diction
,
thi
s
a
ppr
oa
c
h
s
uppor
ts
a
utom
a
ted
t
r
a
f
f
ic
moni
tor
ing
a
nd
da
ta
-
dr
iven
de
c
is
ion
-
making.
How
e
ve
r
,
s
e
ve
r
a
l
pr
a
c
ti
c
a
l
a
s
pe
c
ts
mus
t
be
c
ons
ider
e
d:
‒
S
c
a
labili
ty:
t
he
f
r
a
mew
or
k
is
de
s
igned
to
ha
ndle
l
a
r
ge
volum
e
s
o
f
tr
a
f
f
ic
video
da
ta
,
making
it
s
uit
a
ble
f
or
de
ploym
e
nt
in
u
r
ba
n
e
nvir
onments
with
hig
h
tr
a
f
f
ic
de
ns
it
y.
How
e
ve
r
,
the
a
nnotation
a
nd
tr
a
in
ing
pr
oc
e
s
s
s
ti
ll
r
e
quir
e
c
ons
ider
a
ble
e
f
f
or
t
,
whic
h
ma
y
ne
e
d
a
utom
a
ti
on
o
r
s
e
mi
-
s
upe
r
vis
e
d
tec
hniques
f
or
br
oa
de
r
s
c
a
labili
ty.
‒
C
omput
a
ti
ona
l
r
e
quir
e
ments
:
r
eal
-
ti
me
de
tec
ti
on
us
ing
YO
L
O
a
nd
pr
e
diction
wi
th
L
S
T
M
-
K
M
A
de
mands
s
uf
f
icie
nt
c
omput
a
ti
ona
l
r
e
s
our
c
e
s
,
pa
r
ti
c
ular
ly
dur
ing
model
t
r
a
ini
ng
a
nd
opt
im
iza
ti
on.
De
ploym
e
nt
in
the
f
ield
would
r
e
quir
e
e
dge
c
om
puti
ng
or
c
loud
-
ba
s
e
d
inf
r
a
s
tr
uc
tur
e
to
mee
t
late
nc
y
c
ons
tr
a
int
s
,
e
s
pe
c
ially
f
or
c
onti
nuous
t
r
a
f
f
ic
f
low
a
na
lys
is
.
‒
I
ntegr
a
ti
on
c
ha
ll
e
nge
s
:
i
ntegr
a
ti
ng
thi
s
model
int
o
e
xis
ti
ng
I
T
S
inf
r
a
s
tr
uc
tur
e
may
invol
ve
c
ha
ll
e
nge
s
s
uc
h
a
s
da
ta
c
ompatibi
li
ty,
s
ync
hr
oniza
ti
on
a
c
r
o
s
s
s
e
ns
or
s
a
nd
c
a
mer
a
s
,
a
nd
e
ns
ur
ing
r
e
li
a
bil
it
y
in
va
r
iable
c
ondit
ions
(
e
.
g.
,
we
a
ther
,
li
g
hti
ng
,
a
n
d
oc
c
lus
ion)
.
R
obus
t
pr
e
pr
oc
e
s
s
ing
a
nd
a
da
pti
ve
r
e
tr
a
ini
ng
s
tr
a
tegie
s
c
ould
he
lp
mi
ti
ga
te
thes
e
is
s
ue
s
.
3.
7.
Com
p
ar
is
on
wit
h
r
e
lat
e
d
s
t
u
d
ies
T
he
r
e
s
ult
s
of
thi
s
s
tudy
we
r
e
c
ompar
e
d
with
s
e
ve
r
a
l
e
xis
ti
ng
a
ppr
oa
c
he
s
in
the
li
ter
a
tur
e
:
‒
B
a
s
e
li
ne
a
nd
c
onve
nti
ona
l
models
:
tr
a
dit
ional
L
S
T
M
models
of
ten
s
tr
uggle
with
opt
im
izi
ng
hype
r
pa
r
a
mete
r
s
e
f
f
e
c
ti
ve
ly
,
lea
ding
to
s
ubopti
mal
pr
e
dictions
.
T
he
int
e
gr
a
ti
on
of
KM
A
in
thi
s
s
t
udy
outper
f
or
ms
the
s
tanda
r
d
L
S
T
M
by
a
c
hieving
hi
ghe
r
a
c
c
ur
a
c
y
a
nd
be
t
ter
ge
ne
r
a
li
z
a
ti
on
,
pa
r
ti
c
ula
r
ly
unde
r
c
ompl
e
x
tr
a
f
f
ic
s
c
e
na
r
ios
.
