TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 86
5~8
7
4
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.362
865
Re
cei
v
ed Se
ptem
ber 1, 2014; Re
vi
sed
No
vem
ber 5,
2014; Accept
ed No
vem
b
e
r
20, 2014
Large Crowd Count Ba
sed on Improved SURF
Algorithm
Haining Zha
ng*
1
, Huanbo Gao
2
Schoo
l of Elect
r
onic Informati
on Eng
i
n
eeri
n
g
,
Xi'
an T
e
chnol
og
y Un
iversit
y
2 Xuefuzh
o
n
g
l
u
Avenu
e,W
e
i
y
ang D
i
strict, Xi'
an 71
00
21, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhn19
64@
16
3.com
1
, 1198
3
096
86@
qq.co
m
2
A
b
st
r
a
ct
T
h
is pap
er us
es an an
alysis
of Speede
d u
p
Ro
b
u
st F
eature (SURF
),
based o
n
the method o
f
Lin
ear Interpo
l
ation for ca
mer
a
distortio
n
cali
bratio
n,
for hig
h
-de
n
sity crow
d counti
ng. T
he eig
enva
l
u
e
s are
built o
n
the Gray Level C
o
-
o
ccurre
nce Matrix (GLC
M) features a
nd the SURF
features. T
h
roug
h
the
m
e
tho
d
o
f
l
i
nea
r in
te
rp
ol
a
t
io
n, we
i
g
h
t
va
l
ues a
r
e i
n
t
e
rpo
l
a
t
ed to
red
u
ce
the err
o
r, w
h
ic
h is
ca
used
b
y
camera d
i
storti
on ca
libr
a
tion.
T
he opti
m
i
z
e
d
crow
d
’
s fe
ature
vector can b
e
got t
hen. T
h
ro
u
gh the
method
of
supp
ort vector
regr
essio
n
, th
e crow
d
’
s
n
u
m
ber c
a
n
be
for
e
cast by
trai
ni
ng
mode
l. T
h
e
exp
e
ri
me
nt re
sult
show
s that the meth
od of this
pap
er has
a h
i
gher acc
u
racy
than the pr
evio
us met
hods.
Ke
y
w
ords
: crowd count, SURF, GLCM, pers
pective-correct
, SVR
1. Introduc
tion
With the incre
a
se of the wo
rld pop
ulation
,
unfortunate accide
nts
in
p
ublic pl
aces
caused
by high-d
e
n
s
i
t
y crowd occur freq
uently in re
cent
years. At the sam
e
time, the vi
deo surveilla
nce
system
s are
ubiquito
us [1]
.
If we make
use of
the ex
isting resources,
these inte
lligent syste
m
s
can
help
us effectively fore
wa
rn a
n
d
avoid di
sa
ster event
s. Compa
r
ed
wit
h
the tra
d
itio
nal
approa
ch, th
e intelligent
system of
counting
and
den
sity estim
a
tion can al
so imp
r
ove t
he
utilization
rat
e
of publi
c
facilities, a
n
d
arra
n
ge t
he allo
catio
n
of manp
ower a
nd mate
rial
r
e
sour
ces
eff
e
c
t
ively.
The alg
o
rith
m of cro
w
d
counting
can
be divided
int
o
two catego
ries: di
re
ct a
nd indi
re
ct
mean
s. Th
e
dire
ct
way util
ize
s
p
eopl
e’s
cha
r
a
c
teri
st
ics di
re
ctly, su
ch a
s
colo
r,
sh
ape, et
c, to
g
e
t
the cro
w
d’
s
numbe
r. The
people’
s he
ad and fa
ce
and som
e
other cha
r
a
c
teristi
cs
can
be
sele
cted
a
s
the
statistical
feature
vect
or. Thi
s
m
e
thod i
s
u
s
u
a
ll
y very co
mpl
e
x, and i
s
m
o
re
suitabl
e for
monitori
ng lo
w den
sity po
pulation
s
. Th
e major
re
se
arch metho
d
of countin
g h
i
gh-
den
sity cro
w
d
in the world i
s
the indirect
way
[2]. With this method, the numb
e
r
can be obtain
e
d
by the metho
d
of reg
r
essio
n
throu
gh extractin
g
the whole cro
w
d’
s
feature
s
[3]. But the statistical
pre
c
isi
on
of this m
e
thod
i
s
curre
n
tly n
o
t accu
rate
e
noug
h. The
method
still
need
s fu
rthe
r in
-
depth re
se
arch.
This pap
er u
s
e
s
a
stati
s
tical re
gressio
n
method. Fi
rst, the cro
w
d’
s
foreg
r
ou
nd i
m
age i
s
extracted
fro
m
the in
put
image.
