Indonesian J
ournal of Ele
c
trical Engin
eering and
Computer Sci
e
nce
Vol. 1, No. 3,
March 20
16, pp. 635 ~ 6
4
6
DOI: 10.115
9
1
/ijeecs.v1.i3.pp63
5-6
4
6
635
Re
cei
v
ed O
c
t
ober 3
0
, 201
5; Revi
se
d Ja
nuar
y 29, 20
1
6
; Acce
pted
February 9, 2
016
Real-Time Automatic Obstacle Detection and Alert
System for Driver Assistance on Indian Roads
Sachin Shar
ma
1
, Dharme
sh Shah
2
Electron
ics & Commun
i
cati
o
n
Dep
a
rtment, Gu
jarat T
e
chnolo
g
ica
l
Univ
er
sit
y
, Ahme
dab
ad
India
e-mail: sh
arma
.f@gmail.com, djsh
ah9
9@
gmail.com
A
b
st
r
a
ct
Roa
d
cras
hes
hav
e
bee
n
a
maj
o
r pr
obl
e
m
in
Indi
a in recent
ti
mes. T
he
occ
u
rrenc
es
h
a
v
e
incre
a
sed
cons
ider
ably
ow
ing
to the
influx
of f
our-w
hee
lers
a
nd tw
o-w
heel
er
s on In
di
an r
o
a
d
s. T
he
nu
mb
e
r
of road traffic collis
io
ns has
also i
n
creas
ed
due to
the a
b
senc
e of aut
omatic hi
ghw
a
y
safety and a
l
er
t
systems o
n
major ro
ads c
o
n
nectin
g
cities
and tow
n
s. Th
e interi
or ro
ad
s conn
ecting v
illa
ges
and to
w
n
s
have b
e
e
n
inst
rumenta
l
in
mu
ltiple
ani
mal-ve
hicle c
o
ll
isio
ns. Althoug
h the fi
gure is n
o
t too larg
e co
mpar
e
d
to other caus
e
s
of road-rel
a
te
d inj
u
ries, they
are sign
ifica
n
t in nu
mb
er. T
he associ
ated n
u
mber of fatalit
ie
s
and
inj
u
ri
es ar
e substa
ntia
l too.
T
hou
gh
nu
mer
ous
efforts have
be
en i
n
p
r
ogress to s
o
lv
e an
d re
duce
t
h
e
nu
mb
er of c
o
ll
isions,
lack
of
practica
l a
ppl
ic
ations
an
d r
e
s
ources
al
on
g
w
i
th qua
lity a
n
a
lytical
d
a
ta (f
or
traini
ng
and
testing) r
e
l
a
ted
to an
i
m
al-v
e
h
icle
co
l
lisi
on has
i
m
pe
ded
any ma
jor bre
a
kthrou
gh in
t
h
e
scenar
io.
I
n
o
u
r curre
nt w
o
rk, w
e
have
pr
opos
ed
an
d d
e
sig
ned
a syst
em bas
ed
on
histogr
a
m
res
e
arch
inclu
d
i
ng ori
ent
ed gra
d
ie
nts a
nd bo
osted ca
scade cl
assi
fie
r
s for automati
c
cow
detection. T
he India
n
cow
has b
e
e
n
the
big
gest o
b
stacl
e
co
mp
are
d
to
other
ani
mals
on In
dia
n
ro
a
d
s. T
he d
i
stan
ce betw
e
e
n
a
cow
and the ve
hicl
e
is calculate
d
p
r
omptin
g an al
ert signa
l to
no
tify the driver for app
lyin
g bra
k
es or und
erta
ke
any si
mil
a
r act
i
on. T
he
meth
od is i
m
p
l
e
m
e
n
ted in Op
enc
v softw
are and tested on v
a
rio
u
s vide
o cl
ip
s
involv
in
g cow
mov
e
me
nts in
vario
u
s scen
a
ri
os. The pr
opos
ed syste
m
has
achi
eved
an
ov
erall
efficie
n
cy
of
80%
in ter
m
s
of cow detection. The
propos
ed system
is a low-cost, highl
y r
e
liable system
which can easily
be i
m
ple
m
ente
d
in
auto
m
ob
ile
s for detecti
on
of cow
or a
n
y o
t
her an
i
m
al
after pro
per tra
i
ni
ng a
nd t
e
sting
o
n
the hig
h
w
a
y.
