Inter
national
J
our
nal
of
Electrical
and
Computer
Engineering
(IJECE)
V
ol.
7,
No.
6,
December
2017,
pp.
3037
–
3045
ISSN:
2088-8708
3037
I
ns
t
it
u
t
e
o
f
A
d
v
a
nce
d
Eng
ine
e
r
i
ng
a
nd
S
cie
nce
w
w
w
.
i
a
e
s
j
o
u
r
n
a
l
.
c
o
m
P
edestrian
Detection
using
T
riple
Laser
Range
Finders
Abdul
Hadi
Abd
Rahman
1
,
Khairul
Akram
Zainol
Ariffin
2
,
Nor
Samsiah
Sani
3
,
and
Hairi
Zamzuri
4
1,2,3
Center
for
Artificial
Intelligence
T
echnology
,
F
aculty
of
Information
Science
and
T
echnology
,
,
Uni
v
ersiti
K
ebangsaan
Malaysia.
4
V
ehicle
System
Engineering
Research
Lab,
Uni
v
ersiti
T
eknologi
Malaysia,
Jalan
Semarak,
54100
K
uala
Lumpur
,
Malaysia.
Article
Inf
o
Article
history:
Recei
v
ed:
Jun
6,
2017
Re
vised:
Aug
21,
2017
Accepted:
Sep
3,
2017
K
eyw
ord:
Pedestrian
Detection
Laser
Range
Finder
Autonomous
ABSTRA
CT
Pedestrian
detection
is
one
of
the
important
features
in
autonomous
ground
v
ehicle
(A
GV).
It
ensures
the
capability
for
safety
na
vi
g
a
tion
in
urban
en
vironment.
Therefore,
the
detec-
tion
accurac
y
became
a
crucial
part
which
leads
to
implementation
using
Laser
Range
Finder
(LRF)
for
better
data
representation.
In
this
study
,
an
impro
v
ed
laser
configuration
and
fusion
techni
que
is
introduced
by
implementation
of
triple
LRFs
in
tw
o
layers
with
Pedestrian
Data
Analysis
(PD
A)
t
o
recognize
multiple
pedestrians.
The
PD
A
inte
grates
v
arious
features
from
feature
e
xtraction
process
for
all
clusters
and
fusion
of
multiple
lay-
ers
for
better
recognition.
The
e
xperiments
were
conducted
in
v
arious
occlusion
scenarios
such
as
intersection,
closed-pedestrian
and
combine
scenarios.
The
analysis
of
the
laser
fu-
sion
and
PD
A
for
all
scenarios
sho
wed
an
impro
v
ement
of
detection
where
the
pedestrians
were
represent
ed
by
v
arious
detection
cate
gories
which
solv
e
occlusion
issues
when
lo
w
number
of
laser
data
were
obtained.
Copyright
c
2017
Institute
of
Advanced
Engineering
and
Science
.
All
rights
r
eserved.
Corresponding
A
uthor:
Abdul
Hadi
Abd
Rahman
Center
for
Artificial
Intelligence
T
echnology
,
F
aculty
of
Information
Science
and
T
echnology
,
Uni
v
ersiti
K
ebangsaan
Malaysia
(UKM)
43600
UKM,
Bangi
Selangor
,
Malaysia
Phone:
+603-8921
6712
Email:
abdulhadi@ukm.edu.my
1.
INTR
ODUCTION
Pedestrian
Detection
and
T
racking
(PDT)
for
autonomous
ground
v
ehicle
has
attracted
more
attention
no
w
a-
days.
A
reliable
PDT
contrib
utes
to
a
significant
impro
v
ement
for
other
scenarios
such
as
obstacle
a
v
oidance,
path
planning
and
collision
a
v
oidance.
The
presence
of
Laser
Range
Finder
(LRF)
whic
h
is
capable
of
pro
viding
accurate
range
information,
wide
co
v
erage
area
and
a
lo
w
time
interv
al
permits
implementations
in
real
time
system.
