IAES Inter
nat
iona
l
Journ
al
of A
r
tifici
al I
nt
el
li
gence
(I
J
-
AI
)
Vo
l.
8
, No
.
2
,
J
un
e
201
9
, pp.
107
~
119
IS
S
N: 22
52
-
8938
,
DOI: 10
.11
591/ijai.
v
8
.i
2
.pp
107
-
119
107
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/o
nline/i
nd
ex
.ph
p/I
J
AI
Adapti
ve re
al tim
e traffi
c predict
ion
usin
g deep n
eural
network
s
Parinith
R
I
ye
r
1
,
Shru
th
ee
sh
R
aman I
yer
2
,
Ragha
vendra
n Rames
h
3
, A
na
la
MR
4
,
K.N. Subr
am
anya
5
1,2,3,4
Depa
rtment
of
Com
pute
r
Sci
enc
e
and
Engi
n
e
eri
ng,
R.
V.
Col
l
ege
of
Eng
ine
e
ri
ng,
Indi
a
5
Princi
pal,
R
.
V.
Coll
ege of
Enginee
ring
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
5
, 201
9
Re
vised
Ma
r
6
,
201
9
Accepte
d
Ma
y
9
, 2
01
9
The
eve
r
-
inc
r
eas
ing
sale
of
vehi
cles
and
the
stea
d
y
inc
r
ea
se
in
popula
ti
o
n
density
in
m
et
r
opoli
ta
n
ci
t
ie
s
have
ra
ised
m
an
y
growing
con
ce
rns,
m
ost
important
l
y
com
m
ute
ti
m
e,
a
ir
a
nd
noise
poll
ut
i
on
le
vel
s
.
Tra
ff
i
c
conge
st
ion
ca
n
be
al
l
eviated
b
y
opt
ing
adapt
ive
tr
aff
i
c
li
gh
t
s
y
stems
,
instea
d
of
fixe
d
-
ti
m
e
tra
ffi
c
signal
s.
In
thi
s
paper,
a
s
y
stem
is
proposed
which
ca
n
detec
t
,
cl
assif
y
and
cou
nt
vehi
cles pa
ss
ing
through
an
y
t
raf
fic
jun
ct
ion
u
sing a
single
ca
m
era
(
as
opposed
to
m
ult
i
-
sensor
appr
oa
che
s).
The
de
t
e
ction
and
cl
assifi
ca
t
ion
a
re
done
using
SSD
Neura
l
Network
obje
ct
det
e
ction
al
gorit
hm
.
The
count
of
e
ac
h
c
l
ass
(2
-
whee
le
rs,
ca
rs,
tru
cks,
bu
ses
et
c.
)
is
used
to
pre
dic
t
the
signal
gre
en
-
ti
m
e
for
the
ne
xt
c
y
cle.
Th
e
m
odel
self
-
adj
usts
eve
r
y
c
y
c
le
b
y
ut
il
i
zi
n
g
weight
ed
m
oving
ave
r
age
s.
Thi
s
s
y
stem
works
well
bec
ause
the
cha
nge
i
n
the
density
of
tra
ffi
c
on
an
y
gi
ven
roa
d
is
gra
dual,
spann
in
g
m
ult
iple
tr
aff
i
c
stops throu
ghou
t
th
e
d
a
y
.
Ke
yw
or
d
s
:
In
te
ll
igent t
ransportat
ion
syst
e
m
s
Neural
netw
ork
Object
detect
io
n/cla
ssific
at
ion
and co
unt
Traffic
sig
nal t
i
m
e
Weig
hted
m
oving
a
ve
rag
e
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Sh
r
ut
heesh Ra
m
an
Iyer,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd
E
ng
i
ne
erin
g,
R.V.
C
ollege
of E
ng
i
neer
i
ng
,
Be
ng
al
uru,
Ka
rn
at
a
ka 560
059,
I
nd
ia
.
Em
a
il
:
sh
ru
the
esh.
ir
@
gm
ail.co
m
1.
INTROD
U
CTION
The
tra
ff
ic
c
on
gestio
n
pr
ob
le
m
,
especial
ly
i
n
urban
areas
,
has
been
w
orse
ning
ov
e
r
the
la
st
20
ye
ars.
Ba
sed
on
the
s
urvey
by
Eco
nom
ic
Ti
m
es
[1
]
,
it
has
been
obser
ve
d
that
as
of
20
16,
the
num
ber
of
veh
ic
le
s
in
Be
ng
al
uru,
In
di
a (6
6.6
5
la
khs
)
has
rise
n
to 6.7
tim
es o
f
what
it
was
in 1
99
6
(9.93 lak
hs),
and
2.4 ti
m
es t
hat o
f
2006
(28.0
2
la
kh
s
).
Sim
i
la
r
s
ta
ti
sti
cs
are
found
i
n
m
ulti
pl
e
othe
r
surveys
on
India
as
w
el
l
[2
-
3].
The
r
esult
of
this
ra
pid
gro
w
th
has
m
any
adv
erse
ef
fects,
i
nclu
ding
deteri
or
at
in
g
healt
h
conditi
ons,
a
ve
rag
e
c
omm
ute
tim
e
of
ov
e
r
th
ree
hours
a
day
a
nd
poor
ai
r
qual
it
y.
This
sh
a
rp
gro
wth
a
nd
i
nadequ
at
e
m
eans
to
ha
nd
le
it
have
le
ft
m
any
ro
a
ds
in
a
sta
te
of
dis
rep
ai
r
.
T
hese
issues
ide
ntify
the
s
hortcom
ing
s
of
tra
ff
ic
si
gn
al
s
us
e
d
on
m
os
t
ro
a
ds
t
od
ay
,
in
handlin
g
tra
ff
ic
.
M
os
t
ci
ti
es
ha
ve
fixe
d
traff
ic
li
ght
c
ontr
ol
syst
em
s
in
place
w
hich
don’
t
accom
m
od
at
e the
dynam
ic
n
at
ur
e
of tra
ff
ic
fl
ow.
The
fi
xed
-
ti
m
e
traff
ic
sig
nals
are
inef
fici
en
t,
as
the
durat
ion
of
ti
m
e
for
w
hich
sig
nal
is
op
en
is
us
ua
ll
y
insu
f
fi
ci
ent
f
or
busy
ro
a
ds,
an
d
e
xcessive
f
or
le
ss
busy
on
es
(w
it
h
fe
wer
veh
ic
le
s)
.
T
he
curre
nt
syst
e
m
has
le
ft
m
uch
to
be
d
esi
red
.
A
va
ria
nt
of
t
he
fixe
d
syst
e
m
is
us
e
d
s
om
et
i
m
es.
It
involves
assi
gn
i
ng
diff
e
re
nt
green
-
tim
es
through
ou
t
the
day
to
accom
m
od
at
e
ru
s
h
hours,
but
they
we
re
s
ti
ll
“fixed
”.
A
viable
so
luti
on wo
uld be to
check
th
e change in
tra
ff
ic
d
e
ns
it
y fr
e
qu
e
ntly
(
sa
y, ever
y fe
w
m
inu
te
s),
d
et
e
rm
ine
wh
ic
h
ro
a
d
is
cl
ogge
d
a
nd
deci
de,
i
n
re
al
-
tim
e,
the
tim
e
req
uire
d
to
cl
ear
t
he
tra
f
fic
on
e
ver
y
r
oa
d.
The
pro
ble
m
this
pap
e
r
ai
m
s
to
so
lve
is
the
e
xten
ded
wait
in
g
tim
e
at
traffi
c
intersect
ions,
it
do
es
s
o
by
pr
op
os
in
g
a
nov
el
al
gorithm
to
m
ake
t
he
al
lott
e
d
gr
ee
n
tim
es
adap
ti
ve
with
t
he
tra
ff
ic
volu
m
e
at
that
poin
t
in
ti
m
e.
The
go
al
is
to cr
eat
e a
syst
e
m
that i
m
plem
ents the idea
s put fo
rth
a
bo
ve.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
IJ
-
AI
V
ol.
8
,
No.
2, June
201
9
:
107
–
119
108
The
ai
m
of
t
his
pa
per
is
to
present
a
no
vel
appr
oach
to
t
he
dynam
ic
traf
fic
sig
n
al
syst
em
based
on
intuit
ive
ob
j
ec
t
detect
ion
te
chn
i
qu
e
s
usi
ng
Co
nvol
utional
Neural
Net
works
co
uple
d
wit
h
Sin
gle
Shot
Detect
or
s
a
nd
pr
ovide
a
pr
edict
ive
m
od
e
l
to
decide
up
on
the
green
-
tim
e
of
the
t
raffic
si
gn
al
,
for
the
resp
ect
ive
la
ne
s.