Ke
y
w
ords
: A
n
i
m
al
detecti
o
n
syste
m
; Ca
scade c
l
assifi
er; His
togr
a
m
of orie
nted
grad
ient; Intel
l
i
ge
nt
hig
h
w
a
y safety; Road acci
de
n
t
s
1. Introduc
tion
Today’s auto
m
obile
de
sig
n
p
r
imarily
d
epen
ds
on
safety mea
s
ures,
se
cu
rity tools an
d
comfo
r
t mech
anism. Th
e a
ppro
a
ch ha
s facilitated
the
developm
ent of several int
e
lligent vehicl
es
that rely on mode
rn tools and techn
o
lo
gy for t
heir perform
an
ce. The safety of an automobil
e
is
of the hi
ghe
st
pri
o
rity a
c
cording to
a
re
ce
nt re
port
[1]. The rep
o
rt co
mmission
ed by
Wo
rld
Hea
l
th
Orga
nization
in its
Global
Status stu
d
y
on
Road
Saf
e
ty 2013,
rev
ealed th
at th
e main
cau
s
e
o
f
death fo
r yo
u
ng p
eopl
e
(15
-
29
ag
e) glob
ally is due
to
road
traffic
co
llision
s
. Even
though
vario
u
s
cou
n
trie
s h
a
ve initiated
an
d taken
step
s to redu
ce
ro
ad traffic
colli
sion
s
and
a
c
cide
nts, the
total
numbe
r of co
llision
s
and traffic accident
s remai
n
as
high as 1.2
4
million per ye
ar [2]. By 2020,
the road
traffi
c fatalitie
s
are p
r
oje
c
ted
to in
crea
se
by aroun
d 6
5
% glob
ally [3].
In
a co
untry li
ke
India, 1 in
20,
000
peopl
e di
e an
d 12
in
7
0
,000
peopl
e
su
stain
se
rio
u
s i
n
juri
es ev
ery yea
r
du
e
to
road a
c
cide
nts [4].
India h
a
s the
se
con
d
la
rge
s
t road
net
wo
rk in th
e
worl
d an
d i
s
al
so
kno
w
n
for the
high
est
numbe
r of ro
ad accide
nts
and fatalities
in the wo
rl
d [5]. Data publi
s
he
d by the Nation
al Cri
m
e
Re
cords Bureau
(NCRB),
Minist
ry of
Home
Affair
s, Governmen
t
of India,
shows th
at ro
ad
accide
nt fatalities have b
e
en ste
adily ri
sing e
a
ch
ye
ar an
d in ye
ar 20
08, there we
re 1
18,2
3
9
fatalities d
u
e
to road
a
cci
dents [6]. A
signifi
cant
po
rtion
of the
s
e
ro
ad
crashe
s
a
nd
accid
e
n
ts
involve cars a
nd two-wh
eel
ers.
The ro
ad a
c
cidents a
r
e in
cre
a
si
ng du
e
to increa
se i
n
numbe
r of vehicle
s
on t
he roa
d
day by day a
nd al
so the
d
ue to the a
b
sence of
any i
n
telligent hig
h
way
safety and al
ert sy
stem.
Acco
rdi
ng to
data given in
a study [7], the num
ber
of people
wh
o l
o
st their live
s
in India du
e
to
road a
c
cide
nts wa
s alm
o
st
0.11 million death
s
in 200
6, which wa
s
almost 10%
of the total road
accide
nt dea
ths in the world. Accordi
ng to
the accide
nt research
study co
ndu
cted by JP
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
1, No. 3, March 20
16 : 635 – 646
636
Re
sea
r
ch In
dia Pvt. Ltd. for the Ah
medab
ad
-G
a
ndhin
aga
r re
gion (citie
s
of India), for the
duratio
n Feb
r
uary 20
14 to
Jan
uary 2
0
1
5
, total 206 road traffic
accide
nts
were
recorded a
n
d
these
we
re
in
fluenced by t
h
ree
main
factors i.e.
huma
n
, vehicle, i
n
frast
r
u
c
ture
o
r
a combi
natio
n
of them [8].
Figure 1. Influences o
n
roa
d
traffic accid
ents
The nu
mbe
r
in the figure
1 is p
e
rcenta
ge of
the tot
a
l numb
e
r
of
accide
nts su
rveyed.