‒
M
e
tahe
ur
is
ti
c
-
ba
s
e
d
model
s
:
c
ompar
e
d
to
other
meta
he
ur
is
ti
c
-
int
e
gr
a
ted
models
s
uc
h
a
s
L
S
T
M
-
I
DB
O
a
nd
L
S
T
M
-
P
S
O,
the
pr
opos
e
d
L
S
T
M
-
KM
A
model
pr
ovides
c
ompetit
ive
or
s
upe
r
ior
pe
r
f
o
r
manc
e
i
n
ter
m
s
of
p
r
e
diction
a
c
c
ur
a
c
y.
How
e
ve
r
,
li
ke
other
meta
he
ur
is
ti
c
a
ppr
oa
c
he
s
,
i
t
incur
s
a
hi
ghe
r
c
omput
a
ti
ona
l
c
os
t
due
to
i
ts
it
e
r
a
ti
ve
s
e
a
r
c
h
mec
h
a
nis
m.
‒
Adva
nc
e
ments
a
nd
dif
f
e
r
e
nc
e
s
:
un
li
ke
p
r
e
vious
w
or
ks
that
f
oc
us
e
it
he
r
on
de
tec
ti
on
or
p
r
e
diction
a
l
one
,
thi
s
s
tudy
pr
e
s
e
nts
a
c
ompl
e
te
pipeline
f
r
om
r
e
a
l
-
t
im
e
ve
hicle
de
tec
ti
on
to
t
r
a
f
f
ic
f
low
pr
e
diction.
T
his
int
e
gr
a
ted
de
s
ign
c
ontr
ibut
e
s
to
im
pr
ove
d
pe
r
f
or
m
a
nc
e
a
nd
pr
a
c
ti
c
a
l
a
ppli
c
a
bil
it
y
f
or
I
T
S
,
a
li
gn
ing
with
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
T
r
aff
ic
fl
ow
pr
e
diction
us
ing
long
s
hor
t
-
ter
m
me
mor
y
-
K
omodo
M
li
pir
A
lgor
it
hm:
…
(
I
mam
A
hmad
A
s
har
i)
3351
the
dir
e
c
ti
on
o
f
r
e
c
e
nt
s
tudi
e
s
while
int
r
oduc
in
g
a
nove
l
opti
mi
z
a
ti
on
a
lgor
it
hm
tailo
r
e
d
to
tr
a
f
f
ic
d
a
ta
c
ha
r
a
c
ter
is
ti
c
s
.
Ove
r
a
ll
,
the
s
tudy
c
ontr
ibut
e
s
to
b
r
idgi
ng
th
e
ga
p
be
twe
e
n
a
c
a
de
mi
c
models
a
nd
r
e
a
l
-
wor
ld
I
T
S
im
pleme
ntation
by
a
ddr
e
s
s
ing
both
de
tec
ti
on
a
c
c
ur
a
c
y
a
nd
pr
e
diction
r
obus
tnes
s
,
while
a
c
knowle
dging
the
tr
a
de
-
of
f
s
in
c
omput
a
ti
on
a
nd
int
e
gr
a
ti
on
c
ompl
e
xit
y.
4.
CONC
L
USI
ON
M
ult
i
-
tar
ge
t
ve
hicle
de
tec
ti
on
in
ur
ba
n
tr
a
f
f
ic
f
a
c
e
s
s
igni
f
ica
nt
c
ha
ll
e
nge
s
,
including
poor
li
ghti
ng
,
s
mall
objec
t
s
ize
s
,
a
nd
va
r
iations
in
ve
hicle
types
,
a
ll
of
whic
h
a
f
f
e
c
t
the
a
c
c
ur
a
c
y
of
tr
a
f
f
ic
f
low
pr
e
dictions
.
T
o
a
ddr
e
s
s
thes
e
c
ha
ll
e
nge
s
,
thi
s
s
tudy
pr
opos
e
s
the
us
e
of
a
L
S
T
M
model
opti
m
ize
d
with
the
K
M
A.
T
he
a
na
lys
is
r
e
s
ult
s
s
ho
w
that
the
L
S
T
M
-
KM
A
model
a
c
hieve
s
the
lowe
s
t
R
M
S
E
of
14.
5319,
outper
f
o
r
mi
ng
the
ba
s
e
li
ne
L
S
T
M
(
16.
6827)
,
L
S
T
M
-
I
DB
O
(
15.
094
6)
,
a
nd
L
S
T
M
-
P
S
O
(
15
.
0368)
.
F
ur
ther
mor
e
,
L
S
T
M
-
K
M
A
a
ls
o
de
li
ve
r
s
the
be
s
t
pe
r
f
or
manc
e
ba
s
e
d
on
the
M
AE
,
with
the
lowe
s
t
va
lue
of
8
.