The
n
, the G
L
CM features
and S
U
RF f
eature
of
crowd’
s
foreg
r
ou
nd i
m
age a
r
e extracte
d [4]-[7]. Thoug
h the method of lin
ear inte
rpol
ation, weig
ht value
s
that cau
s
e
d
by came
ra
di
stortion
cali
bration, are
interpol
ated to
redu
ce the
error. Th
roug
h t
h
e
method of suppo
rt vector regres
sio
n
, the crowd’s
numbe
r can
be
finally foreca
st by trai
ning
model. Figu
re
1 sho
w
s the
block dia
g
ra
m of whole al
gorithm.
2. SURF Fe
a
t
ure Ex
trac
tion
The re
se
arch
object of thi
s
pap
er i
s
hi
gh-d
e
n
s
ity crowd. SURF
algorith
m
is
use
d
to
descri
be the
cha
r
a
c
teri
stics of popul
atio
n.
In 2006, He
rbert Bay pro
posed a mo
re pra
c
ti
cal fe
ature dete
c
ti
on algo
rithm
of SURF.
SURF i
s
a lo
cal featu
r
e p
o
int detecto
r
with hi
gh
rob
u
stne
ss, and
the operatin
g spe
ed of this
algorith
m
i
s
highe
r. Be
ca
use
of its g
o
od inva
ri
an
ce of
scale
transfo
rmatio
n
and
pe
rspe
ctive
transfo
rmatio
n, it has beco
m
e an impo
rtant f
eature ex
traction al
go
ri
thm in many ways.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 865
– 874
866
Figure 1. The
block diag
ra
m of whole al
gorithm
The processi
ng of crowd’
s SURF f
eatu
r
e extraction i
s
as follo
ws:
(1)
Run
n
ing the
global
scanni
ng of the origi
nal
image, an
d obtainin
g
the integral im
a
ge [8].
In Figure 2, to any point
,
ij
in the imag
e, its value of
,
ii
i
j
in the integ
r
al
image i
s
the sum
of all points’ g
r
ay
value on the
di
ago
nal. And
the diago
nal
is from the
po
int
,
ij
to the
top left vertex of original im
age. The valu
e
,
ii
i
j
is
as
follows
:
j
j
i
i
j
i
p
j
i
ii
,
)
,
(
)
,
(
(1)
The su
m of all pixels’ gray
value in any wi
nd
ow W
ca
n be obtaine
d by the value of the
four point
s
11
,
ij
,
22
,
ij
,
33
,
ij
,
44
,
ij
in the integral image.
44
3
3
22
1
1
,,,
,
W
I
i
ii
j
i
ii
j
i
ii
j
i
ii
j
(2)
Figure 2. The
calculation of
integral ima
g
e
and t
he su
m of all pixels’ gray value in wind
ow
W
Whe
n
all the
points’ g
r
ay
value in o
r
igi
nal imag
e is
1, the value
of
,
ii
i
j
in integral
image re
presents the re
ct
angul
ar area
from the poin
t
,
ij
to the top l
e
ft vertex.
w
I
mean
s the
area of wi
ndo
w W.
(2)
The
extrem
e points
of scal
e-spa
c
e can be
got
throug
h He
ssian ma
trix approxim
ation. The
s
e
extreme poi
nts are th
e featur
e poi
nts of what we nee
d.
The He
ssian
matrix
x,
H
of
point
x,
y
in the image I is define
d
as follo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Larg
e
Cro
w
d
Cou
n
t Based
on Im
prove
d
SURF Algo
rit
h
m
(Hainin
g
Zhang
)
867
,,
,
,,
xx
xy
xy
yy
Lx
L
x
Hx
Lx
L
x
(3)
2
de
t
,
,
,
xx
yy
x
y
HL
x
L
x
L
x
(4)
means sc
ale.
(,
)
xx
Lx
i
s
the
L
apla
c
ian
of
Gau
ssi
an
of the im
age.
It is t
h
e
convol
ution o
f
the Ga
ussia
n
second
ord
e
r d
e
rivative
22
()
/
g
x
with th
e ima
ge.
(,
)
xx
Lx
a
nd
(,
)
yy
Lx
have
the
sa
me me
anin
g
.
Whe
n
the
val
ue of
det
(
)
H
is
g
r
eate
r
th
an
z
e
r
o
, if
(,
)
xx
Lx
is
greate
r
than
zero, the poin
t
x
is the local
minimum poi
n
t; if
(,
)
xx
Lx
is less t
han zero, the point
x
is the local
minimum p
o
int. The feature poi
nt can b
e
judg
ed throu
gh
comp
uting th
e
determi
nant o
f
each pixel in
the image [9].