Acco
rdi
ng to
the re
co
rd, h
u
man fa
ctor i
n
fluen
ce
o
n
road traffic a
c
cide
nts wa
s 92%,
vehicle
9%
and i
n
fra
s
tru
c
ture 4
5
%. Ou
t of total 45%
(91
a
cci
dent
s) infra
s
tru
c
tu
re influ
e
n
c
ed
road
a
cci
dent
s,
6% (1
2 a
c
cid
ents) were d
ue to
animal
s
o
n
the
ro
a
d
wherea
s o
u
t of total 9
2
%
(17
1
)
hum
an
factor influe
n
c
ed road a
c
cidents, 14% (24) were
d
u
e
to driver inattention and a
b
se
nce of any
timely alert
system for preventing
the collision.
Similar types of
surveys were co
nducted for the
Mumbai
-Pun
e expre
s
swa
y
and Coimb
a
tore by JP
Re
sea
r
ch Ind
i
a Pvt. Ltd. and the co
ncl
u
sion
s
hinted at
a si
gnifica
nt
pe
rcentage
of
roa
d
a
cci
dent
s
resultin
g d
ue
to an
obje
c
t
(animal
)
o
n
th
e
road, d
r
iver in
attention and
absen
ce of a
n
intelligent hi
ghway safety alert sy
stem
.
2. Literature
Sur
v
e
y
Applicatio
ns
based o
n
ani
mal dete
c
tion
have an
im
p
o
rtant role in
providin
g sol
u
tions to
many real-life
problem
s. S
o
me of
the
s
e
appli
c
atio
n
s
a
r
e
detect
an
d
tra
c
k anim
a
l
s
li
ke
elep
ha
nts
in forest
s for und
erstandi
ng thei
r
beh
avior
with
th
e environm
e
n
t, preve
n
tin
g
anim
a
l ve
hicle
colli
sion
on road
s, and fo
r preventin
g the entry of
d
ange
rou
s
a
n
i
m
als in
a residential a
r
ea
[9].
The ba
se for most of the appli
c
ation
s
i
s
the det
e
c
ti
on of animal
s
in the vide
o or imag
e. Many
appli
c
ation
s
require hum
an
intervention.
A rece
nt stu
d
y [10] rece
n
t
ly also reve
al
ed that hu
man bein
g
s
need to take
the final
deci
s
io
n duri
ng driving
wh
ether they ca
n cont
rol
thei
r ca
r to prevent colli
sion
with a re
sp
o
n
se
time of 150m
s. The proble
m
with this m
e
thod is
that
human eye
s
get tired ea
sil
y
and need
some
rest
co
nsi
s
te
ntly which is
why this
met
hod i
s
not
th
at effective. Re
sea
r
che
r
s
in [11] reve
aled a
techni
que fo
r the identifica
t
ion of sala
m
ande
rs
by
do
rsal
skin p
a
ttern
s. Thi
s
techni
que
reve
als
key point
s al
ong the skel
eton of the a
n
imal to
be l
abelle
d physi
cally (m
a
nual
ly) by the user.
Similar
user i
nput [12]
is required
for th
e dete
c
ti
on
a
nd id
entificati
on of
ele
pha
nts from th
e
ear
profile.
Some scientif
ic re
se
arche
r
s [13] have
categor
i
z
e
d
an
imals u
s
in
g a
n
arrang
eme
n
t with a
still ca
mera
mounted
at o
ne si
de of
a strip (corrido
r).
This
arra
ng
ement ma
ke
s the dete
c
tion
of
animal
s
pa
ssing the st
rip
(co
rri
do
r) trivi
a
l. Some
me
thods [14] n
eed the a
n
i
m
als to
strike a
particula
r a
n
g
l
e o
r
a
po
se
t
o
wa
rd
s th
e
camera fo
r the
trigg
e
r, in
clu
d
ing fa
ce
det
ection. A
po
p
u
lar
clue for the
detectio
n
of animal
s
is m
o
tion. T
he fundame
n
tal h
y
pothesi
s
he
re [15] is that th
e
default po
sition is
station
a
ry and
ca
n
simply
be
subtracte
d
. All blobs
wh
ich remain
a
fter
backg
rou
n
d
subtra
ction
are
co
nsi
d
e
r
ed
a
s
regi
on
of
int
e
re
st (ca
ndid
a
te)
dete
c
tion
s. Th
oug
h thi
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Real
-Tim
e Autom
a
tic Obst
acle
Dete
ctio
n and Alert
System
for Dri
v
er A
ssi
stan
ce … (Sachin
S.)