7041,
s
upe
r
i
or
to
the
ba
s
e
li
ne
L
S
T
M
(
9.
9903)
,
L
S
T
M
-
I
DB
O
(
9.
032
8)
,
a
nd
L
S
T
M
-
P
S
O
(
9.
0015)
.
T
hi
s
de
mons
tr
a
tes
that
opti
mi
z
a
ti
on
us
ing
KM
A
s
igni
f
ica
ntl
y
im
p
r
ove
s
the
a
c
c
ur
a
c
y
o
f
the
L
S
T
M
model's
p
r
e
diction
s
whe
n
a
ddr
e
s
s
ing
the
c
ompl
e
xit
y
of
mul
ti
-
tar
ge
t
ve
hicl
e
de
tec
ti
on
in
ur
ba
n
t
r
a
f
f
ic
.
T
hus
,
thi
s
r
e
s
e
a
r
c
h
make
s
a
s
igni
f
ica
nt
c
ontr
ibut
ion
to
the
de
ve
lopm
e
nt
of
p
r
e
dictive
models
that
not
only
a
ddr
e
s
s
the
c
ha
ll
e
nge
s
in
mul
ti
-
tar
ge
t
ve
hicle
de
tec
ti
on
but
a
ls
o
s
uppor
t
r
e
a
l
-
ti
me
tr
a
f
f
ic
mana
ge
ment
s
ys
tems
.
AC
KNOWL
E
DGE
M
E
NT
S
W
e
would
li
ke
to
e
xpr
e
s
s
our
g
r
a
ti
tude
to
the
De
pa
r
tm
e
nt
of
T
r
a
ns
por
tation
o
f
S
e
mar
a
ng
C
it
y
f
o
r
their
a
s
s
is
tanc
e
a
nd
s
uppor
t
in
pr
ovidi
ng
the
va
lua
ble
da
tas
e
t
us
e
d
in
thi
s
r
e
s
e
a
r
c
h.
F
UN
DI
NG
I
NF
ORM
AT
I
ON
T
he
a
uthor
s
a
f
f
i
r
m
that
thi
s
r
e
s
e
a
r
c
h
wa
s
c
onduc
t
e
d
with
pe
r
s
ona
l
f
unding
a
nd
did
not
r
e
c
e
ive
a
ny
f
inanc
ial
s
upp
or
t
f
r
om
e
xter
na
l
ins
ti
tut
ions
,
f
und
i
ng
a
ge
nc
ies
,
or
gr
a
nt
p
r
ogr
a
ms
.
All
e
xpe
ns
e
s
r
e
late
d
to
the
e
xe
c
uti
on
of
thi
s
s
tudy
we
r
e
bor
ne
s
olely
by
the
a
u
thor
s
.
AU
T
HO
R
CONT
RI
B
U
T
I
ONS
S
T
AT
E
M
E
N
T
T
his
jour
na
l
us
e
s
the
C
ontr
ibut
o
r
R
oles
T
a
xo
nomy
(
C
R
e
diT
)
to
r
e
c
ogni
z
e
indi
vidual
a
uthor
c
ontr
ibut
ions
,
r
e
duc
e
a
utho
r
s
hip
dis
putes
,
a
nd
f
a
c
il
it
a
te
c
oll
a
bor
a
ti
on.
Nam
e
of
Au
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
I
mam
Ahma
d
As
ha
r
i
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
W
a
hyul
Amien
S
ya
f
e
i
✓
✓
✓
✓
✓
✓
✓
✓
Adi
W
ibowo
✓
✓
✓
C
:
C
onc
e
pt
ua
li
z
a
ti
on
M
:
M
e
th
odol
ogy
So
:
So
f
twa
r
e
Va
:
Va
li
da
ti
on
Fo
:
Fo
r
ma
l
a
na
ly
s
is
I
:
I
nve
s
ti
ga
ti
on
R
:
R
e
s
our
c
e
s
D
:
D
a
ta
C
ur
a
ti
on
O
:
W
r
it
in
g
-
O
r
ig
in
a
l
D
r
a
f
t
E
:
W
r
it
in
g
-
R
e
vi
e
w
&
E
di
ti
ng
Vi
:
Vi
s
ua
li
z
a
ti
on
Su
:
Su
pe
r
vi
s
io
n
P
:
P
r
oj
e
c
t
a
dmi
ni
s
tr
a
ti
on
Fu
:
Fu
ndi
ng a
c
qui
s
it
io
n
CONF
L
I
CT
OF
I
NT
E
RE
S
T
S
T
AT
E
M
E
N
T
T
he
a
uthor
s
de
c
lar
e
that
they
ha
ve
no
known
c
omp
e
ti
ng
f
inanc
ial
int
e
r
e
s
ts
,
pe
r
s
ona
l
r
e
lations
hips
,
o
r
non
-
f
inanc
ial
c
ompeting
int
e
r
e
s
ts
that
c
ould
ha
v
e
a
ppe
a
r
e
d
to
inf
luenc
e
the
wor
k
r
e
por
ted
in
thi
s
pa
pe
r
.