For th
e
conv
enien
ce
of a
p
p
licatio
ns,
He
rbe
r
t Bay p
r
o
posed
app
rox
i
mating
se
co
nd o
r
de
r
derivatives
wi
th box filters. Us
in
g the b
o
x filters ope
ration to re
pl
ace the
conv
olution op
era
t
ion
L
. The differen
t
scale
s of scale sp
ace are
formed by expandi
ng the size of the box
.
Usi
ng
xx
D
,
y
y
D
and
x
y
D
to repla
c
e
L
, th
en the determinant is si
m
p
lified as:
2
de
t
xx
y
y
xy
HD
D
w
D
(5)
The wei
ght
w
chang
es
with the ch
ang
e of scale.
The featu
r
e
points
sho
u
ld
be furthe
r confirme
d afte
r prelimina
r
y testing. In o
r
de
r to
verify the extreme poi
nts i
n
t
he scal
e
-space, each
sampling p
o
int
should b
e
compa
r
ed with
all
its adja
c
ent p
o
ints. In othe
r wo
rd
s, ea
ch point
is
co
mpared with
26 point
s, wh
ich me
an
s those
18 point
s in the adja
c
e
n
t scale
-
spa
c
e a
nd 8 point
s in
the same im
age. If the point is gre
a
ter
or
less than the
s
e 26 p
o
ints,
it
is the final feature p
o
int.
(3)
The p
r
inci
pal
orientatio
n of
each feature
point is d
e
te
rmine
d
. After that, the 64-d
i
mensi
onal
c
h
arac
teriz
a
tion vec
t
or is
formed.
In this pa
pe
r, the first p
r
o
c
e
ssi
ng of S
URF
feature
points
extra
c
tion is th
at th
e bina
ry
foreg
r
ou
nd i
m
age i
s
extra
c
ted fro
m
the
input
image
by the backg
round
subt
ra
ction metho
d
a
nd
the sliding a
v
erage m
e
th
od. Then the
SURF feat
u
r
e is extra
c
te
d from the binary fore
gro
u
nd
image.
Com
p
ared
with
the
SURF featu
r
e extrac
tio
n
of the
overall imag
e, th
e SURF fe
ature
extraction
of
binary fo
re
ground
imag
e
redu
ce
s
the
computation
complexity. Figure
3
sh
ows the
result of SURF feature extraction.
Figure 3. SURF feature extracti
on of bin
a
ry foreg
r
ou
n
d
image
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 865
– 874
868
The
white
p
a
rt of
bina
ry
fore
gro
und
i
m
age
is the
moving
regi
on. Acco
rdin
g to th
e
prin
ciple
of t
he SURF
alg
o
rithm, mo
st
of the
SURF
feature
point
s a
r
e i
n
the
motion a
r
e
a
,
but
there
will
be
a fe
w fe
ature poi
nts
aro
und
regi
on
[10]. As
sh
own in
Figu
re
3
,
the n
u
mbe
r
of
feature p
o
ints cann
ot effect
ively reflect the ch
arac
t
e
ri
st
ic
s of
cr
ow
d
.
S
o
the feature poi
nts in th
e
non-i
n
terest region
sho
u
ld
be reje
cted.
This se
le
cting process o
n
ly need
s to scan all feat
ure
points, which are di
stingui
shed
by their value of pixel.
1
,
,
255
,
0,
,
0
ix
y
su
r
f
x
y
ix
y
(6)
,
sur
f
x
y
is the discri
minant of
poi
nt
,
x
y
. When the value of
,
sur
f
x
y
i
s
1
,
t
h
i
s
point will be
kept; when it is 0, this point will be re
mov
ed.
,
ix
y
is the pixel values of this poi
nt.
Figure 4. After reje
cting fea
t
ure point
s in non interest region
In Figure 4, we can
see tha
t
the kept SURF
feature po
ints ca
n refle
c
t the cha
r
a
c
t
e
risti
cs
of cro
w
d reall
y
and effectively.
3. Eigen
v
ector Cons
tru
c
tion
w
i
th SURF Fea
t
ur
e a
nd GLCM F
e
ature
SURF fe
atu
r
e ha
s g
o
o
d
invaria
n
ce
of scale
transfo
rmatio
n and p
e
rspective
transfo
rmatio
n, and
can
re
flect the
cha
r
acteri
stic
s
of cro
w
d. But fo
r large p
opul
ations, a
nd
whe
n
there i
s
a d
ense cove
rin
g
, the SURF
feature ca
nn
o
t
r
e
flec
t th
e
c
h
ar
ac
te
r
i
stic
s
ve
r
y
we
ll.