A
reli-
able
and
ef
ficient
pedestrian
detection
in
urban
area
is
one
of
the
crucial
for
successful
autonomous
na
vig
ation.
Laser
range
finder
pro
vides
v
aluable
data
of
the
surrounding
especially
for
pedestria
n
detection
b
ut
there
are
crucial
limita-
tions
that
need
to
be
considered:
a
pedestrian
could
be
se
gmented
into
se
v
eral
se
gments
caused
by
partial
occlusions
and
laser
-absorbed
such
as
glassy
or
blac
k
surf
aces,
and
only
parts
of
the
objects
f
acing
the
sensor
are
visible
which
often
changes
as
the
object
mo
v
es
which
could
de
grade
the
detection
result.
It
w
as
suggested
that
the
LRF
placement
on
a
v
ehicle
or
robot
is
important
in
determining
detection
of
body
parts
either
w
aist
or
le
gs
of
pedestrians.
W
aist
and
le
gs
are
tw
o
of
the
most
ob
vious
features
which
could
be
v
ery
helpful
in
classification
of
a
pedestrian
especially
in
LRF
implementation.
Both
implementation
ha
v
e
their
o
wn
adv
antages
and
dra
wbacks.
A
small
le
g
size
af
fects
the
amount
of
detected
laser
data
especially
in
long
range
implementation
thus
may
produce
misclassification
between
le
g
and
measurement
noises.
Meanwhile
the
w
aist
part
data
may
contain
data
of
pedestrians
hands
which
may
sometimes
cause
occlusion
for
full
w
aist
data.
The
v
arious
orientation
of
pedestrian
could
easily
ha
v
e
af
fected
the
detection
misclassification
and
isolation
of
feature
motion.
A
single
planar
approached
using
LRF
is
not
suf
ficient
enough
for
observing
dif
ferent
object
whi
ch
are
closed
to
each
other
[1,
2,
3,
4].
The
measurement
quality
of
detected
object
is
unequal.
A
high
quality
measurement
is
achie
v
ed
when
an
object
is
in
a
clear
vie
w
to
the
scanner
.
The
measurement
obtained
is
complete
and
good
shape
for
further
e
v
aluation.
Contradictly
for
block
objects
or
when
the
sensor
is
block
ed,
it
may
be
represented
by
partially
J
ournal
Homepage:
http://iaesjournal.com/online/inde
x.php/IJECE
I
ns
t
it
u
t
e
o
f
A
d
v
a
nce
d
Eng
ine
e
r
i
ng
a
nd
S
cie
nce
w
w
w
.
i
a
e
s
j
o
u
r
n
a
l
.
c
o
m
Evaluation Warning : The document was created with Spire.PDF for Python.
3038
ISSN:
2088-8708
and
ambiguous
shape
[5,
6].
The
representati
on
of
pedestrian
could
be
impro
v
ed
using
multiple
LRFs
in
multi-
layer
implementation.
It
increased
the
possibilities
for
pedestrian
detection
in
an
en
vironment
for
state
estimation.
Ho
we
v
er
,
it
is
challenging
to
handle
data
from
v
arious
sources
in
term
of
data
inte
gration.
There
are
f
actors
that
need
to
be
considered
such
as
sensor
calibration,
resource
allocation
and
fusion
technique
while
maintaining
a
lo
w
computation
cost.
Thorpe
[7]
implemented
a
three
LRFs
configuration
mounted
on
v
arious
positions
placed
at
front
and
both
sides
of
a
v
ehicle
to
pro
vide
horizontal
profiling
while
W
ang
[8]
co
v
ered
for
both
horizontal
and
v
ertical
profiling.
The
y
suggested
that
pedestrian
detection
using
laser
scanners
were
dif
ficult
because
the
number
of
measurement
points
associated
wi
th
a
pedestrian
is
often
small
in
the
applications.
Recognition
algorithms
can
be
used
to
confirm
the
results
of
lidar
-based
detection.