637
works fine in
rest
ricted
areas, e.g., for unde
rwat
er
videos, such
mech
ani
sm d
oes n
o
t hold
in
gene
ral settings.
Re
sea
r
che
r
s [16] used thre
shol
d se
g
m
entation a
ppro
a
ch for
getting the targete
d
animal’
s deta
ils from b
a
ckgrou
nd. Recent re
sea
r
ch
es
[17] al
so
reveale
d
that
it is difficult to
deci
de the th
reshold valu
e
as the ba
ckgrou
nd
chan
ges often. A
method a
ppli
c
abl
e to moving
backg
rou
n
d
s
(e.g., due to came
ra motio
n
) is p
r
e
s
ente
d
in sub
s
e
q
u
ent studie
s
[1
8] and [19]. The
tracking lig
ht (sp
a
rse
)
feature
s
point
s over
time and applie
s RA
NSAC to separate backg
rou
nd &
foreg
r
ou
nd
motion. The
backg
rou
nd
motion is ta
ken as the m
a
in (do
m
ina
n
t
) motion in
th
e
scene. T
he
rest of th
e mo
tion is taken
as
a
singl
e
o
b
ject
whi
c
h
can b
e
the
ani
mal/ ob
stacl
e
of
intere
st. The probl
em with
this metho
d
is that
other
moving obje
c
ts woul
d dist
urb the a
ppro
a
ch
and may be falsely dete
c
te
d as anim
a
ls.
Re
sea
r
che
r
s in [20] trie
d to di
scove
r
an
anim
a
l’
s p
r
e
s
en
ce i
n
the
scene
(imag
e
)
affecting the
powe
r
spe
c
trum of the
image. This method of
animal detection was
also
con
s
id
ere
d
n
o
t app
rop
r
iat
e
si
nce q
u
icker
re
sults
wit
h
this metho
d
wo
uld i
n
volve gig
antic am
ount
of image p
r
o
c
e
ssi
ng in a
sho
r
t peri
od
of time
[21].
Some re
se
arche
r
s [2
2] propo
sed a m
e
thod
for anim
a
l sp
ecie
s d
e
tecti
on an
d ma
de
simila
r rest
riction
s
rega
rdi
ng the fo
reg
r
ound
obje
c
ts in
the video.
Th
e d
r
awba
ck o
f
this p
r
o
p
o
s
e
d
meth
od
wa
s that it
was
highly
spe
c
ifi
c
to th
e
sp
eci
f
ic
setting
and
n
o
t valid in
the
co
ntext of th
e India
n
roa
d
s
whe
r
e
seve
ral
animal
s
m
a
y be
present
a
t
the s
a
me time.
Re
sea
r
che
r
s
in [23] al
so u
s
ed th
e fa
ce
detecto
r te
ch
nique i
n
itiate
d by Viola a
n
d
Jo
ne
s
for a
spe
c
ific
animal type.
Once an
ani
mal face
is
d
e
tected
and i
dentified, the
resea
r
chers t
r
y to
track it
over time. The
p
r
o
b
lem
with thi
s
te
chni
que
i
s
that
face d
e
tection
ne
ed
s the
a
n
imal
s to
look i
n
to the
cam
e
ra
whi
c
h i
s
, in g
e
n
e
ral, n
o
t ne
cessarily
capt
ured
by the
road travel video.
Animals
can
arrive fro
m
a scene fro
m
variou
s di
re
ctions a
nd in va
riou
s si
ze
s, pose
s
an
d col
o
rs.
An intere
sting approa
ch for the animal
detection
an
d tracking u
s
es a texture
descri
p
tor ba
se
d
on SIFT that tries to mat
c
h it against a
pred
efi
ned li
bra
r
y of anim
a
l textures [2
4]. The probl
em
with thi
s
m
e
thod i
s
th
at it
is
re
stri
cted
to video
s h
a
ving
singl
e
animal
only
and ve
ry min
i
mal
backg
rou
nd clutter. Both condition
s are not met
espe
cially with ani
mals p
r
e
s
ent
on roa
d
sid
e
s.