Becau
s
e th
e
GLCM feat
ure
s
can effectively
overcome th
e occlu
s
ion p
r
obl
em, this pap
er
prop
oses co
mbining
the
GLCM
featu
r
e with S
U
RF
feature.
T
he eigenve
c
tor
i
s
comp
osed of
four
uncorrelate
d GLCM featu
r
e vector
s (en
e
rgy, entropy
, contra
st, co
rrel
a
tion) a
n
d
the numbers of
SURF featu
r
e
points [11].
The four u
n
correlated G
L
CM
feature vectors a
r
e:
(1) Energy: A statistic reflect
s
the con
s
iste
ncy.
2
11
00
,,
NN
ij
AS
M
p
i
j
d
(7)
Energy refle
c
ts the level
of texture
co
a
r
sene
ss and the unif
o
rmity level of gray
distrib
u
tion. Whe
n
the texture is
coa
r
se
, the
energy is high. Othe
rwise, the ene
rgy is low.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Larg
e
Cro
w
d
Cou
n
t Based
on Im
prove
d
SURF Algo
rit
h
m
(Hainin
g
Zhang
)
869
(2) Contra
st: A statistic of cont
rast.
11
2
00
,,
NN
ij
Co
n
i
j
p
i
j
d
(8)
The
cont
ra
st refle
c
ts th
e
sh
arp
n
e
s
s
of the ima
g
e
.
Whe
n
the
texture i
s
co
arse, the
contrast i
s
small. Otherwi
se, the co
ntra
st is big.
(3) Entropy: A param
eter calculating the
ra
ndomn
e
ss di
stributio
n of gray-level
d
.
11
00
,,
l
o
g
,
,
NN
ij
E
n
t
p
ij
d
p
ij
d
(9)
Entropy indicates the leve
l of non-unif
o
rmity
texture or the com
p
lexity of
the image.
Whe
n
the texture is
coa
r
se
, the entropy is
sm
all. Othe
rwi
s
e, the ent
ropy is la
rge.
(4) Homog
eneit
y
: A correlatio
n
statistic of g
r
ay value.
11
2
00
,,
1
LL
H
ij
pi
j
d
S
ij
(10
)
The h
o
mog
e
neity refle
c
ts
the direction
of t
he texture
,
and
sho
w
s the si
milarity
degree of
rows o
r
colu
mns. Th
e differen
c
e
of pixel value
s
bet
wee
n
elem
en
ts is bi
gge
r, the hom
oge
n
e
ity
value is sm
all
e
r.
Thro
ugh the
above theo
ry, 6-dime
nsi
o
n
a
l featur
e vector is form
ed i
n
this pap
er,
and the
SURF featu
r
e
is the main chara
c
te
risti
c
:
(
num
surf
, s
,
fe
ature
entropy
, fe
ature
energy
, fe
ature
contrast
, feature
hom
ogen
ei
ty
).
num
surf
is the numb
e
r of
SURF
point
s. s i
s
the
area
of moving pe
ople in
binary
foreg
r
ou
nd i
m
age.
featu
r
e
entropy
is the entro
py of
GLCM
matrix
.
feature
energy
is the
ene
rg
y of
GLCM matrix
.
feature
contrast
is the
co
ntra
st of GL
CM
matrix.
feature
homogen
eity
is the ho
mog
e
n
e
ity
of GLCM mat
r
ix
4. Linear Interpolation
Weights for Ca
mera Distortion Calibrati
on
Came
ra dist
ortion calib
ra
tion
is cau
s
e
d
by
the in
crea
sing
dista
n
ce
between
movin
g
obje
c
tive and
the camera. As we all kn
ow, the are
a
of people ne
ar ca
mera is
bigge
r than the
area
of o
ne f
a
r a
w
ay from
cam
e
ra. In o
r
de
r to
red
u
ce the influ
e
n
c
e cau
s
ed
by the lo
sing
de
p
t
h
informatio
n o
f
image, this
pape
r ad
opts the metho
d
of linear i
n
terpolation
weig
hts for
cam
e
ra
distortio
n
cali
bration.
Thi
s
method
ha
s str
ong
ada
p
t
ability and
h
i
gh real
-time. And in
a
c
tu
al
appli
c
ation, the re
sea
r
che
r
s do n
o
t
nee
d measureme
n
t of the environment.
Figure 5. The
theory of linear interpolatio
n weig
hts
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 865
– 874
870
In Figure 5,
the imag
e is divided i
n
to
four
regi
ona
l grid
s. Th
e
width a
nd
h
e
ight of
obje
c
t’s mini
mum en
clo
s
ing re
ctangl
e (MER) ca
n be
got in the entran
c
e a
n
d
exit of each grid.