Hashimoto
et
al.
[9]
further
enhanced
the
configuration
by
allocating
three
LRFs
in
dif
ferent
layer
to
co
v
er
knee,
thigh
and
w
aist
part
of
pedest
rian.
A
decentralised
architecture
approaches
is
chosen
to
pro
vide
a
de
gree
of
scalability
and
rob
ustness
compared
to
centralised
architecture.
Sato
et
al.
[10]
has
de
v
eloped
a
pedestrian
tracking
method
in
v
arious
urban
en
vironments
to
impro
v
e
the
pedes
trian
detection
rate
for
f
alse/miss
alarm
using
a
six-layer
-LRFs.
Carballo
[2,
11]
implemented
a
fusion
of
tw
o
LRFs
which
arranged
f
acing
in
opposite
directions
to
co
v
er
360
surrounding.
Then
it
w
as
e
xtended
in
multiple
layers
to
create
3D
model
of
people.
Elliptical
shape
computed
using
Romanujanss
approximation
for
chest
area
while
small
circular
shapes
detected
in
le
g
part.
The
centroid
estimation
e
xtracted
from
the
w
aist
part
w
as
projected
to
lo
wer
body
part
to
find
correspond
le
gs
using
a
v
erage
w
alking
steps.
The
inclusion
of
reflection
intensity
data
of
LRF
arranged
in
multiple
layers
w
as
introduced
in
[12]
for
people
detection.
The
y
included
a
calibrated
intensity
feature
to
the
e
xisting
Adaboost
to
train
better
and
strong
classifiers.
Carballo
[13]
further
e
xtended
the
fusion
method
by
combining
tw
o
LRFs
in
tw
o
layers
to
co
v
er
measurement
of
tw
o
dif
ferent
body
parts.
Ho
we
v
er
,
a
reduction
of
50
scan
points
has
to
be
done
for
g
athering
simultaneous
range
and
intensity
from
the
multi
LRFS.
McKinle
y
et
al.
[14]
and
Kim
et
al.
[15]
highlighted
the
use
of
multiple
LRFs
to
impro
v
e
the
performance
of
the
detection
algorithms
due
to
the
increased
amount
of
data
for
better
rob
ustness
ag
ainst
occlusion.
Ho
we
v
er
,
the
scheme
w
as
highly
dependent
on
the
correct
alignment
of
multi
LRFs
and
could
cause
system
f
ailure
if
the
misalignment
w
as
lar
ge
enough.
Mozos
et
al.
[16]
allocated
three
LRFs
in
dif
ferent
layers
for
detecting
head,
w
aist
and
le
g
parts
of
pedestrian.
The
y
assorted
detection
result
by
the
le
v
el
of
confidence
of
each
detected
se
gment.
A
higher
le
v
el
confidence
is
defined
when
all
pedestrian
part
is
detected
while
lo
w
confidence
referre
d
to
enough
detection
of
an
y
body
part
s.
The
approach
w
as
able
to
produce
significant
impro
v
ement
to
e
xisting
configurations.
Gidel
et
al.
[17]
presented
a
pedestrian
detection
system
using
4
horizontal
plane
layer
of
LRFs
for
f
alse
detections
w
as
reduced
in
comparison
with
application
using
a
single
laser
scanner
.
An
e
xtremum
map
is
computed
by
calculating
all
related
probabilities
of
a
pedestrian
weighted
by
the
intersections
of
number
of
layers.
The
fusion
of
the
four
layers
of
LRFs
were
e
x
ecuted
in
decentralised
architecture.
Experimental
results
pro
v
ed
that
the
usage
of
four
laser
planes
has
impro
v
ed
the
pedestrian
detection
with
a
lo
wer
f
alse
alarms.
Impro
v
ement
of
detection
approach
is
still
an
ongoing
process.
There
are
still
limitations
on
the
pre
vious
implementation
for
pedestrian
detection.