In Saudi Ara
b
ia, the num
ber of collisi
ons b
e
twe
e
n
the camel
and a vehi
cl
e we
re
estimated to
rea
c
h mo
re t
han a h
und
re
d each year.
To preve
n
t these
colli
sion
s, an intelligen
t
Camel
Vehi
cle Accide
nt
Avoidance S
y
stem (CVA
AS) was de
sign
ed
usi
n
g
GPS (glob
a
l
pos
itioning sys
tem) [25].
For find
ing
th
e corre
c
t p
o
sition of fi
she
s
in th
e
sea,
rese
arche
r
s [
26]
desi
gne
d a tech
niqu
e usi
ng LIDAR (li
ght detec
tion
and rangi
ng
). Usin
g the micro
-
Doppl
er
techni
que [27
], resea
r
chers also tried av
oiding ri
sky animal intru
s
io
ns in the ho
u
s
ing a
r
ea.
Lack of re
so
urces
and q
u
a
lity database of image
s
(for trainin
g
a
nd testing
)
of
animal
s
on the road i
s
still o
ne of t
he challen
g
in
g tasks.
D
ue
to this fact a
nd a
s
pe
r ou
r latest survey
s,
very less
wo
rk ha
s bee
n re
ported
so far i
n
this are
a
at least in conte
x
t to
Indian hi
ghways. In this
pape
r, we
prese
n
t a nove
l
approa
ch fo
r the dete
c
tio
n
of animal
s
(co
w
) on Indi
an ro
ad
s and
a
method fo
r
calcul
ating the
dista
n
ce of
detecte
d
a
n
imal from
the
came
ra
mou
n
ted vehi
cle
(after
detectio
n
) wh
ich can hel
p the drive
r
to av
oid the colli
sion vehicl
e-a
n
imal colli
sio
n
.
Intelligent hig
h
way safety and d
r
iver a
s
sista
n
ce sy
stems a
r
e ve
ry helpful to re
duce the
numbe
r of a
c
cide
nts that a
r
e ha
ppe
ning
due to vehi
cl
e-ani
mal colli
sion
s. With
resp
ect to Indi
an
road
s, two types of anim
a
ls – the co
w and t
he dog
are foun
d more often tha
n
others on the
road
s. Spe
c
ific obje
c
tives o
f
the rese
arch
work are:
To develop a
n
automatic a
n
imal dete
c
ti
on syste
m
in context to Indian roa
d
s.
Finding the a
pproxim
ate di
stan
ce of ani
mal fr
om the
vehicle in
whi
c
h came
ra is
mounted.
To develop a
n
alert sy
ste
m
once th
e a
n
imal gets d
e
t
ected on the
road which will help the
driver in
appl
ying b
r
akes
or ta
kin
g
oth
e
r
ne
ce
ssa
ry
actio
n
fo
r a
v
oiding
colli
si
on b
e
twe
e
n
vehicle a
nd a
n
imal.
3. Brief Ov
erv
i
e
w
and Ad
v
a
ntages of
HOG a
nd Ca
scade
Classi
fier
A histogram
of oriente
d
gradient
s (HO
G
) is a
c
tually
a feature de
scripto
r
used
to detect
obje
c
ts in
co
mputer visi
on
and imag
e p
r
ocessin
g
[28
]. The HOG
descri
p
tor te
chniqu
e re
cou
n
ts
occurre
n
ces
of gradi
ent o
r
ientation i
n
locali
ze
d po
rtions
of an im
age -
dete
c
tion win
d
o
w
, or
region of interes
t
(ROI). Fi
g
u
re 2 sho
w
s t
he algo
rithmi
c implem
enta
t
ion scheme
of HOG.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
1, No. 3, March 20
16 : 635 – 646
638
Figure 2. Algorithmi
c
impl
ementation
schem
e of hog
The
HO
G de
scripto
r
i
s
p
a
rticularly
app
ropria
te
for an
imal dete
c
tio
n
or hu
man
d
e
tection
in video o
r
im
age
s du
e to some
key adv
antage
s
com
pare
d
to othe
r de
scripto
r
s.
First, it ope
rat
e
s
on lo
cal
cell
s so it i
s
i
n
variant to ge
om
etric
and
ph
o
t
ometric tran
sform
a
tion
s.