It
can
al
so
r
ect
if
y
it
sel
f
qu
ic
kly
with
the
ongoin
g
tre
nd
of
traff
ic
,
as
fas
t
as
on
e
cy
cl
e
of
the
sign
al
(
one
cy
cl
e
inv
ol
ves
al
l
fo
ur
r
oa
ds
ge
tt
ing
green
ti
m
e
on
ce
).
T
he
de
ns
it
y
of
tra
ff
i
c
in
each
la
ne
is
us
ed
to
con
t
ro
l
traf
f
ic
sign
al
s
by
predict
in
g
the
am
ou
nt
of
tim
e
the
sign
al
nee
ds
to
ope
n.
T
hus,
the
durati
on
of
"gr
ee
n
si
gn
al
"
(signal
is
op
e
n)
is
dicta
te
d
by
the
c
ontrib
ution
of
the
sa
id
la
ne
t
o
the
traff
ic
bu
il
d
up.
T
he
m
od
el
p
r
opos
e
d here
can
pote
ntial
ly
b
e d
e
pl
oyed o
n
a
sm
all, por
ta
ble sys
tem
li
ke
the
Nvidia
Jetso
n Kit
.
2.
RELATE
D
W
ORK
Pr
e
vious
w
ork
in
In
te
ll
igent
Transp
or
ta
ti
on
Syst
e
m
s
(I
TS)
hav
e
been
ve
ry
prom
isi
ng
with
a
few
caveat
s
w
hich
can
pote
ntial
ly
cause
pr
ob
le
m
s
in
the
im
ple
m
entat
ion
phas
e
in
certai
n
loc
at
ion
s.
Our
pr
e
viou
s
work
[
4]
pr
e
se
nts
ve
hicle
det
ect
ion
a
n
d
c
ou
nting
us
i
ng
im
age
pr
ocessin
g
te
chn
iq
ues
.
T
his
pa
per
e
xpl
or
es
the
dee
p
le
a
rn
i
ng
te
ch
niques
to
achie
ve
the
sam
e
ta
sk
.
Ma
j
ori
ty
of
the
w
ork
ta
c
kling
th
is
pro
blem
has
bee
n
sens
or
ori
ente
d
[5
]
.
T
her
e
ha
ve
bee
n
pro
po
sit
ion
s
involvi
ng
placi
ng
se
nsors
on
ei
th
e
r
side
of
the
ro
a
d,
or
in
oth
e
r
places
al
m
os
t
a
t
the
gr
ound
le
vel
to
de
te
rm
ine
the
num
ber
of
ve
hicle
s
m
ov
in
g
or
to
ge
ner
al
ly
est
i
m
at
e
the
de
ns
it
y
of
veh
ic
le
s
a
nd
he
nce
deci
de
ho
w
m
uch
green
-
tim
e
to
al
locat
e.
The
m
os
t
prom
inent
of
the
se
are
the
loop
detect
or
s
wh
ic
h
are
excell
ent
in
cal
culat
ing
tra
ff
ic
volum
e.
Yet,
they
do
no
t
perform
well
in
est
i
m
ating
the
traff
ic
de
ns
it
y.
In
[
6],
the
use
of
RF
ID
se
nsors
has
been
pro
po
se
d
to
de
te
ct
veh
ic
le
s
m
ov
ing
thr
ough
so
m
e
portio
n
of
the
ro
a
d.
A
ny
ap
proac
h
in
volvin
g
de
p
l
oying
se
ns
ors
a
ppr
ox
i
m
at
ely
at
the
gro
un
d
le
vel
will
be
i
nf
easi
ble
as
not
al
l
places
en
able
su
c
h
de
plo
ym
ent
of
se
nsors
.
T
he
de
ve
lop
in
g
c
ount
ries
ha
ve
ro
a
ds
w
hich so
m
et
i
m
es are indist
inguisha
ble
f
r
om
the foo
t
pa
th as sh
own
i
n
Fi
gure
1.
Figure
1
.
A Fr
equ
e
ntly
co
m
m
uted
ro
a
d
in
Be
ng
al
uru
In
[
7],
the
pro
po
s
ed
syst
em
m
akes
us
e
of
Ultraso
nic
sen
so
r
a
nd
cl
ai
m
s
to
place
it
"on
top
of
the
ro
a
d"
but
the
m
axi
m
u
m
ran
ge
of
the
se
ns
or
is
just
4
m
et
ers,
wh
ic
h
is
not
feasible
.
I
n
[
8]
,
the
syst
em
us
es
I
R
sens
or
s
an
d
A
rduin
o
a
nd
pla
ns
to
m
ou
nt
t
he
se
nsors
on
ei
ther
si
de
of
the
r
oa
d
on
the
po
le
s.
De
pl
oying
sens
or
s
in
s
uc
h
an
en
vir
on
m
ent
is
no
t
po
ss
ible
and
su
c
h
appr
oach
es
ar
e
no
t
best
su
it
ed
to
so
lve
the
issue.
Re
centl
y,
appr
oach
e
s
ha
ve
s
hifted
t
o
cam
e
ra
-
base
d
so
l
ution
s
.
T
hese
m
e
thods
at
te
m
pt
to
fin
d
the
ve
hicle
densi
ty
on
eac
h
la
ne
a
nd
c
onseq
uen
tl
y
deter
m
ine
the
traf
fic
sign
al
ti
m
e.
The
ea
rly
ap
proach
e
s
dealt
w
it
h
bl
ob
extracti
on
m
eth
ods
su
c
h
a
s
in
[
9].
Alth
ough
they
we
re
m
or
e
ve
rsati
le
than
the
se
nsor
m
eth
ods,
their
acc
ur
ac
y
le
ft
a
lot
to
be
desire
d.
I
n
re
cent
tim
es,
research
base
d
on
detect
ion
,
tra
c
king
an
d
cl
assi
ficat
ion
of
vehi
cl
es
from
vid
eo
im
ages
has
be
com
e
feasible.
Thanks
to
br
eakth
rou
gh
s
i
n
com
pu
te
r
vi
sion
te
ch
nolo
gy
and
incr
em
ental
a
dv
a
ncem
ents
i
n
com
pu
te
powe
r.
T
he
re
su
lt
s
are
f
oun
d
to
be
ver
y
prom
is
ing
as
well
.
Conv
olu
ti
onal
Neural
Netw
orks
(CN
N
)
pro
vid
e
nea
r
hum
an
le
vel
accur
acy
in
the
field
of
obj
ect
det
ect
io
n.
In
[
10]
,
a
su
cc
essfu
l
cam
era
base
d
co
un
ti
ng
m
et
ho
d
was
im
ple
m
ented
by
the
m
eans
of
m
ulti
-
ob
j
ect
tr
ackin
g.
Each
ve
hicle
on
the
r
oa
d
is
t
r
acked
un
ti
l
it
goes
out
of
fr
a
m
e.
This
he
nc
e
give
s
a
n
acc
ur
at
e
c
ount.
H
ow
e
ve
r,
this
process
i
s
extrem
el
y
c
om
pu
ta
ti
on
al
ly
exp
e
ns
ive
si
nce
ob
j
ect
trackin
g
in
vo
l
ve
s
redu
nd
a
ncy.
This
be
com
es a problem
in
syst
e
m
s that
hav
e
con
strai
ned h
a
r
dware r
e
source
s.
In
[
11]
,
a
deep
neural
netw
ork
was
us
ed
on
vid
e
os
with
a
r
egr
es
sio
n
ap
proach
t
o
colle
ct
ively
coun
t
the
num
ber
of
veh
ic
le
s
t
o
be
counted
,
t
hu
s
no
t
al
lo
wi
ng
r
oo
m
for
cl
assi
ficat
ion
.
I
n
m
ajo
rity
of
the
ca
m
era
-
base
d
ap
proac
hes
s
uch
a
s
th
e
on
e
s
disc
us
s
ed
ab
ove,
at
te
m
pts
are
m
ade
to
cal
culat
e
th
e
red
ti
m
e
in
a
giv
e
n
traff
ic
ju
nctio
n
fo
r
a
la
ne
.
T
his
syst
e
m
is
no
t
feasible
f
or
a
f
ew
reas
ons.
F
r
om
the
view
of
the
sign
al
,
vehi
cl
es
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
AI
IS
S
N:
22
52
-
8938
Ad
ap
ti
ve re
al t
ime
traff
ic
p
r
e
dicti
on
us
i
ng dee
p neural
net
works
(
P
ar
init
h
R
I
yer
)
109
far
off
i
n
the
r
o
ad
,
wait
in
g
in
the
sign
al
,
wil
l
no
t
be
ide
ntif
ie
d
accuratel
y.
As
the
distanc
e
increases,
th
e
siz
e
of
th
e
ob
j
ect
s
captu
red
by
th
e
ca
m
era
is
bo
th
sm
al
l
and
no
t
e
ntirel
y
visi
ble
(as
pa
rts
m
ay
be
cut
off
by
cars
in
fro
nt),
this,
the
refo
re,
exposes
the
c
onstrai
nts
of
s
uc
h
m
od
el
s.