Pre
vious
researches
as
mentioned
earlier
sho
wed
that
LRF
configurations
and
placement
is
one
of
the
f
actor
that
af
fected
the
detection
performance.
More
specific
application
could
be
further
e
xplored
for
performance
e
v
aluation
using
LRF
.
There
were
less
in
v
estig
ations
on
implementing
LRFs
in
a
multilayer
configuration
in
outdoor
en
vironment
to
deal
with
high
measurement
noises.
Therefore,
this
paper
analyses
on
the
performance
enhancement
in
detection
by
proposing
ne
w
laser
fusion
approach
using
mult
i
LRFs
in
a
multilayer
configuration.
2.
RESEARCH
METHODOLOGY
In
this
study
,
Hokuyo
Laser
Range
Finder
(LRF)
URG04-LX
model
has
been
selected
as
main
sensor
to
pro
vide
en
vironment
data.
The
LRF
w
as
selected
based
on
light-weight
and
easy
to
mount
on
an
y
v
ehicle.
It
has
a
wide
co
v
erage
area
for
LRF
which
are
240
with
0.33
angle
resolution,
maximum
distance
co
v
erage
of
5.6
meters
in
0.1
sec
time
interv
al.
F
or
that,
a
custom
mounted
for
has
been
de
v
eloped
to
place
all
LRFs
as
sho
wn
in
Figure
1.
The
three
LRFs
were
configured
in
tw
o
multilayer
co
v
erage.
One
LRF
w
as
placed
at
the
center
in
front
of
the
v
ehicle
with
height
1.2
meter
to
co
v
er
the
width
part
of
pedestrians.
F
or
bottom
layer
,
tw
o
LRFs
were
positioned
on
both
side
in
front
of
the
v
ehicle
at
0.4
meter
abo
v
e
ground
to
produce
an
interlace
of
pedestrians
le
gs
data.
The
distance
between
right
and
left
LRFs
w
as
set
as
1.0
meter
considering
the
v
ehicle
width
and
mounting
limitation.
The
LRFs
were
filtered
to
co
v
er
180
angle
and
5
meters
in
distance
for
all
LRFS
to
co
v
er
focused
area
of
tar
geted
pedestrian.
A
set
of
calibration
tests
were
done
to
ensure
the
accurac
y
of
the
produced
data
to
represent
le
gs
[18,
19]
.
The
calibration
results
could
not
perform
a
100%
accurac
y
due
to
sensor
noise
of
LRFs
b
ut
it
achie
v
ed
considerably
reliable
output.
The
fusion
technique
to
solv
e
pedestrian
detection
in
an
outdoor
en
vironment
from
a
ground
mo
ving
v
ehicle
w
as
equipped
with
three
Hokuyo
Laser
Range
Finders
(LRFs)
which
were
configured
in
tw
o
dif
ferent
layers.
The
detection
process
in
v
olv
ed
a
consecuti
v
e
processing
steps
containing
pre-process,
pedestrian
IJECE
V
ol.
7,
No.
6,
December
2017:
3037
–
3045
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISSN:
2088-8708
3039
Figure
1.
Mounted
LRFs
on
v
ehicle
platform
where
(a)
LRF
position
on
top
center
(b)
LRF
position
on
bottom
right
(c)
LRF
position
on
bottom
left.
analysis,
map
matching
and
feature
e
xtracti
o
n.
Ra
w
data
for
each
laser
w
as
pre-processed
before
being
fused
to-
gether
.
Then,
the
pedestrian
data
analysis
(PD
A)
w
as
performed
to
produce
the
observ
ation
output
after
passing
the
matching
process
with
the
de
v
eloped
online
feature
e
xtraction
mapping.
There
are
four
steps
in
pre-processing
tech-
nique
which
includes
data
clustering,
pedestrian
analysis,
map
matching
and
feature
e
xtraction.
The
pre-processed
(clustering)
of
streaming
dat
a
for
each
LRF
w
as
e
x
ecuted
using
parallel
processing
technique.