Secon
d
ly coa
r
se
(sp
a
tial)
sam
p
ling, fine ori
entation sam
p
ling an
d strong lo
cal ph
otometri
c no
rmalizatio
n all
o
w
individual
bo
dy moveme
nt of anim
a
ls/p
ede
strian
s
to
be
overlo
oked if they m
a
intain a
ro
ug
hly
uprig
ht positi
on [29].
Ca
scadin
g
is a
pa
rticul
ar case
of
gro
up l
earni
ng
ba
se
d
on
the con
c
aten
ation of several
cla
ssifie
r
s, u
s
ing all info
rm
ation collecte
d
from t
he
o
u
tput from a
given cl
assifier a
s
ad
ditio
nal
information for the next
c
l
as
s
i
fier in the
c
a
s
c
ade
[30]. The
key adv
antage
s
of b
ooste
d
ca
sca
de
cla
ssifie
r
s over mo
nolithi
c cla
ssifie
r
s are that it
is a
fast lea
r
ne
r a
nd requi
re
s l
o
w
com
putation
time. Ca
scad
ing al
so elim
inates
ca
ndid
a
tes
(false
po
s
i
tive
s)
e
a
r
l
y o
n
,
s
o
la
ter
s
t
ag
es
d
on’t
bother a
bout
them.
Figure 3. Boosted casca
d
e
classifier
As sh
own in the figure 3, e
a
ch filter
reje
cts n
on-obje
c
t windo
ws
an
d let obje
c
t windo
ws
past to
the n
e
xt layer of t
he
ca
scade.
A wind
ow
is
con
s
id
ere
d
a
s
a
n
obj
ect if
and
only of
all
layers of the
ca
scade
cla
s
sifies it as o
b
j
e
ct
[31]. The filter i of the cascad
e is de
signed to
Reje
ct the large po
ssi
ble n
u
mbe
r
of non
-obje
c
t win
d
o
w
s
To allow la
rg
e possibl
e nu
mber of obj
ect windo
ws for
quick evalu
a
tion
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Real
-Tim
e Autom
a
tic Obst
acle
Dete
ctio
n and Alert
System
for Dri
v
er A
ssi
stan
ce … (Sachin
S.)
639
4. Rese
arch
Metho
dolog
y
As sho
w
n a
b
o
ve in Fig
u
re
4, a video i
s
taken f
r
om
a
forward
-
fa
cin
g
ca
mera in
whi
c
h a
moving anim
a
l is pre
s
ent
apart from ot
her stat
io
nary and non-stationary obj
e
c
ts. This vide
o is
store
d
in th
e
comp
uter
and
conve
r
ted i
n
to diffe
re
nt
fra
m
es. We are
usin
g
a com
b
ination
of HO
G
and
b
o
o
s
ted ca
scade cla
s
sifiers
for ani
mal
dete
c
ti
on
. All the imag
e pro
c
e
s
sing
techni
que
s a
r
e
impleme
n
ted
in Op
en
CV
software.
On
ce the
anim
a
l gets dete
c
ted
in
th
e
vide
o, the
next step
is
to find the
distance
of the
a
n
imal fro
m
th
e testin
g vehi
cle
and th
en
alert the
drive
r
so that
he
can
apply the
bra
k
e
s
o
r
p
e
rfo
r
m any oth
e
r
necessa
ry
a
c
tion which i
s
displ
a
yed o
n
comm
and
pro
m
p
t
as a me
ssag
e.
Figure 4. Block di
agram of
the propo
se
d
method
Followi
ng is
the pro
posed
pro
c
edu
re f
o
r traini
ng a
nd testing of
the data for animal
detectio
n
:
Colle
ct all po
sitive and ne
gative image
s in the data folder
Gene
rate
An
notation
Cre
a
te sa
mpl
e
i.e. generat
e .vec file (figure 5
)
Train d
a
ta i.e. generating xml file (figure
6, 7)
Testing
(figure
8)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
1, No. 3, March 20
16 : 635 – 646
640
Figure 5. Cre
a
te sampl
e
Figure 6. Trai
n data
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Real
-Tim
e Autom
a
tic Obst
acle
Dete
ctio
n and Alert
System
for Dri
v
er A
ssi
stan
ce … (Sachin
S.)