Author
s
s
tate
no
c
onf
li
c
t
of
int
e
r
e
s
t.
DA
T
A
AV
AI
L
A
B
I
L
I
T
Y
T
he
da
ta
that
s
uppor
t
the
f
ind
ings
of
thi
s
s
tudy
we
r
e
obtaine
d
f
r
om
the
De
pa
r
tm
e
nt
o
f
T
r
a
ns
por
tation
of
S
e
mar
a
ng
C
it
y.
R
e
s
tr
ictions
a
pply
to
t
he
a
va
il
a
bil
it
y
of
thes
e
da
ta,
whic
h
we
r
e
us
e
d
unde
r
pe
r
mi
s
s
ion
f
or
thi
s
s
tudy.
Da
ta
a
r
e
a
va
il
a
ble
f
r
om
the
c
or
r
e
s
ponding
a
utho
r
upon
r
e
a
s
ona
ble
r
e
que
s
t
a
nd
with
the
pe
r
mi
s
s
ion
of
the
De
pa
r
tm
e
nt
o
f
T
r
a
ns
por
tatio
n
of
S
e
mar
a
ng
C
it
y.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
334
3
-
3353
3352
RE
F
E
RE
NC
E
S
[
1]
W
.
Z
ha
ng,
R
.
Y
a
o,
Y
.
Y
ua
n,
X
.
D
u,
L
.
W
a
ng,
a
nd
F
.
S
un,
“
A
tr
a
f
f
ic
-
w
e
a
th
e
r
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
f
o
r
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
f
or
r
oa
d
ne
twor
ks
unde
r
ba
d
w
e
a
th
e
r
,”
E
ngi
n
e
e
r
in
g
A
ppl
ic
at
io
ns
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
,
vol
.
137,
2
024,
doi
:
10.1016/j
.
e
nga
ppa
i.
2024.109125.
[
2]
D
.
O
la
di
me
ji
,
K
.
G
upt
a
,
N
.
A
.
K
os
e
,
K
.
G
undoga
n,
L
.
G
e
,
a
nd
F
.
L
ia
ng,
“
S
ma
r
t
tr
a
ns
por
ta
ti
on:
a
n
ove
r
vi
e
w
of
te
c
hnol
ogi
e
s
a
nd
a
ppl
ic
a
ti
ons
,”
Se
n
s
or
s
, vol
. 23, no. 8, pp. 1
–
32, 2023, doi:
10.3390/s
23083880.
[
3]
K
.
M
.
A
lm
a
ta
r
,
“
S
ma
r
t
tr
a
ns
por
ta
ti
on
pl
a
nni
ng
a
nd
it
s
c
ha
ll
e
nge
s
in
th
e
K
in
gdom
of
S
a
udi
A
r
a
bi
a
,”
Sus
ta
in
abl
e
F
ut
ur
e
s
,
vol
.
8,
no. De
c
e
mbe
r
2023, 2024, doi:
10.1016/j
.s
f
tr
.2024.100238.
[
4]
S
.
K
ha
n,
S
.
K
ha
n,
A
.
S
ul
a
im
a
n,
M
.
S
.
A
l
R
e
s
ha
n,
H
.
A
ls
ha
hr
a
ni
,
a
nd
A
.
S
ha
ik
h,
“
D
e
e
p
ne
ur
a
l
ne
twor
k
a
nd
tr
us
t
ma
na
g
e
me
nt
a
ppr
oa
c
h
to
s
e
c
ur
e
s
ma
r
t
tr
a
ns
por
ta
ti
on
da
ta
in
s
us
t
a
in
a
bl
e
s
ma
r
t
c
it
ie
s
,”
I
C
T
E
x
pr
e
s
s
,
vol
.
10,
no.
5,
pp.
1059
–
1065,
2
024,
doi
:
10.1016/j
.i
c
te
.2024.08.006.
[
5]
K
.
J
a
li
l,
Y
.
X
i
a
,
Q
.
C
h
e
n
,
M
.
N
.