T
hus,
this
pap
e
r
presents
a
n
in
novativ
e
appr
oach
t
hat
has
borro
wed
so
m
e
el
e
m
ents
from
the
afo
r
e
m
entione
d
re
search
t
o
cre
a
te
a
pr
e
dicti
ve
and
adap
ti
ve
tra
ff
i
c
con
tr
ol
sig
nal.
I
n
this
appr
oach,
base
d
on
the
nu
m
ber
of
ve
hicle
s
passing
thr
o
ugh
the
jun
ct
io
n,
t
he
gree
n
tim
e
for
the
par
ti
cu
la
r
la
ne
is
est
im
at
ed
fo
r
t
he
nex
t
cy
cl
e.
I
n
this
way,
it
is
bo
t
h
sel
f
-
c
orrecti
ng
and
a
reali
sti
c
im
ple
m
entat
ion
.
The
basic
pri
nc
iple
in
any
so
luti
on
for
thi
s
sp
eci
fic
pro
blem
is
c
le
ar,
the
nu
m
ber
of
veh
ic
l
es
is
the
m
os
t
i
m
po
rtant
pa
ram
et
e
r
one
s
hould
obta
in
to
proce
ed
f
ur
t
her
i
n
any
way
an
d
to
ultim
at
ely
al
l
ocate
appr
opriat
e
gr
een
-
ti
m
e.
It
is
cl
ear
th
at
us
i
ng
se
ns
ors
is
no
t
feasi
ble,
s
o
m
or
e
strai
ght
-
for
ward,
hu
m
an
-
li
ke
countin
g
a
ppr
oa
ches
cam
e
ab
ou
t.
I
n
[1
2],
de
te
rm
ining
t
he
de
ns
it
y
of
ob
je
ct
s
in
a
c
row
ded
sce
ne
us
in
g
the
Hydra
-
CN
N
a
nd
t
he
Co
unti
ng
-
CN
N
is
pr
opos
e
d
w
hic
h
is
revoluti
on
ary
in
te
rm
s
of
m
ass
public
place
su
r
veill
ance
sy
stem
s
and
j
am
-
pac
ke
d
r
oa
ds
.
This
sti
ll
fall
s
short
of
gro
und
trut
h
nu
m
ber
s
wh
e
n
t
he
s
cene
i
t
evaluates
is
cr
owde
d
beyo
nd
an
uppe
r
-
t
hr
e
sh
ol
d
or
w
he
n
the
sce
ne
is
ver
y
li
ghtl
y
pa
cked
bel
ow
a
lowe
r
thres
ho
l
d
.
A
ddin
g
to
this,
the
m
od
el
do
e
s
no
t
pro
vid
e
real
-
ti
m
e
cl
ass
ific
at
ion
an
d
countin
g
of
diff
e
ren
t
cl
asses
of
obj
e
ct
s
in
the
crow
ded
sce
ne.
T
hi
s
will
be
i
m
po
rtant
w
hen
pro
cessi
ng
a
li
ve
f
eed
f
ro
m
a
m
o
un
te
d
ca
m
era
m
eant
for
s
urveil
la
nc
e
pur
poses
c
om
es
into
pictu
re
.
T
his
pa
per
discusse
s
a
fe
w
a
dd
it
io
nal
f
eat
ur
es
adopted
which
m
igh
t help
the
ITS
s
ect
or in
s
om
e w
ay
s:
1.
Veh
ic
le
detect
ion an
d
cl
assi
ficat
ion
from
a ca
m
era'
s v
ideo f
eed.
2.
A
pro
bab
il
ist
ic
syst
e
m
of
pre
dicti
ng
the
gre
en
-
ti
m
e
based
on
the
obta
ine
d
c
ount
of
different
cl
asse
s
of
veh
ic
le
s.
3.
Con
sta
ntly
ada
pting
the g
ree
n
-
li
gh
t
ti
m
e
based
o
n
t
he
c
hangin
g
nu
m
ber
s o
f
ve
hicle
s
res
ulti
ng
i
n
opt
im
al
tim
e eff
ic
ie
ncy f
or
a
cyc
le
.
3.
RESEA
R
CH MET
HO
D
The
entire
pro
cedure
is
div
i
ded
i
nto
3
ste
ps
.
T
he
pri
m
a
ry
ste
p
is
the
veh
ic
le
detect
i
on,
in
w
hic
h
cl
asses
of
ve
hicle
s
are
detect
ed
in
a
vid
e
o
f
eed.
T
his
is
th
en
fe
d
to
t
he
s
econd
ste
p
w
hi
ch
in
vo
l
ves
c
ountin
g
the
veh
ic
le
s
in
a
fr
am
e
and
su
bse
que
ntly
cou
ntin
g
the
tota
l
nu
m
ber
of
ve
hicle
s
in
the
vid
eo
f
eed
.
The
final
ste
p
c
om
pr
ise
s
of
pr
e
dicti
ng the
‘green
ti
m
e’
of the
n
e
xt cy
cl
e. In short
, t
he
steps a
re list
e
d
as:
a.
Veh
ic
le
detect
ion
b.
Counti
ng
of
ve
hicle
s in
a
fra
m
e and
he
nce
f
ro
m
the v
i
deo f
eed.
c.
Pr
e
dicti
ng
t
he “
gr
ee
n
-
ti
m
e” of
the
n
e
xt cycl
e.
3
.
1.
V
e
hicl
e d
etectio
n
Veh
ic
le
de
te
ct
ion
is
a
n
insta
nc
e
of
obj
ect
de
te
ct
ion
in
im
ages
an
d
vi
deo
s
.
Object
detect
io
n
ha
s
m
ade
gr
eat
stri
des
in
the
la
st
5
ye
ar
s
since
the
a
dvent
of
Ale
xN
et
[13]
for
im
age
cl
assifi
cat
ion
.
Im
age
cl
assifi
cat
ion
is
the
ta
sk
of
assigni
ng
or
identify
in
g
t
he
cl
ass
of
a
n
ob
j
ect
in
a
n
im
age.
Deep
le
ar
ning
ap
proac
hes
,
par
ti
cula
rly
C
onvoluti
onal
Neural
Netw
orks
ha
ve
pro
duced
near
hu
m
an
-
le
vel
accura
cy
in
this
fiel
d
.
Every
CNN
arc
hitec
ture
has
f
our
par
ts
–
c
onvo
luti
on
,
non
-
li
ne
arit
y,
su
bs
am
pling
fo
ll
owe
d
by
cl
ass
ific
at
ion
.
Object
detect
io
n
is
the
pr
oce
dure
of
cl
assify
ing
obj
ect
s
in
an
i
m
age
and
lo
cal
iz
ing
the
ex
te
nt
of
the
obje
ct
in
the
im
age
by
dr
a
wing
boun
ding
boxes
a
r
ound
it
.
T
he
f
irst
m
od
el
,
Re
gion
Ba
se
d
C
onvoluti
onal
Neural
Netw
orks
(R
-
CNNs
)
[14]
in
tuit
ively
beg
i
n
with
t
he
reg
i
on
sea
rch
an
d
t
hen
pe
rfo
rm
the
cl
assifi
cat
ion.
Since
then
s
eve
ral
m
od
el
s
ha
ve
bee
n
dev
el
op
e
d
t
ha
t
hav
e
great
ly
i
m
pr
ove
d
perf
or
m
ance
su
c
h
as
Fast
R
-
C
N
N
[
15]
,
Faste
r
R
-
C
NN
[16],
Re
gion
Ba
sed
F
ully
Conv
olu
ti
onal
Neural
Netw
orks
(R
-
FC
N)
[17]
.
T
hese
m
et
ho
ds
m
od
el
ob
j
ect
detect
ion
as
a
cl
assifi
cat
ion
prob
le
m
.
The
se
m
et
ho
ds
a
r
e
ver
y
acc
ur
at
e
bu
t
c
om
e
at
a
big
com
pu
ta
ti
on
al
cost (
l
ow fram
e
-
rate)
, in ot
he
r
w
ords
,
th
ey
a
re
no
t
fit t
o be
us
e
d on em
bed
ded d
e
vices.
Re
cent
ap
proa
ches
com
bin
e
these
tw
o
ta
sk
s
into
one
net
w
ork.
In
ste
a
d
of
hav
i
ng
a
netw
ork
pr
oduce
reg
i
on
proposa
l,
a
set
of
pr
e
-
de
fine
d
boxes
a
r
e
init
ia
li
zed
to
look
f
or
ob
j
ect
s.
This
cl
ass
of
detect
or
s
is
kn
own
as
Sin
gle
S
hot
Detect
or
s
.
T
he
al
go
rithm
t
his
pa
pe
r
us
es
is
the
Si
ng
le
Shot
Mult
iB
o
x
Detect
or
(
SSD)
,
by
G
oogle
[
18]
.