A
tw
o
le
v
el
fusion
w
as
proposed
in
v
olving
fusion
of
both
LRFS
at
the
bottom
to
calculate
the
distance
of
each
e
xtracted
points
for
pedestrian
le
gs.
The
details
on
fusion
processing
step
are
e
xplained
in
Algorithm
1.
Algorithm
1
Laser
fusion
using
decentralised
multi
threaded
process
1:
pr
ocedur
e
L
A
S
E
R
F
U
S
I
O
N
2:
Start
streaming
data
from
LRF
top
centre.
3:
Start
streaming
data
from
LRF
bottom
left.
4:
Start
streaming
data
from
LRF
bottom
right.
5:
Multithreaded
decentralized
laser
data
pre-processing.
6:
f
or
each
node
in
bottom
LRFs
do
7:
data
clustering
8:
pedestrian
data
analysis
9:
add
point
e
xtraction
to
ArrayList
10:
end
f
or
11:
Find
possible
association
for
right
and
left
le
gs
12:
f
or
top
LRF
do
13:
data
clustering
14:
pedestrian
data
analysis
15:
find
association
with
bottom
ArrayList
16:
end
f
or
17:
if
association
==
true
then
18:
do
point
e
xtraction
fusion
using
2
19:
Add
to
observ
ation
list
20:
end
if
21:
if
association
==
f
alse
then
22:
calculate
point
e
xtraction
based
on
top
LRF
of
bottom
LRFs
only
23:
Add
to
observ
ation
list
24:
end
if
25:
end
pr
ocedur
e
P
edestrian
Detection
using
T
riple
Laser
Rang
e
F
inder
s
(A.H.A.
Rahman)
Evaluation Warning : The document was created with Spire.PDF for Python.
3040
ISSN:
2088-8708
All
cl
usters
resulted
f
rom
the
pre
vious
process
could
possibly
containing
pedestrian
or
similar
-lik
e
object.
Further
analysis
for
each
cluster
is
required
to
determine
the
final
observ
ation
for
tracking
process.
This
process
is
important
to
produce
high
quality
observ
ation
while
reducing
the
f
alse
alarm
containing
undesirable
objects.
There
are
fe
w
criteria
ha
v
e
been
identified
to
determine
detection
of
pedestrian
including
feature
analysis,
inte
gration
of
results
for
top
and
bottom
laser
,
and
occupanc
y
grid
in
determine
static
or
dynamic
pedestrian.
Based
on
conclusion
and
suggestion
in
Arras
[5],
fe
w
best
features
had
been
selected
and
e
v
aluated
for
implementation
for
feature-based
analysis
for
pedestrian
detection
described
as
follo
ws;
i)
number
of
elements
(N)
and
width
(W),
curv
ature,
mean
an-
gular
dif
ference,
radius,
boundary
length,
and
multi-layer
as
sociation.
The
Pedestrian
Data
Analysis
(PD
A)
process
is
sho
wn
in
Figure
2.
Figure
2.
Pedestrian
Data
Analysis
for
pedestrian
detection
confirmation
3.
EV
ALU
A
TION
OF
DETECTION
AND
F
ALSE
RA
TE
There
are
tw
o
most
important
parameters
for
pedestrian
detection
e
v
aluation
which
are
detection
rate
(DR)
and
f
alse
alarm
rate
(F
AR).
Detection
rate
repre
sents
the
detection
accurac
y
of
the
implemented
approach.
Imple-
mentation
using
2
LRFs
w
as
chosen
as
benchmark
for
e
v
aluation
of
the
proposed
approach.
Since
there
are
no
ground
truth
data
for
the
testing
dataset,
the
analyses
of
the
detection
and
f
alse
alarm
rate
were
done
manually
for
each
testing
scenario.
In
each
frame
of
data,
all
mo
ving
objects
for
each
data
frame
were
identified
with
v
alidation
from
the
recorded
video
sequences.
F
or
each
mo
ving
object,
true
and
f
alse
detection
were
identified
and
counted
for
analysis.