Z
a
h
id
,
T
.
M
a
nz
oor
,
a
nd
J
.
Z
h
a
o
,
“
I
nt
e
gr
a
ti
v
e
r
e
vi
e
w
o
f
da
ta
s
c
i
e
n
c
e
s
f
or
dr
iv
in
g
s
ma
r
t
mo
bi
li
ty
i
n
i
nt
e
l
li
ge
nt
t
r
a
n
s
p
or
t
a
t
io
n
s
y
s
t
e
m
s
,
”
C
o
m
p
ut
e
r
s
an
d E
le
c
tr
i
c
al
E
ng
in
e
e
r
i
ng
,
vo
l.
1
19
,
20
24
,
doi
:
10
.1
01
6/
j.
c
o
mp
e
l
e
c
e
n
g.
20
24
.1
09
62
4
.
[
6]
E
.
D
il
e
k
a
nd
M
.
D
e
ne
r
,
“
C
omput
e
r
vi
s
io
n
a
ppl
ic
a
ti
o
ns
in
in
te
ll
ig
e
nt
tr
a
ns
por
ta
ti
on
s
ys
te
ms
:
a
s
ur
ve
y,”
Se
ns
o
r
s
,
vol
.
23,
no.
6,
2023, doi:
10.3390/s
23062938.
[
7]
H
.
C
hi
,
Y
.
L
u,
C
.
X
ie
,
W
.
K
e
,
a
nd
B
.
C
he
n,
“
S
pa
ti
o
-
te
mpor
a
l
a
tt
e
nt
io
n
ba
s
e
d
c
ol
la
bor
a
ti
ve
lo
c
a
l
–
gl
oba
l
le
a
r
ni
ng
f
or
tr
a
f
f
ic
f
l
ow
pr
e
di
c
ti
on,”
En
gi
ne
e
r
in
g A
ppl
ic
at
io
ns
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
, vol
. 139, 2025, doi:
10.1016/j
.e
nga
ppa
i.
2024.109575.
[
8]
H
.
X
in
g,
A
.
C
he
n,
a
nd
X
.
Z
ha
ng,
“
R
L
-
G
C
N
:
tr
a
f
f
ic
f
lo
w
pr
e
d
ic
ti
on
ba
s
e
d
on
gr
a
ph
c
onvolut
io
n
a
nd
r
e
in
f
or
c
e
me
nt
le
a
r
ni
ng
f
or
s
ma
r
t
c
it
ie
s
,”
D
is
pl
a
y
s
,
vol
. 80, 2023, doi:
10.1016/j
.di
s
pl
a
.202
3.102513.
[
9]
C
.
M
a
,
Y
.
H
u,
a
nd
X
.
X
u,
“
H
ybr
id
de
e
p
le
a
r
ni
ng
mode
l
w
it
h
V
M
D
-
B
iL
S
T
M
-
G
R
U
ne
twor
ks
f
or
s
hor
t
-
te
r
m
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on,”
D
at
a Sc
ie
nc
e
and M
anage
m
e
nt
, N
ov. 2024, doi:
10
.1016/j
.ds
m.2024.10.004.
[
10]
J.
Z
he
ng,
M
.
W
a
ng,
a
nd
M
.
H
u
a
ng,
“
E
xpl
or
in
g
th
e
r
e
la
ti
ons
hi
p
be
twe
e
n
da
t
a
s
a
mpl
e
s
iz
e
a
nd
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
a
c
c
ur
a
c
y,”
T
r
ans
por
ta
ti
on E
ngi
ne
e
r
in
g
, vol
. 18, 2024, doi:
10.1016/j
.t
r
e
ng.2024.100279.
[
11]
M
.
A
bde
l
-
A
ty
,
Z
.
W
a
ng,
O
.
Z
he
ng,
a
nd
A
.
A
bde
lr
a
ouf
,
“
A
d
va
nc
e
s
a
nd
a
ppl
ic
a
ti
ons
of
c
omput
e
r
vi
s
io
n
te
c
hni
que
s
in
ve
h
ic
le
tr
a
je
c
to
r
y
ge
ne
r
a
ti
on
a
nd
s
ur
r
oga
te
tr
a
f
f
ic
s
a
f
e
ty
in
di
c
a
to
r
s
,”
A
c
c
id
e
nt
A
nal
y
s
is
and
P
r
e
v
e
nt
io
n
,
vol
.
191,
2023,
doi
:
10.1016/j
.a
a
p.2023.107191.
[
12]
H
.
G
a
o
e
t
al
.