SSD
perfor
m
s
con
side
rabl
y
bette
r
than
it
s
co
m
petit
or
s
,
both
in
te
rm
s
of
s
pe
e
d
as
well
as
accuracy.
The
perf
or
m
ance
in
te
rm
s
of
s
pe
ed
is
m
easure
d
on
t
he
F
ra
m
es
Per
Sec
ond
(FPS)
it
processes
.
Fo
r
300
×
300
inp
ut,
SS
D
ac
hieve
s
74.
3%
m
AP
on
the
VO
C2
007
datas
et
at
59
FPS
on
an
N
Vidia
T
it
an
X.
This
is
in
co
m
par
ison
to
Faste
r
R
-
CN
N
at
7
FP
S
with
m
A
P
73.2
%
a
nd
Y
O
LO
at
45
FP
S
wit
h
m
AP
6
3.4%
[1
6].
The SSD
Mult
iB
ox m
et
ho
d ha
s 3
aspects:
Sing
le
S
hot:
The
ta
sk
s
of
obje
ct
local
iz
at
ion
and
cl
assifi
cat
ion
ar
e
done
i
n
a
sing
le
f
orwa
rd
pass
of
th
e
netw
ork
Mult
iB
ox
: T
he
techn
i
qu
e
for
boundi
ng bo
x r
egr
es
sio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
IJ
-
AI
V
ol.
8
,
No.
2, June
201
9
:
107
–
119
110
Detect
or
:
The
netw
ork
is a
n o
bj
ect
detect
or t
hat cla
ssifie
s t
ho
s
e
detect
ed object
s
The
SS
D
arc
hi
te
ct
ur
e
this
pa
per
em
plo
ys,
as
sh
ow
n
in
Fi
gure
2,
is
the
sam
e
as
the
on
e
presente
d
by
[1
8],
wh
ic
h
ba
ses
it
sel
f
on
VGG
a
rch
it
ect
ure.
T
hi
s
pap
e
r
us
es
a
n
al
rea
dy
exist
ing
m
od
el
that
su
it
ed
the
nee
ds
a
nd
hav
e
util
iz
ed
the
Ke
ras
port
of
Sin
gle
Shot
Mult
iB
ox
Detect
or
,
by
Pierl
uig
i
Fe
rr
a
ri
[
19]
.
This
is
buil
t
on
the K
e
ras fram
ewor
k
that
run
s Tenso
rf
l
ow in th
e
b
ac
k
-
en
d.
Figure
2
.
Th
e
SSD
net
work a
rch
it
ect
ure
[18]
The
SS
D
ap
proach
is
base
d
on
a
fee
d
-
f
orward
co
nvol
ution
al
netw
ork
that
produces
a
fixed
-
siz
e
colle
ct
ion
of boun
ding box
e
s an
d
sc
or
e
s f
or
the p
rese
nce
of o
bject
class i
ns
ta
nces in
th
ose
b
oxes, foll
owed by
a
non
-
m
axim
u
m
su
ppressi
on
ste
p
to
pro
du
ce
the
final
de
te
ct
ion
s.
Eac
h
add
e
d
la
ye
r
c
an
pro
duce
a
set
of
pr
e
dicti
on
s
.
Mult
iB
ox
is
a
reg
res
sio
n
te
ch
nique
that
sta
r
ts
with
the
pr
i
or
s
as
pre
dicti
on
s
a
nd
at
tempts
to
regress
cl
os
e
r
to
the
groun
d
truth
bo
unding
boxes,
base
d
on
tw
o
loss
f
unct
ions,
the
co
nf
i
den
ce
loss
,
an
d
locat
ion
lo
ss.
Confide
nce
lo
ss
m
easur
es
how
c
onfide
nt
it
is
of
fin
ding
an
obj
ect
wh
il
e
locat
io
n
loss
determ
ines
how
fa
r
a
way
it
is
from
the
act
ual
box.
Ba
se
d
on
this,
non
-
m
axi
m
u
m
su
ppressi
on
is
use
d,
i
n
wh
ic
h
th
res
ho
l
ding
of co
nf
i
de
nce loss i
s
done
a
nd
boxes
ar
e d
isc
ar
de
d.
Fo
r
cl
assifi
cat
ion
of
ve
hicle
s
on
t
he
r
oa
d,
th
e
arch
it
ect
ure
is
al
te
red
by
subsam
pling
the
neur
on
s
i
n
the
netw
ork
s
uch
that
the
num
ber
of
cl
as
ses
include
d
a
nd
trai
ned
is
f
or
car
s,
m
oto
rcycl
es,
truck
s
,
bu
ses
,
per
s
on
a
nd
bac
kgr
ound.
E
ve
r
y
obj
ect
detect
ed
in
the
fr
am
e
belo
ngs
to
one
of
t
he
6
cl
a
sses.
B
oundin
g
boxes
are
dr
a
wn
ar
ound
t
hem
to
ind
ic
at
e
t
heir
posit
ion
as
well
.
T
his
ob
j
ect
detect
ion
is
ap
plied
to
the
vi
deo
of
veh
ic
le
s
passi
ng
thr
ough
t
he
jun
ct
io
n
durin
g
the
green
-
ti
m
e.
As
lo
n
g
a
s
the
sig
nal
is
gr
ee
n
f
or
a
la
ne,
t
he
m
ov
ing
ve
hicle
s
are
detect
ed
by
the
m
od
el
.
This
pro
vid
es
var
i
ou
s
a
dv
a
nt
ages
com
par
ed
to
detect
ing
ve
hicle
s
durin
g
red
li
gh
t,
as
it
no
t
only
i
m
pr
ov
e
s
the
accuracy
of
de
te
ct
ion
due
t
o
cl
ear
visibil
it
y
of
eac
h
veh
i
cl
e
in
the
fr
am
e,
bu
t
al
so
re
du
ce
s
c
om
pu
ta
ti
on
ne
eded
since
on
l
y
on
e
la
ne
nee
ds
to
be
obse
rved
du
rin
g
sin
gle
gr
ee
n
tim
e
instea
d
of
obser
ving
the
3
la
nes
wait
in
g
in
t
he
re
d
-
sign
al
.
Sho
w
in
Figure
3
O
utput
of
the
SS
D
Objec
t
Detect
ion
Al
gorithm
.
Figure
3
.
O
utput
of t
he
S
SD
obj
ect
detect
io
n
al
go
rithm
[20]
3
.
2.
Cou
nt
in
g v
ehic
le
s
I
t
w
ou
l
d
be
ve
ry
be
nef
ic
ia
l
to
co
unt
the
nu
m
ber
of
ve
hicle
s
w
hile
they
are
passi
ng
t
hro
ug
h
the
jun
ct
i
on
duri
ng
gr
ee
n
-
li
ght
instea
d
of
c
on
si
der
i
ng
a
sing
le
im
age
durin
g
re
d
-
li
ght,
therefo
re
,
it
be
com
es
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
AI
IS
S
N:
22
52
-
8938
Ad
ap
ti
ve re
al t
ime
traff
ic
p
r
e
dicti
on
us
i
ng dee
p neural
net
works
(
P
ar
init
h
R
I
yer
)
111
pivotal
to
c
ome
up
with
a
w
ay
to
co
unt
th
e
num
ber
of
ve
hicle
s
of
eac
h
cl
ass
t
hat
pa
ssed
t
hro
ugh
a
scen
e
wh
il
e
m
ai
ntain
ing
real
-
tim
e
s
peeds
of
obj
ec
t
detect
ion
.
Co
un
ti
ng
the
num
ber
of
obj
ect
s
in
a
sti
ll
i
m
a
ge
i
s
si
m
ple.
It
is
com
m
on
place
no
w
in
e
ver
y
ob
je
ct
detect
ion
a
lgorit
hm
bu
t
countin
g
f
ro
m
a
vid
e
o
fee
d
is
tric
ky
because
eve
ry
fr
am
e
is
treat
e
d
in
dep
e
ndentl
y
and
m
igh
t
cou
nt
duplica
te
s.
In
the
w
or
st
-
ca
se
scenari
o,
it
m
igh
t
keep
co
unti
ng
duplica
te
s
un
ti
l
the
obj
ect
is
ou
t
of
the
fr
am
e.
The
fol
lowing
m
et
ho
d
is
a
n
at
te
m
pt
a
t
al
le
viati
ng
this
problem
an
d o
btainin
g
a
n
a
ppr
oxim
a
te
g
r
ound tr
uth
c
ount.
The fre
quently
u
se
d
te
rm
s ar
e
:
OP
T
-
Op
ti
m
al
Pr
oc
essin
g
Ti
m
e.
F
i
–
The
i
th
f
ra
m
e in the vide
o feed
.
T
i
–
Tim
e
ta
ke
n
to
pr
ocess
t
he
i
th
fr
am
e.