641
Figure 7. XML file
Figure 8. Testing
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
1, No. 3, March 20
16 : 635 – 646
642
Figure 9. Dist
ance cal
c
ul
ation
As sho
w
n in the figure 9, video is taken
and co
nverte
d into frames (image of si
ze 640 *
480). Follo
wi
ng is the pro
c
ed
ure for
ca
lculatin
g t
he distan
ce of the detecte
d
animal from t
he
came
ra
-mo
u
n
ted vehicl
e:
Image re
sol
u
tion is 64
0 × 4
8
0
X range i
s
0 to 640
Y range i
s
0 to 480
Let the right b
o
ttom coo
r
din
a
te of the detected
co
w be
(x, y)
Then the di
stance of co
w from the lo
wer edge (ca
r
/ca
m
era
)
is 48
0 – y
5. Data
Anal
y
s
is and Interpretation
W
e
a
r
e us
in
g H
O
G
de
scr
ip
to
r
s
wh
ich
ar
e
fe
a
t
u
r
e
de
scripto
r
s
an
d are used
i
n
compute
r
vision an
d image processi
ng for the
pu
rpose of obje
c
t detection. F
o
r obje
c
t cla
s
sificatio
n
, we
are
usin
g boo
ste
d
ca
scade cl
assifiers. For prepa
ring
th
e requi
r
ed d
a
taba
se, we
are pe
rformi
ng
animal detect
ion in relation to the Indian
scenario
as no research has been
performed till date i
n
this are
a
, an
d not many sou
r
ces a
r
e
pre
s
ent relat
ed to this scenari
o
. A good so
urce fo
r th
e
animal im
age
s is the KT
H
dataset
[32] and NEC dat
aset
[33]
th
at
inclu
ded i
m
a
ges
of cows
and
dog
s (of our interest
). Some more a
n
i
m
al image
s
have been
cl
icked for cre
a
ting a healt
h
y
databa
se of
almost 9
00 i
m
age
s con
s
i
s
ting of po
si
t
i
ve image
s
i
n
whi
c
h the
target ani
mal
is
pre
s
ent a
nd
negative ima
ges i
n
which
there i
s
n
o
target
animal
for feature e
x
traction
and
for
training the
cl
assifier.
After the cla
s
sifier i
s
train
e
d
and the d
e
t
ecti
on sy
ste
m
is built, we
tested the
same on
variou
s vide
o
s
. We te
sted
our meth
od
on at l
e
a
s
t 2
8
to 3
0
vide
os
(fra
me
s of
si
ze
640*
48
0),
inclu
d
ing
80
animal
s
(co
w
) in th
e video
. A detectio
n
rate of
almo
st 80% was
achieved
with l
o
w
false
dete
c
tio
n
rate in
case of
singl
e o
b
ject
(c
ow) i
n
the f
r
ame
(testing
video
)
. Training
a
n
d
testing on l
a
rge datasets
with diffe
rent orientations
a
nd different weather
co
nditi
ons will
improve
the detectio
n
rate and ove
r
all
efficien
cy of the system
.
Some of the
scre
en
shot
s of the
cam
e
ra
-mou
nted
vehicle
an
d
re
sults (with
different
climate condit
i
ons a
nd with
different sp
ee
ds) a
r
e
sho
w
n in figure
s
1
0
, 11, 12, 13, 14 and 1
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
Real
-Tim
e Autom
a
tic Obst
acle
Dete
ctio
n and Alert
System
for Dri
v
er A
ssi
stan
ce … (Sachin
S.)
643
Figure 10. Ca
mera mo
unte
d
vehicle
Figure 11. Animal detectio
n
at 0 kmph
speed
wi
th obj
ect station
a
ry
in mornin
g condition
Figure 12. Animal detecte
d
at a speed of
40 kmph in a
fternoon
con
d
i
tion
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
1, No. 3, March 20
16 : 635 – 646
644
Figure 13. Animal detecte
d
at a distance
of 228
pixels from the cam
e
ra mo
unted
vehicle
with
the spe
ed of 60 kmp
h
in e
v
ening time
Figure 14. Multiple animal
s
dete
c
ted in
one of the testing video
Figure 15. Multiple animal
s
dete
c
ted in
the se
con
d
testing video
Evaluation Warning : The document was created with Spire.PDF for Python.