,
“
A
hybr
id
de
e
p
le
a
r
ni
ng
mode
l
f
or
u
r
ba
n
e
xpr
e
s
s
w
a
y
la
ne
-
le
v
e
l
mi
xe
d
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on,”
E
ngi
ne
e
r
in
g
A
ppl
ic
at
io
ns
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
, vol
. 133, 2024, doi:
10.1
016/
j.
e
nga
ppa
i.
2024.108242.
[
13]
R
.
Z
ha
o,
S
.
H
.
T
a
ng,
J
.
S
he
n,
E
.
E
.
B
in
S
upe
ni
,
a
nd
S
.
A
.
R
a
hi
m,
“
E
nha
nc
in
g
a
ut
onomous
dr
iv
in
g
s
a
f
e
ty
:
a
r
obu
s
t
tr
a
f
f
ic
s
ig
n
de
te
c
ti
on a
nd r
e
c
ogni
ti
on mode
l
T
S
D
-
Y
O
L
O
,”
Si
gnal
P
r
oc
e
s
s
i
ng
, vol
. 225, 2024, doi:
10.1016/j
.s
ig
pr
o.2024.109619.
[
14]
S
.
X
u,
M
.
Z
ha
ng,
J
.
C
he
n,
a
nd
Y
.
Z
hong,
“
Y
O
L
O
-
H
ype
r
V
is
io
n:
a
vi
s
io
n
tr
a
ns
f
or
me
r
ba
c
kbone
-
ba
s
e
d
e
nha
nc
e
me
nt
of
Y
O
L
O
v5
f
or
de
te
c
ti
on of
dyna
mi
c
t
r
a
f
f
ic
i
nf
o
r
ma
ti
on,”
E
gy
pt
ia
n I
nf
or
m
a
ti
c
s
J
our
nal
, vol
. 27, 2024, doi:
10.1016/j
.e
ij
.2024.100523.
[
15]
H
.
A
.
S
a
pu
tr
i,
M
.
A
vr
i
ll
io
,
L
.
C
hr
is
to
f
e
r
,
V
.
S
im
a
nj
a
ya
,
a
nd
I
.
N
.
A
la
m,
“
I
mpl
e
me
nt
a
ti
on
of
Y
O
L
O
v7
a
lg
o
r
it
hm
in
e
s
ti
ma
ti
ng
tr
a
f
f
ic
f
lo
w
i
n M
a
la
ng,”
P
r
oc
e
di
a C
om
put
e
r
S
c
ie
nc
e
, vol
. 245,
pp. 117
–
126, 2024, doi:
10.1016/j
.pr
oc
s
.2024.10.235.
[
16]
J
.
T
a
ng,
C
.
Y
e
,
X
.
Z
ho
u,
a
nd
L
.
X
u,
“
Y
O
L
O
-
f
us
io
n
a
nd
in
te
r
ne
t
of
th
in
gs
:
a
dva
nc
in
g
obj
e
c
t
de
te
c
ti
on
in
s
ma
r
t
tr
a
ns
por
ta
ti
on,”
A
le
x
andr
ia
E
ngi
ne
e
r
in
g J
our
nal
, vol
. 107, pp. 1
–
12, 2024, doi:
10.1016/j
.a
e
j.
2024.09.012.
[
17]
R
.
R
ona
r
iv
,
R
.
A
nt
oni
o,
S
.
F
.
J
or
ge
ns
e
n,
S
.
A
c
hma
d,
a
nd
R
.
S
ut
oyo,
“
O
bj
e
c
t
de
te
c
ti
on
a
lg
or
it
hms
f
or
c
a
r
t
r
a
c
ki
ng
w
it
h
e
uc
li
de
a
n
di
s
ta
nc
e
t
r
a
c
ki
ng a
nd Y
O
L
O
,”
P
r
o
c
e
di
a C
om
put
e
r
Sc
ie
nc
e
, vol
. 245, pp. 627
–
636, 2024, doi:
10.1016/j
.pr
oc
s
.2024.10.289.
[
18]
X
.
Z
ha
i,
Z
.
H
ua
ng,
T
.
L
i,
H
.
L
iu
,
a
nd
S
.
W
a
ng,
“
Y
O
L
O
-
d
r
one
:
a
n
opt
im
iz
e
d
Y
O
L
O
v8
ne
twor
k
f
or
t
in
y
U
A
V
obj
e
c
t
de
te
c
ti
on,”
E
le
c
tr
oni
c
s
, vol
. 12, no. 17, 2023, doi
:
10.3390/e
le
c
tr
oni
c
s
1217
3664.
[
19]
I
.
S
ha
mt
a
,
F
.