W
i
–
T
he
c
ontr
ibu
ti
on
of the i
th
fr
am
e.
Veh
ic
le
sCo
unte
d
–
St
ru
ct
ur
e
ho
l
ding the
nu
m
ber
o
f ve
hicle
s of eac
h
cl
as
s cou
nted
i
n
th
e lat
est
cyc
le
.
SideBy
Side
–
Stru
ct
ur
e
desc
ribing
t
he
num
ber
of
ve
hic
le
s
of
eac
h
cl
ass
capa
ble
of
sim
ultaneou
sl
y
cro
ssi
ng the
int
ersecti
on side
by side
.
Ti
m
eTaken
–
Stru
ct
ur
e
holdi
ng the tim
e taken for a
ve
hicle
of eac
h
cl
ass
to cross
the i
ntersecti
on.
Ti
m
eGiven
–
S
tructu
re
ho
l
ding the
r
at
io
of
T
i
m
eTaken
to
Si
deBy
Side.
Pr
e
vP
re
dict
–
The
green
ti
m
e
tha
t
was
al
locat
ed
f
or
t
he
pr
ese
nt
gr
e
en
cy
cl
e
afte
r
fee
db
ac
k
fro
m
the pre
vious
green cycl
e.
Pr
ese
ntAppr
ox
–
The
ap
pro
xim
a
te
t
i
m
e
t
hat
had
to
be
giv
en
to
the
pr
ese
nt
cy
cle,
con
si
der
i
ng
the v
e
hicle
s th
at
p
asse
d.
Nex
tP
re
dict
–
The pre
dicti
on
m
ade for
t
he n
ex
t g
ree
n
cy
cl
e.
a
m
in
–
The
m
ini
m
u
m
accel
erat
i
on of a
v
e
hicle
whil
e cr
os
sin
g t
he
inte
rsecti
on.
a
m
ax
–
The
m
axi
m
i
m
acce
le
rati
on
of a
ve
hicle
whil
e cr
os
sin
g
the
interse
ct
ion.
Con
si
der
i
ng
th
at
an
obj
ect
de
te
ct
ion
al
go
rithm
isn’t
al
wa
ys
perfect
a
nd
m
igh
t
detect
non
-
existe
nt
obj
ect
s
i
n
a
sc
ene,
an
intuit
iv
e
ap
proac
h
is
t
o
c
onsider
that
an
ob
j
ect
is
ac
tuall
y
pr
ese
nt
i
n
the
sce
ne
only
if
it
per
sist
s
thr
ough
so
m
e
nu
m
b
er
of
f
ram
es.
Al
so
ta
king
in
to
account
the
un
sta
ble
nat
ure
of
ha
r
dw
a
r
e
that
perform
s o
bj
ec
t detec
ti
on
, e
ve
ry f
ram
e takes a d
iffe
re
nt am
o
un
t
of
ti
m
e to
process
depen
di
ng
on
c
onditi
ons at
that
instant,
t
he
nu
m
ber
of
obj
ect
s
i
n
the
s
cene,
et
c.
This
m
igh
t
le
ad
to
cases
w
her
e
t
he
obj
ect
is
pr
e
sent
in
the
sce
ne
for
the
require
d
am
ou
nt
of
tim
e
bu
t
is
pr
esent
on
ly
in
a
few
fr
am
es
becau
se
on
ly
few
ha
ve
be
e
n
captu
red
an
d
proces
sed
.
L
ocki
ng
t
he
f
ram
e
rate
w
on’t
help
ei
the
r
as
pro
cessi
ng
tim
e
m
igh
t
en
d
up
higher
than
th
e
de
fin
ed
ti
m
e
interv
al
betwee
n
f
r
a
m
e
captur
e.
To
s
olv
e
t
his,
there
had
t
o
be
s
om
e
kin
d
of
a
“con
t
rib
ution
”
par
am
et
er
or
"weig
ht"
to
e
ve
ry
fr
am
e
that
was
processe
d
base
d
on
w
hi
ch
a
decisi
on
can
be
m
ade
wh
et
he
r
to
ta
ke
the
vehi
cl
es
in
this
fr
am
e
seriou
sly
or
no
t.
Id
eal
ly
,
a
fr
am
e
is
capt
ur
e
d
at
interval
s
suc
h
that
no
veh
ic
le
is
pr
esent
in
t
wo
c
onsecuti
ve
fr
am
es
and
that
al
l
veh
ic
le
s
passing
thr
ough
a
scene
e
xist
in
exactl
y
on
e
fra
m
e.
This
interval
is
cal
le
d
“Op
ti
m
al
Pr
ocessing
Tim
e”
(O
PT,
t
he
tim
e
it
ta
kes
to
process
a
fr
am
e and
t
he
n
fetch
the
ne
xt
on
e
fro
m
the c
a
m
era).
Ob
ta
ini
ng
OPT
is
it
sel
f
a
matt
er
of
co
nce
r
n
as
it
is
no
t
stric
tl
y
def
ined
neither
ca
n
it
be
m
easur
ed
accuratel
y.
T
hi
s
has
m
uch
to
do
with
la
r
ge
-
scal
e
data
ac
quisi
ti
on
a
nd
da
ta
analy
sis
pro
cedures
t
o
fig
ure
ou
t
how
m
uch
ve
hi
cl
es
acce
le
rate
w
hile
cr
os
si
ng
ju
nctio
ns
a
nd
if
it
is
relat
ed
to
tra
ff
ic
de
ns
it
y
it
sel
f,
how
this
tren
d
cha
ng
es
acro
s
s
dif
fer
e
nt
den
sit
ie
s
of
traf
fic.
The
siz
e
of
the
ju
nction
al
s
o
com
es
into
play
as
t
he
OPT
increases
as
th
e
siz
e
do
es.
T
he
siz
e
of
the
jun
ct
io
n
can
be
easi
ly
cal
culated
base
d
on
how
m
uc
h
the
ca
m
era
can
see
in
t
he
scene
.
T
he
ac
cel
erati
on
of
ve
hicle
s
is
bo
und
to
be
a
dat
a
-
sci
ence
pro
bl
e
m
that
need
s
huge
a
m
ou
nts
of
obs
erv
at
io
ns
.
S
how
in
Fig
ure
4
I
ll
us
trat
ing
t
he defi
niti
on
of OP
T.
Figure
4
.
I
ll
us
t
rati
ng the
de
fin
it
ion
of
OP
T
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
IJ
-
AI
V
ol.
8
,
No.
2, June
201
9
:
107
–
119
112
Con
si
der
i
ng
duri
ng
O
PT
,
n
nu
m
ber
of
fr
a
m
es
wer
e
capt
ur
e
d
a
nd
eac
h
fram
e
F
i
took
tim
e
T
i
to
process
.
Ther
e
f
or
e,
=
∑
=
1
(
1
)
If
a
fr
am
e
Fi
ta
kes
lo
ng
e
r
t
o
pr
ocess,
t
he
i
m
m
ediat
e
ne
xt
fr
am
e
F
i+1
mi
gh
t
be
m
iss
ing
ou
t
on
ca
pturin
g
the
ve
hicle
in
the
scene.
So,
to
com
pen
sat
e
fo
r
t
his,
F
i
shou
l
d
ha
ve
a
hi
gh
e
r
co
ntri
bu
ti
on
wh
il
e
co
ns
i
der
i
ng
the
ve
hicle
s
in
it
.
Si
m
il
arly
,
i
f
F
i
is
pr
ocesse
d
to
o
qu
i
c
kly,
there
m
igh
t
be
m
or
e
occ
urre
nc
es
of
the
veh
i
cl
e
in
the
scene.
T
his
is
co
m
pen
sat
e
d
by
giv
i
ng
a
lowe
r
co
ntribut
ion
to
F
i
.
It
is
ob
s
er
ved
that
t
he
“weig
ht”
W
i
of
a
fr
am
e
F
i
is dire
ct
ly
p
rop
or
ti
on
al
to
the ti
m
e T
i
ta
ken to
proce
ss it
.
∝
(2)
The
i
niti
al
assum
ption
is
t
hat
the
ve
hicle
s
a
pp
ea
r
i
n
t
he
sc
ene
only
once
durin
g
OP
T
,
t
he
n
t
he
weig
ht
W
i
of
a
giv
e
n fr
am
e
is def
i
ned b
y
:
=
(3)
This
de
rivati
on
sh
i
fts
the
re
qu
i
rem
ent
fr
om
“A
veh
ic
le
sh
oul
d
be
pr
e
s
ent
in
a
s
pecif
ie
d
num
ber
of
f
ram
es
captu
red
by
th
e
ca
m
era”
to
“A
ve
hicle
shoul
d
be
pr
ese
nt
in
the
cam
era'
s
view
f
or
a
sp
e
ci
fic
interval
of
tim
e”.