The
wei
ghts
of ea
ch
gri
d
can
be
obtai
ned
by th
e
a
r
ea
chang
e
rate of th
e M
E
R. As sho
w
n in
Figure 5, the area
cha
nge
rate of grid
a,
b, c and d
is
as follows:
22
11
hw
k
hw
(11
)
11
,
hw
is the width a
nd heig
h
t of obj
ect’
s MER i
n
the entra
nce.
22
,
hw
is the widt
h and
height of obje
c
t’s MER in t
he exit.
Figure 6
sho
w
s a
video
seque
nce of
single
peo
ple
wal
k
ing
in P
E
TS video
seque
nce
s
[12]. This pap
er sepa
rate
s the monito
r sp
ace into
4 pa
rts in ea
ch vid
eo frame. T
h
e purpo
se is t
o
improve the
accuracy of
came
ra di
sto
r
tion
calibration. It should
be noted th
e accu
ra
cy of
calib
ration i
s
more a
c
curate whe
n
the more pa
rts
sep
a
rated. But the more
weig
h
t
s interpol
ate
d
,
the pro
c
e
ssi
n
g
of determini
ng t
hese wei
ghts is m
o
re
compl
e
x.
(a) The 9th frame
(b) The 72th frame
Figure 6. PETS2009
singl
e peopl
e wal
k
ing video seq
uen
ce
For fo
ur
part
i
tions, the
r
e
are fo
ur i
n
te
rpolatio
n
wei
ghts. As sho
w
n in
Figu
re
6, the
numbe
r of S
URF
feature
points i
s
diffe
rent in di
ffe
re
nt frame
s
. Th
e 4 weight
s can be
cal
c
ul
a
t
ed
by the method of regressio
n
.
(a) Before correc
tion
(b) After c
o
rrec
t
ion
Figure 7. The
corre
c
tion effect of SURF
numbe
r
0
10
20
30
40
50
60
70
80
90
100
0
5
10
15
20
25
30
35
40
45
50
fr
a
m
e
ac
r
e
ag
e
0
10
20
30
40
50
60
70
80
90
100
0
5
10
15
20
25
30
35
40
45
50
fr
a
m
e
a
c
r
eag
e
28 points
16 points
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Larg
e
Cro
w
d
Cou
n
t Based
on Im
prove
d
SURF Algo
rit
h
m
(Hainin
g
Zhang
)
871
Figure
7 sho
w
s obviou
s
d
i
fferences of
SURF
numb
e
r befo
r
e
an
d after corre
c
tion. The
absci
ssa rep
r
ese
n
ts the n
u
mbe
r
of frames, or
dinat
e rep
r
e
s
ent
s SURF n
u
mb
ers. Fig
u
re 7
(a)
sho
w
s that th
e SURF
num
ber re
du
ce
s
grad
ually a
s
the p
ede
stria
n
wal
k
s a
w
ay f
r
om th
e
cam
e
ra
grad
ually. Fi
gure
7
(b
)
sho
w
s that t
he SURF
n
u
mbe
r
remai
n
s i
n
a
sta
b
le rang
e af
ter
interpol
ating the four weigh
t
s for co
rrecti
on.
Figure 8
sh
o
w
s differe
nce
of fore
gro
u
n
d
ar
ea before
and after co
rrectio
n
.
Th
e absci
ssa
rep
r
e
s
ent
s th
e num
be
r of
frame
s
, o
r
din
a
te re
present
s the
fore
gro
und
are
a
[13
]. Figure
8
(a)
sho
w
s that t
he foregroun
d area
redu
ces g
r
a
dually
as the
pe
d
e
stria
n
walks away from
the
came
ra g
r
a
d
ually. Figure
8 (b)
sh
ows t
hat the
foreg
r
ound a
r
e
a
re
mains i
n
a st
able rang
e after
interpol
ating the four weigh
t
s for co
rrecti
on.
Figure 7 a
nd
Figure 8
sho
w
that the m
e
thod of
line
a
r interpol
ation weig
hts can solve
the
probl
em of ca
mera di
stortio
n
rapidly an
d effectively.
(a) Before correc
tion
(b) After c
o
rrec
tion
Figure 8. The
corre
c
tion effect of foreg
r
o
und area
5. The Tes
t
Re
sults,Analy
s
is and Comp
arison
5.1 The Tes
t
Resul
t
s and
Analy
s
is
This expe
rim
ent u
s
e
s
Mi
crosoft Visu
al
C++
6.0
as software
d
e
vel
opment
environment
and O
pen
CV
1.0 as imag
e
pro
c
e
ssi
ng li
bra
r
y in t
he
o
peratin
g
syst
em of Wi
ndo
ws XP.