D
e
mi
r
,
a
nd
B
.
E
.
D
e
mi
r
,
“
P
r
e
di
c
ti
ve
f
a
ul
t
d
e
te
c
ti
on
a
nd
r
e
s
ol
ut
io
n
u
s
in
g
Y
O
L
O
v8
s
e
gm
e
nt
a
ti
on
mode
l
:
a
c
ompr
e
he
ns
iv
e
s
tu
dy
on
hot
s
pot
f
a
ul
ts
a
nd
g
e
ne
r
a
li
z
a
ti
on
c
ha
ll
e
nge
s
in
c
omput
e
r
vi
s
io
n,”
A
in
Sham
s
E
ngi
ne
e
r
in
g
J
our
nal
,
vol
. 15, no. 12, 2024, doi
:
10.1016/j
.a
s
e
j.
2024.103148.
[
20]
K
.
W
a
ng,
C
.
M
a
,
Y
.
Q
ia
o,
X
.
L
u,
W
.
H
a
o,
a
nd
S
.
D
ong,
“
A
h
ybr
id
de
e
p
le
a
r
ni
ng
mod
e
l
w
it
h
1D
C
N
N
-
L
S
T
M
-
a
tt
e
nt
io
n
ne
tw
or
ks
f
or
s
hor
t
-
te
r
m
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on,”
P
hy
s
ic
a
A
:
St
at
is
ti
c
al
M
e
c
hani
c
s
and
it
s
A
ppl
ic
at
io
ns
,
vol
.
583,
2
021,
doi
:
10.1016/j
.phys
a
.2021.126293.
[
21]
J
.
L
u,
“
A
n
e
f
f
ic
ie
nt
a
nd
in
te
ll
ig
e
nt
t
r
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
me
th
o
d
ba
s
e
d
on
L
S
T
M
a
nd
va
r
ia
ti
ona
l
moda
l
de
c
ompos
it
io
n,”
M
e
as
ur
e
m
e
nt
:
Se
n
s
or
s
, vol
. 28, 2023, doi:
10.1016/j
.me
a
s
e
n.20
23.100843.
[
22]
B
.
N
a
h
e
li
y
a
,
P
.
R
e
dh
u,
a
nd
K
.
K
um
a
r
,
“
M
F
O
A
-
Bi
-
L
S
T
M
:
a
n
op
ti
mi
z
e
d
bi
di
r
e
c
ti
on
a
l
l
on
g
s
h
or
t
-
t
e
r
m
m
e
m
or
y
mo
d
e
l
f
or
s
hor
t
-
te
r
m
tr
a
f
f
i
c
f
lo
w
pr
e
d
i
c
ti
on
,
”
P
hy
s
ic
a
A
:
St
at
i
s
ti
c
a
l
M
e
c
h
an
i
c
s
a
nd
i
t
s
A
ppl
i
c
at
io
n
s
,
v
ol
.
63
4,
2
02
4,
d
oi
:
1
0.
10
16
/j
.p
hy
s
a
.
20
23
.1
29
44
8.
[
23]
J
.
D
.
W
a
ng
a
nd
C
.
O
.
N
.
S
us
a
nt
o,
“
T
r
a
f
f
ic
f
lo
w
pr
e
di
c
ti
o
n
w
it
h
he
te
r
oge
nous
da
ta
us
in
g
a
hybr
id
C
N
N
-
L
S
T
M
mod
e
l,
”
C
om
put
e
r
s
, M
at
e
r
ia
ls
and C
ont
in
ua
, vol
. 76, no. 3, pp. 3097
–
3
112, 2023, doi:
10.32604/cmc
.2023.040914.
[
24]
A
.
R
.
S
a
tt
a
r
z
a
de
h,
R
.
J
.
K
ut
a
di
na
ta
,
P
.
N
.
P
a
th
ir
a
na
,
a
nd
V
.
T
.
H
uynh,
“
A
nov
e
l
hybr
id
de
e
p
le
a
r
ni
ng
mode
l
w
it
h
A
R
I
M
A
c
onv
-
L
S
T
M
ne
twor
ks
a
nd
s
huf
f
le
a
tt
e
nt
io
n
la
y
e
r
f
or
s
hor
t
-
te
r
m
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on,”
T
r
ans
por
tme
t
r
ic
a
A
:
T
r
ans
po
r
t
Sc
ie
nc
e
,
vol
. 21, no. 1, pp. 388
–
410, 2025, doi:
10.1080/23249935.
2023.2236724.
[
25]
Y
.
L
uo,
J
.
Z
he
ng,
X
.
W
a
ng,
Y
.
T
a
o,
a
nd
X
.