This
ti
es
the
ca
lc
ulati
on
s
t
o
a real
-
w
orl
d
value
f
ree f
r
om
a
m
big
uity
,
instea
d
of
ty
in
g
it
to
the
al
ways
-
c
hangi
ng
value o
f
f
ram
e
-
rate.
The
“
co
nt
ribu
ti
on”
of a
fr
am
e thr
ou
ghou
t
OPT
ca
n b
e v
is
ualiz
ed
as
sh
ow
n
in
Fi
gur
e 5
.
Figure
5
.
I
ll
us
t
rati
ng the c
ontr
ibu
ti
on
of each
f
ram
e in a
s
pa
n of OP
T
This
m
eans,
t
hat
if
the
ve
hi
cl
e
did
a
ppea
r
in
the
cam
e
ra'
s
view
f
or
as
lo
ng
as
O
PT
a
nd
wa
s
even
t
ually
captur
e
d
in
al
l
the
fr
am
es
that
wer
e
ta
ken
acr
oss
the
tim
e
sp
an
of
OP
T
,
it
w
il
l
ult
i
m
at
e
ly
e
nd
up
bein
g
c
ounte
d
as
one v
e
hicle
,
w
hich
is
the
de
sired
outc
o
m
e.
N
ow
,
to
deci
d
e
whet
her
a de
te
ct
ed
ve
hicle
that
is
a
le
gitim
at
e
on
e,
the
veh
ic
le
s
can
be
co
unte
d
by
c
onside
ring
t
he
total
nu
m
ber
of
veh
ic
l
es
of
e
ach
cl
as
s
in
a
fr
am
e
and
m
ult
iply
ing
them
by
W
i
,
i
.e
.
wei
ght
of
t
he
f
ram
e
.
This
is
essent
ia
ll
y
the
con
tri
bu
t
io
n
of
al
l
ve
hicle
s
in
a
fr
am
e
to
be
acce
pted
as
a
le
gitim
a
te
vehi
cl
e
as
tim
e
go
es
on.
Co
ntinui
ng
this
unti
l
any
desire
d
tim
e,
it
is
po
s
sible
to
obt
ai
n
an
ap
pro
xi
m
at
e
cou
nt
of
t
he
veh
ic
le
s.
F
or
ex
am
ple,
if
‘
Ca
rs
t
’
is
t
he
num
ber
of
ca
rs
pa
ssing
thr
ough
the
sce
ne
in
ti
m
e
t
, an
d ‘
Ca
rs
i
’
is t
he
num
ber
of
ca
rs
i
n
a
fra
m
e
F
i
, and
n
f
ram
es are cap
t
ur
e
d
i
n
ti
m
e
t,
=
∑
(
∗
)
=
1
(
2
)
Si
m
il
arly
,
it
is
po
s
sible t
o
get
the app
roxim
ate
co
un
ts
of e
ve
ry v
e
hicle
clas
s in
ti
m
e
t.
Con
si
der
i
ng
t
he
act
ual
traf
fic
jun
ct
io
n,
t
his
tim
e
t
,
is
the
tim
e
giv
en
f
or
t
he
veh
ic
le
s
to
pass
(gree
n
-
tim
e).
Applyi
ng
the
ab
ove
f
or
m
ula
for
th
e
tim
e
per
io
d
of
t
he
al
locat
e
d
gr
ee
n
-
ti
m
e,
the
tot
al
num
ber
of
veh
ic
le
s
of
eac
h
cl
ass
that
pa
ss
th
rou
gh
the
j
unct
io
n
is
obta
ined.
A
po
i
nt
to
kee
p
in
m
ind
is
t
hat
the
ca
m
era
m
us
t
hav
e
the
j
unct
io
n
a
nd
nothin
g
el
se
in
i
ts
view.
If
othe
r
r
oads,
or
th
e
opposit
e
la
ne
s
are
capt
ur
e
d
by
the
ca
m
era,
this m
i
gh
t l
ea
d
to
discrep
a
ncies in
ob
ta
ining
t
he
e
xa
ct
co
unt
of the
veh
ic
le
s.
3
.
3.
Predic
tin
g
green
time
The dat
a
ob
ta
i
ned can
b
e
v
is
ualiz
ed
as
foll
ows.
Veh
ic
le
sCo
unte
d
{
Ca
rs
=
12.78;
2
-
w
heeler
s
= 19.
54
;
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
AI
IS
S
N:
22
52
-
8938
Ad
ap
ti
ve re
al t
ime
traff
ic
p
r
e
dicti
on
us
i
ng dee
p neural
net
works
(
P
ar
init
h
R
I
yer
)
113
Tr
uc
ks
=
3.
16
;
Buses
=
1.6
2;
Bi
cy
cl
e = 5
.
31
;
}
This
sho
ws
the
approxim
at
e
nu
m
ber
of
ve
hicle
s
of
eac
h
cl
as
s
at
an
assum
ed
tim
e.
An
a
rr
a
y
def
ini
ng
the am
ou
nt of ti
m
e that every v
ehicl
e cla
ss
ne
eds
to c
ross the ju
nctio
n
wil
l be m
a
intai
ned
. T
his ag
ai
n
de
pend
s
on
t
he
siz
e
of
the
jun
c
ti
on
unde
r
co
ns
i
der
a
ti
on
an
d
the
si
ze
of
the
r
oad
as
m
ulti
ple
veh
ic
le
s
of
t
he
s
a
m
e
or
diff
e
re
nt
cl
ass
can
cross
the
jun
ct
io
n
side
by
side
as
sh
own
in
Figure
6
,
and
the
ave
ra
ge
acce
le
rati
on
with
wh
ic
h
the
v
e
hi
cl
es start cr
os
si
ng the
jun
ct
i
on.
Figure
6
.
Top
View o
f vehicl
es cr
os
sin
g
a
n
i
ntersecti
on sim
ultaneo
us
ly
Ob
ta
ini
ng
the
aver
a
ge
acce
le
rati
on
ca
n
be
thou
gh
t
of
as
a
cl
assic
data
s
ci
ence
pro
blem
inv
olv
i
ng
la
rg
e
-
scal
e
dat
a
-
acq
uisit
ion
a
nd
a
naly
sis,
an
d
the
rest
of
th
e
par
am
et
ers
will
be
sp
eci
fi
c
to
ever
y
ju
nc
ti
on
.
It
will
be
a
one
-
t
i
m
e
entry
of
pa
ram
et
ers
and
good
to
go
un
l
ess
the
ro
a
d
w
idth
is
c
hange
d
or
t
he
j
unct
ion
is
com
plete
ly
rec
on
st
ru
ct
e
d.
As
su
m
ing
the
road
unde
r
c
on
si
der
at
io
n
is
wi
de
e
nough
to
occupy
the
f
ol
lowi
ng
nu
m
b
er
of v
e
hi
cl
es o
f
the
sam
e
cl
ass side
b
y
side,
SideBy
Side {
Ca
rs
=
3;
2
-
w
heeler
s
= 6
;
Tr
uc
ks
=
2
;
Buses
=
2;
Bi
cy
cl
e = 7;
}
Ma
intai
nin
g
a
no
t
her
str
uctu
r
e
(arbit
rar
y
val
ues
f
or
now
)
de
fining
ho
w
m
uch
ti
m
e
a
s
ing
le
ve
hicle
of
a
cl
ass
ta
kes
to
cr
os
s a
n
em
pty juncti
on (
i
n
sec
onds)
,
Ti
m
eTaken
{
Ca
rs
=
4.67;
2
-
w
heeler
s
= 3
.
5;
Tr
uc
ks
=
6.
83
;
Buses
=
6.2
9;
Bi
cy
cl
e = 5
.
16
;
}
The
struct
ur
e
Ti
m
eTaken
is
al
so
a
resu
lt
of
la
rg
e
-
scal
e
dat
a
gather
i
ng
an
d
data
analy
s
is
pr
oce
dures
.
O
nce
al
l
the
above
-
m
entioned
data
is
avail
able,
the
tim
e
to
be
give
n
to
each
ve
hicle
to
cl
ear
the
j
unct
io
n
c
an
be
decide
d upo
n.
The
Tim
eGive
n
for
eac
h veh
i
cl
e cla
ss w
il
l h
ave to be:
[
]
=
[
]
[
]
(5)
This
is
to
accou
nt
f
or
a
reali
s
ti
c
traff
ic
flow
scenario.
In
a
ro
a
d
wide
e
nough
to
hold
3
cars,
3
car
s
can
cr
os
s
the
jun
ct
io
n
in
t
he
sam
e
tim
e
as
i
t
ta
kes
for
a
si
ng
le
ca
r
to
cr
oss
a
jun
ct
io
n
if
they
are
al
l
side
by
side.
S
o,
m
aki
ng
this
twea
k
is
i
m
po
rtant
wh
il
e
predict
ing
the
gr
ee
n
-
tim
e
fo
r
the
ne
xt
cy
cl
e.