Hard
ware
experim
ental
platform i
s
a PC ma
chi
ne, and th
e
PC memo
ry
is 2G. T
h
is
experim
ent u
s
e
s
Matlab7.0 a
s
the analysi
s
tool for the co
nclu
sio
n
and
analysi
s
.
The
n
u
mb
er of
crowd can
be
e
s
timated
after
ε
-SVR training. T
he t
r
aining
mod
e
l
can
be
got throu
gh trainin
g
the correctio
n
feature ve
ct
ors. This p
ape
r tests three vid
eo se
que
nce
s
in
the PETS200
9 video
library [12]. Figure
9
sho
w
s th
e
c
u
r
v
e o
f
th
e re
a
l
c
r
ow
d
’
s
nu
mb
er
, th
e
tes
t
cro
w
d’
s nu
m
ber with
out correctio
n
and
t
he test crowd’s num
be
r after co
rrection.
(a) Vide
o 1
0
10
20
30
40
50
60
70
80
90
100
11
0
0
200
400
600
800
1
000
1
200
1
400
1
600
1
800
fr
a
m
e
ac
r
e
ag
e
0
10
20
30
40
50
60
70
80
90
10
0
11
0
0
20
0
40
0
60
0
80
0
10
00
12
00
14
00
16
00
18
00
fr
a
m
e
ac
r
eage
0
20
40
60
80
10
0
120
140
160
180
20
0
0
5
10
15
20
25
30
35
40
fr
a
m
e
c
o
unt
real
nu
m
ber
t
e
s
t
num
ber
af
t
e
r
c
o
r
r
ec
t
i
on
t
e
s
t
num
ber
w
i
t
hout
c
o
r
r
ec
t
i
on
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 865
– 874
872
(b) Vide
o 2
(c) Video 3
Figure 9. The
test result an
d analysi
s
The thre
e in
dexes of a
n
a
l
yzing the te
st result a
r
e: MAE (Mean
Absolute E
r
ror), M
R
E
(Mean Rel
a
tive Error) [14], and
MAXE (Maximum Absolute Error).
0
1
1
n
i
M
AE
N
i
N
i
n
(12
)
0
1
0
1
n
i
Ni
N
i
MR
E
nN
i
(13
)
n
is the frames of video.
N
i
is the test numb
e
r of frame
i
.
N
i
is the real nu
mber
of frame
i
.
Experimental
results befo
r
e
and after correctio
n
s of thi
s
pap
er a
r
e shown in Tabl
e 1.
Table 1. The
test re
sults a
nalysi
s
before corre
c
tion
Name
Non-cor
r
ection
After-c
or
rec
t
i
on
MAE MRE
MAXE
MAE
MRE
MAXE
Video1 8.465
32.2%
14
0.951
5%
3
Video2 6.802
51.9%
10
1.518
14.8%
3
Video3 10.654
50.1%
22
1.788
11.5%
7
0
20
40
60
80
100
12
0
140
16
0
180
200
0
5
10
15
20
25
30
35
40
fr
a
m
e
co
u
n
t
r
eal
num
ber
t
e
s
t
num
ber af
t
e
r
c
o
r
r
e
c
t
i
o
n
t
e
s
t
num
ber w
i
t
h
o
u
t
c
o
rr
ec
t
i
on
0
20
40
60
80
10
0
120
14
0
16
0
18
0
20
0
0
5
10
15
20
25
30
35
40
45
fr
a
m
e
co
u
n
t
real
num
ber
t
e
s
t
num
b
e
r af
t
e
r c
o
rrec
t
i
on
t
e
s
t
num
b
e
r w
i
t
h
out
c
o
rrec
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Larg
e
Cro
w
d
Cou
n
t Based
on Im
prove
d
SURF Algo
rit
h
m
(Hainin
g
Zhang
)
873
Comp
ared
wi
th the no
n-co
rre
cti
on
meth
od, the a
c
curacy after
co
rrection i
s
m
u
ch high
er.
It prove
s
tha
t
the meth
od
of line
a
r int
e
rpol
at
ion
weights can
solve the
pro
b
lem of
cam
e
ra
distortio
n
ra
pi
dly and effect
ively.
Table 2. The
results of GL
CM metho
d
Name
GLCM
SURF
MAE MRE
MAXE
MAE
MRE
MAXE
Video1 3.027
15.3%
7
1.02
5.1%
3
Video2 2.508
21.9%
9
1.21
9.8%
4
Video3 7.402
38.2%
15
4.47
19.4%
9
In Table 2, the GLCM m
e
thod is the
crow
d
cou
n
ting
method ba
sed on GL
CM
feature
s
.
The cro
w
d fe
ature eig
enve
c
tor of this m
e
thod is
o
n
ly made up of t
he 4
GL
CM feature ve
ctors.