J
ia
ng,
“
G
T
-
L
S
T
M
:
a
s
pa
ti
o
-
te
mpor
a
l
e
ns
e
mbl
e
ne
two
r
k
f
or
tr
a
f
f
ic
f
lo
w
p
r
e
di
c
ti
o
n,”
N
e
ur
al
N
e
t
w
or
k
s
, vol
. 171, pp. 251
–
262, 2024, doi:
10.1016/j
.ne
une
t.
2023.12.016.
[
26]
K
.
Z
ha
o,
D
.
G
uo,
M
.
S
un,
C
.
Z
ha
o,
a
nd
H
.
S
hua
i,
“
S
hor
t
-
te
r
m
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
ba
s
e
d
on
V
M
D
a
nd
I
D
B
O
-
L
S
T
M
,
”
I
E
E
E
A
c
c
e
s
s
, vol
. 11, pp
. 97072
–
97088, 2023, doi:
10.1109/AC
C
E
S
S
.2023.3312711.
[
27]
B
ha
r
ti
,
P
.
R
e
dhu,
a
nd
K
.
K
uma
r
,
“
S
hor
t
-
te
r
m
tr
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
ba
s
e
d
on
opt
im
iz
e
d
d
e
e
p
le
a
r
ni
ng
ne
ur
a
l
ne
twor
k:
P
S
O
-
Bi
-
L
S
T
M
,”
P
hy
s
ic
a A
:
St
at
is
ti
c
al
M
e
c
hani
c
s
and it
s
A
ppl
ic
at
io
ns
,
vol
. 625,
2023, doi:
10.1016/j
.phys
a
.2023.129001.
[
28]
C
.
C
ha
our
a
,
H
.
L
a
z
a
r
,
a
nd
Z
.
J
a
r
ir
,
“
T
r
a
f
f
ic
f
lo
w
pr
e
di
c
ti
on
a
t
in
te
r
s
e
c
ti
ons
:
e
nha
n
c
in
g
w
it
h
a
H
ybr
id
L
S
T
M
-
P
S
O
a
ppr
oa
c
h,”
I
nt
e
r
nat
io
nal
J
our
nal
of
A
dv
anc
e
d
C
om
put
e
r
Sc
ie
n
c
e
and
A
ppl
ic
at
io
ns
,
vol
.
15,
no.
5,
pp.
494
–
501,
2024,
doi
:
10.14569/I
J
A
C
S
A
.2024.01
50549.
[
29]
S
.
S
uya
nt
o,
A
.
A
.
A
r
iy
a
nt
o,
a
nd
A
.
F
.
A
r
iy
a
nt
o,
“
K
omodo
M
li
pi
r
a
lg
or
it
hm,”
A
ppl
ie
d
Sof
t
C
om
put
in
g
,
vol
.
114,
2
022,
doi
:
10.1016/j
.a
s
oc
.2021.108043.
[
30]
G
.
K
.
J
a
ti
,
G
.
K
uw
a
nt
o,
T
.
H
a
s
hmi
,
a
nd
H
.
W
id
ja
ja
,
“
D
i
s
c
r
e
te
komodo
a
lg
or
it
hm
f
or
tr
a
ve
li
n
g
s
a
le
s
ma
n
pr
obl
e
m,”
A
ppl
ie
d
Sof
t
C
om
put
in
g
, vol
. 139, M
a
y 2023, doi:
10.1016/j
.a
s
oc
.2023.1102
19.
[
31]
D
.
R
a
im
ondo
e
t
al
.
,
“
D
e
te
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
of
hys
te
r
os
c
opi
c
im
a
ge
s
us
in
g
d
e
e
p
le
a
r
ni
ng,”
C
anc
e
r
s
,
vol
.
16,
no.
7,
M
a
r
.
2024, doi:
10.3390/ca
nc
e
r
s
1
6071315.
[
32]
M
.
P
e
i,
N
.
L
iu
,
B
.
Z
ha
o,
a
nd
H
.
S
un,
“
S
e
lf
-
s
upe
r
vi
s
e
d
le
a
r
ni
n
g
f
or
in
dus
tr
ia
l
im
a
ge
a
noma
ly
de
te
c
ti
on
by
s
im
ul
a
ti
ng
a
noma
lo
us
s
a
mpl
e
s
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
C
om
put
at
io
nal
I
nt
e
ll
ig
e
nc
e
Sy
s
te
m
s
, vol
. 16, no. 1, 2023, doi
:
10.1007/s
44196
-
023
-
00328
-
0.
Evaluation Warning : The document was created with Spire.PDF for Python.