Consi
der
a
real
-
li
fe
scena
r
io
wh
e
re
so
m
e
value
of
gr
ee
n
-
tim
e
fo
r
the
present
cy
cl
e
(
Pr
ev
Pr
e
dict
)
is
pr
e
viously
pr
e
dicte
d
.
Now,
duri
ng
a
tim
e
interval
of
le
ng
t
h
‘
Pr
e
vP
r
e
dict
’,
the
nu
m
ber
of
ve
hi
cl
es
that
did
c
ro
ss
the
j
unct
i
on
a
re
store
d
in
‘
Ve
hi
cl
esCoun
te
d
’.
The
pre
dicti
on
a
bout
Pr
e
vPredict
m
igh
t
ha
ve
bee
n
wrong,
or
the
traf
fi
c
flo
w
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
IJ
-
AI
V
ol.
8
,
No.
2, June
201
9
:
107
–
119
114
m
igh
t
be
changin
g,
ei
ther
w
ay
,
it
’s
i
m
po
rtant
to
know
how
m
uch
tim
e
sh
o
uld
act
uall
y
hav
e
bee
n
gi
ven
to
the
nu
m
ber
of
veh
ic
le
s
t
hat
w
ere
j
ust
co
unte
d
in
the
prese
nt
cy
cl
e
and
sto
red
in
Ve
hicle
sCounted
.
T
his
data
will
serv
e
as
the
"gro
und
-
t
r
uth
"
val
ue,
li
ke
in
al
l
oth
er
sel
f
-
co
rr
ect
in
g
syst
e
m
s.
This
can
end
up
bein
g
the f
ee
dback
th
at
’s
nee
de
d.
T
he
refor
e
,
P
rese
nt
Appro
x
is cal
c
ulate
d
as:
=
∑
(
ℎ
[
ℎ
]
∗
[
ℎ
]
)
ℎ
=
(6)
The
val
ue
of
P
resen
t
App
rox
m
igh
t
be
gr
eat
er
or
le
sser
tha
n
Pr
e
vPredict
,
nonetheless
,
it
is
us
ed
as
a
fee
db
ac
k
m
echan
ism
to
i
m
pr
ove
our
pr
e
dicti
on
'
s
accuracy
f
or
the
green
-
ti
m
e
to
be
giv
e
n
in
the
ne
xt
cy
cl
e
(
Ne
xtPr
e
dict
).
In
t
uiti
vely
,
Ne
xtPr
e
dict
will
hav
e
to
ta
ke
t
he
foll
ow
i
ng tw
o values
into
c
on
si
der
at
io
n:
1.
The pre
dicti
on
m
ade prev
i
ou
sl
y for the
prese
nt cycl
e (
P
revP
red
ic
t
)
2.
The
tim
e
that
had
t
o
be
giv
e
n
to
the
nu
m
ber
of
ve
hicle
s
that
act
ually
cro
ssed
the
j
unct
io
n
in
the
pr
ese
nt
cy
cl
e (
Pr
ese
nt
Appro
x
)
Using
Weig
hte
d
Mo
vi
ng Ave
rag
e
,
Nex
tP
redi
ct
can be cal
c
ulate
d
as:
=
(
∗
)
+
(
(
1
−
)
∗
)
(7)
wh
e
re
is
the
trut
h
value
gi
ven
to
our
ob
s
erv
at
io
ns
duri
ng
the
prese
nt
cy
cl
e.
T
he
hig
he
r
t
he
v
al
ue
of
,
the
faste
r
the
syst
e
m
recti
fies
it
sel
f
but
increasin
g
al
so
dism
isses
the
c
on
t
rib
ution
of
P
revPr
edict
,
so
c
onve
ntio
na
ll
y,
the
val
ue
of
is
ta
ken
t
o
be
0.5.
A
fter
the
value
of
Nex
tP
re
dict
is
cal
culat
ed,
t
he
green
-
tim
e
giv
en
in
the
nex
t
cy
cl
e
wi
ll
be
the
value
of
Ne
xtP
re
dict
and
w
he
n
the
nex
t
cy
cl
e
sta
rts,
the
value
of
Pr
e
vP
re
dict
w
il
l
be
m
ade
e
qu
al
to
Ne
xtP
red
ic
t
,
he
nce
m
ai
ntaining
th
e
con
tri
bu
ti
on
of
our
pr
e
di
ct
ion
s
thr
oughout co
nse
cutive cy
cl
es. A
s the assu
m
ption
is that t
he
d
ensity
o
f
th
e tra
ff
ic
ch
a
nges g
ra
du
al
ly
spann
i
ng
m
ul
ti
ple
cy
cl
es
throu
ghout
th
e
day,
it
can
be
visu
al
iz
ed
ho
w
this
ap
proac
h
w
ould
al
way
s
pro
vid
e
a
ppr
opriat
e
gr
ee
n
-
ti
m
e
pr
e
dicti
on
s
.
Let
’
s
consi
der
a
hypotheti
cal
tim
e
-
var
yi
ng
data
si
m
ula
ti
ng
the
ri
se
in
t
raffic
de
ns
it
y
up un
ti
l a
r
ush
hour a
nd the
n back
do
wn ag
a
in.
Figure
7
.
V
a
riat
ion
of Pr
e
vPre
dict an
d
P
rese
nt
Appro
x
il
lustr
at
ing
acc
ur
acy
and self
-
co
rr
ec
ti
ng
natu
re
A
ty
pical
rise
and
fall
in
traf
fic
den
sit
y
dur
ing
r
ush
hour
would
s
pan
a
r
ound
50
cy
cl
es,
bu
t
f
or
the
sake
of
te
sti
ng,
a
hypoth
et
ic
al
le
ss
grad
ual
c
hange
has
bee
n
co
ns
ide
red.
T
h
e
gr
a
ph
in
Fig
ur
e
7
s
how
s
ho
w
the
pr
e
dicti
on
c
orr
ect
s
it
sel
f
ever
y
cy
cl
e
bit
by
bit
and
m
an
ages
to
kee
p
up
with
the
in
creasin
g
num
ber
of
veh
ic
le
s
durin
g
r
ush
hour.
It
al
so
e
ven
s
out
when
t
he
tra
ffi
c
den
sit
y
sat
urat
es
an
d
a
gai
n
re
duces
as
t
he
rush
hour
com
es
to
an
en
d.
H
ence
,
this
appro
ac
h
of
lo
okin
g
at
traf
fic
li
gh
t
aut
om
ation
w
ould
resu
lt
in
eff
ic
i
ently
util
iz
ing
tim
e
and
c
ut
do
wn
a
lot
of
wait
in
g
tim
e
at
sign
al
s.
This
m
od
el
can
easi
ly
be
scal
ed
up
to
a
4
-
roa
d
intersect
io
n,
a
nd
e
ven
th
ough
t
his
m
od
el
tr
eat
s
eve
ry
r
oa
d
at
a
n
interse
ct
ion
i
nd
e
pe
ndently
,
due
t
o
it
s
sel
f
-
correct
ing
nat
ur
e
,
com
poun
ding
ef
fects
of
s
uch
syst
e
m
s
us
e
d
at
c
onsecuti
ve
tra
ff
i
c
sign
al
s
will
al
so
be
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
AI
IS
S
N:
22
52
-
8938
Ad
ap
ti
ve re
al t
ime
traff
ic
p
r
e
dicti
on
us
i
ng dee
p neural
net
works
(
P
ar
init
h
R
I
yer
)
115
accounte
d
f
or
wh
il
e
pr
e
dicti
ng
t
he
nex
t
cy
cl
e’s
green
-
ti
m
e.
In
t
his
wa
y,
ch
ok
i
ng
of
traff
ic
ac
ro
s
s
m
ul
ti
ple
traff
ic
li
ghts ca
n
al
s
o be all
evi
at
ed.
3
.
4.
Decidi
n
g Hyper
-
P
aram
eter
Va
lues
The
SSD
m
od
el
has
bee
n
tra
ined
on
the
M
S
-
COC
O
(Com
m
on
obj
ect
s
in
co
ntext
)
dataset
[21],
a
nd
then
t
he
weig
ht
s
of
the
ne
ur
a
l
netw
ork
ha
ve
bee
n
s
ubsam
pled
to
fit
the
6
obj
ect
cl
asses
(a
nd
delet
in
g
oth
e
r
cl
asses
in
the
l
ast
la
ye
r)
that
we’re
interest
e
d
in
–
Ca
r,
t
ruck,
per
s
on,
bicy
cl
e,
m
oto
rcycl
e,
bus.
T
he
va
lue
of
OP
T
,
wh
ic
h
i
s
de
pe
nd
e
nt
on
t
he
veh
ic
le
’
s
acce
le
rati
on
and
velocit
y
pro
file
s
at
tra
ffi
c
intersect
io
ns
wa
s
decide
d
base
d
on
previ
ous
re
search
in
t
his
area
[
22
-
24]
.