The S
URF
m
e
thod i
s
th
e
counting
meth
od b
a
sed
on
SURF
alg
o
rit
h
m. The
mai
n
vecto
r
of
crowd
feature eig
e
n
v
ector i
s
SURF numbe
r.
The results
a
nalysi
s
sho
w
s that thi
s
m
e
thod
can
estimate the pe
ople
s
’ num
be
r in te
st
regio
n
ra
pidl
y and accu
ra
tely. Compared with t
he
GLCM m
e
tho
d
, the accura
cy of this pa
per
increa
se
s ob
viously. It shows t
hat SURF num
ber i
s
an impo
rta
n
t feature vector of co
unti
ng
cro
w
d.
Comp
ared
with the
SURF m
e
tho
d
, the ac
cura
cy of video 3 increa
se
s ob
viously. It shows
that the GLCM feature vectors
can effe
ctively overco
me the occlu
s
ion p
r
obl
em.
5.2 Compar
e
d
With The P
i
xel Statistic
Feature Me
thod
In many
ca
ses the
pixel
s
statisti
c feat
ur
e
ca
n de
scribe th
e po
p
u
lation
cha
r
a
c
teri
stics
effectively,
mainly includ
ing foreg
r
ou
nd f
eature a
nd edge fea
t
ure [15]. By introdu
cing
the
perspe
c
tive correctio
n
parameter, it can
calc
ul
ate the weig
ht ratio
calculation p
a
ram
e
ters.
i
i
i
i
wi
r
w
(14
)
r
is the foregro
und pixels o
r
edge pixel
s
o
f
the pixels statistic feature
,
i
w
is the impul
se
respon
se fun
c
tion,
i
is the perspe
c
tive correctio
n
parameter
s
of pixels
. When it is
foreground
or edg
e pixel, the value is 1, otherwi
se t
he value is 0.
Thro
ugh the
comp
ari
s
o
n
of the SURF algorithm
an
d the pixels sta
t
istic feature
metho
d
use
d
in video
3, the result
s are sh
own in
the Figure 1
0
.
Figure 10. Th
e result of co
mpari
s
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 865
– 874
874
It can
be
se
e
n
that the
SURF
algo
rithm
method
i
s
bet
ter than
the
p
i
xels
statistic
feature
method. Ove
r
all, the devi
a
tion of expe
riment resu
lts of the pixe
ls stati
s
tic fe
ature m
e
thod
i
s
bigge
r, while
the SURF
method
which kee
p
s t
he
experim
ent result
s cu
rve
arou
nd the a
c
tual
numbe
r curve
all the way is more p
r
eci
s
e
l
y.
6. Conclusio
n
Acco
rdi
ng to
the
pro
b
le
m of
cou
n
tin
g
hig
h
-den
sity crowd, thi
s
p
ape
r
pro
poses an
improve
d
cro
w
d counting
method ba
se
d on SURF
algorithm a
nd GLCM al
go
rithm. Experim
ental
study find
s t
hat, the lin
ea
r inte
rpol
atio
n weight
s
co
rre
ct
ion
met
h
od i
s
a
sim
p
l
e
an
d ef
f
e
ct
iv
e
method fo
r came
ra
disto
r
tion calibration. This
alg
o
rithm h
a
s
strong a
dapta
b
ility, and can
accurately e
s
timate the n
u
mbe
r
of p
e
ople in
ea
ch
frame
with t
he ave
r
ag
e
error l
e
ss th
an 2
peopl
e per frame. With th
e variety and
compl
e
xity
of the resea
r
ch
environm
ent, the method
still
need
s furthe
r in-depth
re
se
arch.
Referen
ces
[1]
Conte
D, F
oggi
a P, Percan
nel
l
a
G, et al. A Method
for
Cou
n
t
ing Movi
ng P
e
opl
e in V
i
de
o S
u
rveil
l
a
n
ce
Vide
os.
EURA
SIP Journal o
n
Advances i
n
S
i
gn
al Process
i
n
g
.
2010; 5(
1): 1-8.
[2]
Ya Hua
ng, Su
Hang, Z
hen
g
Shib
ao. Lar
ge-
scale Cro
w
d
D
ensit
y
Estimati
on.
Vide
o App
l
icatio
n and
Project
. 201
0; 34(5): 11
3-1
1
6
.
[3]
Antoni B, Ch
a
n
, Nuno V
a
sc
once
l
os. Co
un
ting Peo
p
l
e
W
i
th Lo
w
-
L
e
v
e
l F
eatures
a
nd Ba
yes
i
a
n
Regr
essio
n
.
IEEE Transactio
n
s on Image Pr
ocessi
ng
. 20
12
; 21(4): 216
0-2
177.