I
n
[
22]
,
the
a
cc
el
erati
on
of
th
e
1
st
li
ne
of
ca
rs
was
ob
s
er
ved
ove
r
the
dista
nce
of
a
zeb
ra
c
rossing,
sta
rtin
g
fro
m
wh
en
the
ve
hicle
s
we
re
at
rest.
T
he
value
s
of
acce
le
rati
on
s
r
ecorde
d
acr
os
s
diff
e
re
nt
m
od
el
s
of
car
s
al
th
ough
va
ried
,
f
ollow
e
d
a
nor
m
al
distribu
ti
on
with
a
m
ean
of
arou
n
d
1.8
5
m
s
-
2
.
B
ut
this
stud
y
al
so
sta
te
d
that
the
cars
in
the
3
rd
or
4
th
li
ne
had
ne
gligibl
y
low
acce
le
rati
on
v
a
lues in
the o
r
de
r
of
0.5 m
s
-
2
wh
ic
h
m
akes s
ense if
one i
m
a
gin
es
dr
ivi
ng
a
car
in su
ch
a posit
i
on
at
a
traf
fic
sig
nal.
T
he
veloc
it
y
of
ve
hicle
s
,
h
oweve
r,
is
n’t
eve
r
inc
reas
ing
a
nd
sat
ura
te
s
an
d
s
om
e
wh
at
reaches
a
m
axim
u
m
wh
il
e
cr
os
sin
g
the
i
ntersecti
on.
I
n
[
19]
,
it
is
fo
un
d
that
the
ave
rage
acce
le
rati
on
of
a
ny
veh
ic
le
ste
adily
decr
eases
wi
th
the
increase
in
the
sp
eed
of
the
veh
ic
le
and
assum
ing
the
velocit
y
of
ve
hicle
s
do
e
sn’t
excee
d
30
or
40
kph
wh
il
e
cr
os
si
ng
an
inter
sect
io
n,
it
can
be
in
f
err
e
d
f
ro
m
the
researc
h
in
[
22
-
23]
that t
he
a
ver
a
ge
accel
erati
on
change
w
hile c
ro
ssi
ng of
dif
fe
ren
t cl
asses
of
veh
ic
le
s a
re as
fo
ll
ows.
Ca
r
=
1.85
–
0.
5
m
s
-
2
Tru
c
k, Bus
=
1.0
–
0.2
9m
s
-
2
Motorcycl
e =
0.94
–
0.4
7m
s
-
2
This
acce
le
rati
on
is
no
t
stric
t
ly
a
fu
nctio
n
of
tim
e
or
dista
nce
at
a
traf
fic
intersect
ion.
I
ns
te
ad
,
it
dep
e
nd
s
on
factors
su
c
h
a
s
oth
e
r
ve
hicle
s
on
t
he
r
oa
d,
possible
inte
rf
e
ren
ce
from
people
wal
king,
bad
dr
i
vers,
eve
n
anim
a
ls
on
the
ro
a
ds
at
tim
es.
The
cl
os
est
one
can
get
to
de
fine
the
acce
l
erati
on
pro
file
is
by
an
ass
umpti
on
that
it
is
stric
tly
a
functi
on
of
dista
nce
or
ti
m
e.
Her
e
,
the
assum
ption
is
that
it
is
a
fun
ct
ion
of
distan
ce
is
const
antly
decr
easi
ng
an
d
ev
entuall
y
sat
ur
a
te
s
near
zer
o.
This
kind
of
a
functi
on
re
se
m
bles
an
exp
onentia
l
decay f
unct
ion, so fit
ti
ng the
acce
le
rati
on
pa
ram
et
ers
that were
just
decid
ed upo
n,
(
)
=
max
−
(8)
Wh
e
re
(consi
de
rin
g
X
is t
he
leng
t
h of t
he
i
ntersecti
on un
der co
ns
ide
rati
on)
,
=
(
m
a
x
m
i
n
)
(9)
This
acce
le
rat
ion
c
urve
c
orres
pondin
g
t
o
this
e
quat
ion
an
d
a
25
-
m
et
er
intersect
ion
f
or
a
car
is
sh
ow
n
in
Fi
gur
e
8.
Figure
8
.
A
cce
le
rati
on
pro
file
of a ca
r
acc
ordin
g
to
a(
x)
This
m
akes
sense
f
or
sho
rt
distances
s
uc
h
as
an
inte
rse
ct
ion
w
he
re
the
vel
ocity
functi
on
is
a
re
su
lt
of
integrati
ng t
he a
ccel
erati
on
f
unct
ion wit
h dis
ta
nce ‘
x
’
1
2
(
(
)
)
2
=
∫
(
)
.
(10)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
IJ
-
AI
V
ol.
8
,
No.
2, June
201
9
:
107
–
119
116
The vel
ocity
f
unct
ion o
btaine
d
is:
(
)
=
+
√
2
ma
x
(
1
−
−
)
(11)
Plott
ing
v(x
)
a
gainst
x,
Th
e
gr
a
ph
s
h
own
i
n
Fi
gure
9
w
ould
m
ake
sense
if
one
relat
es
sel
f
-
dr
i
ving
th
rou
gh
an
inter
sect
ion
an
d
check
i
ng
t
he
ve
locit
y
of
the
c
ar
as
one
cr
os
s
es
and
at
th
e
ve
ry
en
d
of
it
.
A
lt
ho
ug
h
this graph
doesn’t
ho
ld
tru
e
for
lo
ng
distan
ces,
on
ly
s
hort
distance
s
a
re
r
el
evan
t
to
this
pro
blem
.
Ti
m
e
t,
ca
n
be
obta
ined
as
a
f
unct
ion
of
distance
x
by i
nteg
rati
ng as,
(
)
=
∫
1
(
)
.
(12)
(
)
=
(
√
1
−
+
1
)
−
(
|
√
1
−
−
1
|
)
√
2
ma
x
+
(13)
as
t(x
)
is
zer
o
w
he
n
x
is
zero,
a
de
finit
e
equ
at
i
on
for
t(x)
is
obta
ined
a
nd
by
pa
ssing
t
he
le
ngth
of
the
intersect
io
n
‘
X
’
i
nto
t(
x)
,
the
ti
m
e
that
a
certai
n
ve
hic
le
ty
pe
ta
kes
to
cr
os
s
t
he
int
ersecti
on
is
ob
ta
ined
wh
ic
h
is
nothi
ng
but
the
O
P
T
that
is
nee
de
d.
T
his
val
ue
c
an
al
so
be
inte
rpreted
as
t
he
‘
Tim
eTaken
’
f
or
t
hat
sp
eci
fic cl
ass
of
veh
ic
le
s
from
wh
ic
h ‘
Tim
eGiven
’
ca
n
al
so
be deri
ved.
Figure
9
.
V
el
oc
it
y pr
ofi
le
of
a car acc
ordi
ng to v(
x)
4.
RESU
LT
A
N
D ANALY
SIS
To
a
pply
the
find
i
ngs
f
r
om
ou
r
stu
dy
ou
tl
in
ed
a
bove
,
a
t
ra
ff
ic
footage
th
at
is
acce
ssible
to
t
he
public
on
Y
ouT
ub
e
was
us
e
d
[25].
The
s
pecific
requirem
ents
f
or
the
c
am
era’s
posit
io
n,
a
ng
le
,
et
c.
a
re
no
t
m
e
t
perfect
ly
in
a
ny
publicl
y
avai
la
ble
vi
deo
f
oota
ge
so
the
vi
deo
is
cr
oppe
d
an
d
m
od
ifie
d
to
sim
ulate
a
near
ly
feasible
in
pu
t
t
o
the
al
gorith
m
.
The
base
vid
eo
[25]
c
hose
n
was
s
uc
h
tha
t
it
of
fer
e
d
a
fron
t
-
to
p
view
t
ow
a
r
ds
incom
ing
veh
i
cl
es
wh
ic
h
yi
el
ds
the
best
resu
lt
s
f
or
an
obj
ect
detect
ion
al
gorithm
with
le
ast
ov
erla
ps.
The
f
arthe
st
po
int
in
the
vid
e
o
is
too
far
a
wa
y
fo
r
the
obj
ec
t
detect
ion
syst
e
m
to
detect
anyt
hin
g
that
s
m
al
l
as
sh
ow
n
in
Fig
ure
10,
an
d
the
re
is
ano
the
r
la
ne
of
tra
ff
ic
in
the
fr
am
e
o
f
the
ca
m
era
wh
ic
h
will
produce
inaccu
rate res
ul
ts, the
vid
e
o w
as cr
oppe
d
to
a
sp
eci
fic
pa
rt a
s sho
wn in
Fig
ur
e
11.
Evaluation Warning : The document was created with Spire.PDF for Python.