[4]
W
ang Yal
i
n. Rese
arch on
Algorit
hm of Cro
w
d D
ens
ity Estimatio
n
Based o
n
Gra
y
L
e
ve
l Co-
occurre
nce Ma
trix.
Master'
s
thesis
. Xi
’a
n: Xi
’a
n Univ
ersit
y
of
T
e
chnolog
y
an
d Scienc
e. 201
3.
[5]
Ba
y
Herb
ert, Ess A,
T
u
y
t
e
l
a
a
rs
T
,
et al. Speede
d-up
Rob
u
s
t F
eatures (SURF
).
Comput
er Visio
n
an
d
Imag
e Un
derst
and
ing
. 2
008;
110(
3): 346-
35
9.
[6]
Nan Ge
ng, D
o
ngji
a
n
He, Y
a
n
s
hua
ng S
o
n
g
.
Camera
Imag
e
Mosaic
ing
Ba
sed o
n
a
n
Opti
mized S
URF
Algorit
hm.
T
E
LKOMNIKA: Indones
ian J
ourn
a
l
of Electrica
l
Engi
neer
in
g
.
2012; 10(
8): 218
3-21
93.
[7]
W
ang W
,
Li
W
H, W
ang C
X, et
al. A N
o
vel W
a
term
ar
king A
l
gor
ithm
base
d
o
n
SU
RF
and S
V
D.
T
E
LKOMNIKA: Indones
ia
n Jo
urna
l of Electri
c
al Eng
i
ne
eri
n
g
. 2013; 1
1
(3): 156
0-15
67.
[8]
K. Zhang,
T
.
F. Xu, P. W
ang, L. Feng. Rea
l
-time
full-fram
e digit
a
l ima
g
e
stabiliz
atio
n s
y
stem B
y
SURF.
Optics
and Prec
isio
n Engi
neer
in
g
. 2011; 19(
8): 196
4-19
72.
[9]
R
y
an D, D
e
n
m
an S, F
ooke
s
C, et al.
Cro
w
d Counti
ng
Using Gro
up T
r
ackin
g
an
d L
o
cal F
e
atures
.
2010 7th IEEE
Internationa
l
Conference on Advanc
ed
V
i
deo and Signal
B
a
sed S
u
rveillance. Boston.
201
0: 218-
224.
[10]
X. L
i
u, B. Da
i, H.G He.
Rea
l
-time o
b
ject s
e
g
m
e
n
tatio
n
for v
i
sual
ob
ject d
e
tection
in
dyna
mic sc
en
es
.
Soft Computin
g and Patter
n
Reco
gniti
on (S
oCPa
R). Dali
a
n
. 2011: 4
23-4
28.
[11]
Sun W
enc
ha
n
g
, Song
Jia
n
sh
e, Yang
Men
g
, et al. C
o
rner
Detectio
n Al
go
rithm Base
d o
n
Entrop
y
an
d
Uniq
ue
ness.
Journ
a
l of Co
mputer App
lic
ations
. 200
9; 29(
2): 225-2
27.
[12]
Cho
udri, S., Ferr
y
m
an, JM.,
Badi
i, A.
Robu
st backgrou
n
d
mod
e
l for pix
e
l bas
ed pe
op
le cou
n
tin
g
usin
g a
sin
g
l
e
u
n
cal
i
brat
ed
ca
mer
a
. Proceedings
of T
w
elfth IEEE
I
n
ternational W
o
rkshop on
Performanc
e Evalu
a
tion of
T
r
acking and Surveil
l
a
n
ce (
PET
S-Winter). IEEE Computer Science.
Miami. 20
09: 1
-
8.
[13]
Ya Li
ho
u, Pa
n
g
G KH. Pe
op
le C
o
u
n
ting
a
nd H
u
ma
n D
e
tection
in
a C
hall
e
n
g
in
g S
i
tu
ation.
IEEE
T
r
ansactio
n
s o
n
Systems
. 20
11; 11(4): 2
4
-3
3.
[14]
K. Zhang,
T
.
F. Xu, P. W
ang, L. Feng. Rea
l
-time
full-fram
e digit
a
l ima
g
e
stabiliz
atio
n s
y
stem B
y
SURF.
Optics
and Prec
isio
n Engi
neer
in
g
. 2011; 19(
8): 196
4-19
72.
[15]
Yang
Hu
a, Su
Han
g
, Z
h
e
n
g
Shib
ao.
Lar
g
e
-scal
e
Cr
o
w
e
d
De
nsit
y Esti
mation.
V
i
de
o Engi
neer
in
g
.
201
0; 34(5): 11
3-11
6.
Evaluation Warning : The document was created with Spire.PDF for Python.