The
parameters
for
Pedestrian
Data
Analysis
(PD
A)
are
described
in
pre
vious
section.
This
e
v
aluation
is
important
to
get
the
best
range
for
all
parameters
in
v
olv
ed
in
determination
of
detected
pedestrians
based
on
the
laser
data
input.
A
total
number
of
948
scans
in
an
outdoor
en
vironment
were
collected
where
it
in
v
olv
ed
pedestrian
in
both
mo
ving
and
standing
still.
The
total
number
of
se
gments
e
xtracted
were
5349
se
gments.
F
or
each
scan,
the
laser
data
w
as
clustered
and
analyzed
using
the
proposed
pedestrian
data
analysis
module.
Results
obtained
from
e
xperiments
using
a
modified
Pedestrian
Data
Analysis
and
compared
with
commonly
used
approach
from
literature.
Collection
of
data
obtained
from
the
e
xperiment
in
both
static
and
mo
ving
v
ehicle
using
the
proposed
configuration
of
3
LRF
were
compared
with
pre
vious
selected
multilayer
implementation.
T
abl
e
1
lists
the
results
of
the
analysis
has
been
sorted
based
on
PD
A
into
5
cate
gories
as
follo
ws:
C1
-
w
aist
with
one
le
g
for
either
one
of
the
bottom
LRFs,
C2
-
w
aist
with
one
le
g
for
both
LRFs
bottom,
C3
-
w
aist
with
tw
o
le
gs
for
either
one
of
the
bottom
LRF
,
C4
-
w
aist
with
tw
o
le
gs
for
both
bottom
LRFs
and
C5
-
w
aist
only
.
From
Pedestrian
Data
Analysis
(PD
A)
results,
it
can
be
concluded
that
the
most
informati
v
e
feature
is
the
radius
of
the
circle
that
fitted
into
the
se
gment.
The
mean
angular
dif
ference
is
the
second
most
important
feature
which
quantified
the
con
v
e
xity
of
the
se
gment.
The
combination
of
curv
ature
and
radius
does
not
measure
the
de
gree
of
circularity
b
ut
pro
vide
e
xtra
information
of
det
ected
pedestrians.
IJECE
V
ol.
7,
No.
6,
December
2017:
3037
–
3045
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISSN:
2088-8708
3041
T
able
1.
A
v
erage
v
alues
for
all
parameters
in
pedestrian
data
analysis.
total
cluster
cluster
boundary
mean
circle
clusters
element
width
(mm)
length
(mm)
angular
(
)
radius(mm)
C1
711
20
495
617
54.9
183
C2
1095
17
532
636
62.3
188
C3
828
24
619
613
42
182
C4
1842
22
600
669
53
183
C5
873
13
523
513
59.4
186
3.1.
P
edestrian
Detection
Ev
aluation
The
pedestrian
detection
and
tracking
during
intersection
w
as
performed
to
demonstrate
the
pedestrian
identification
and
track
association
capabilities.
T
w
o
method
were
in
v
olv
ed
in
this
e
v
aluation.
Con
v
entional
laser
configuration
(2LRF)
represents
combination
of
laser
configuration
using
2
LRFs
and
e
xisting
con
v
entional
Mul-
tiple
Hypothesis
T
racking
(MHT)
approach.
The
proposed
laser
configurations
(3LRF)
approach
consists
of
laser
fusion
using
3
LRFs
in
tw
o
layers
with
an
impro
v
e
MHT
method
from
pre
vious
research
in
[19].
In
the
simulation
and
e
xperiment,
a
multiple
pedestrian
situation
from
dif
ferent
direction
and
v
elocity
with
intersecting
trajectory
is
simulated
and
e
xperimented.
T
o
ensure
the
ef
fecti
v
eness
of
the
propose
dyMHT
algorithm,
the
e
v
aluation
on
the
per
-
formance
were
conducted
separately
before
an
y
further
tests
were
proceed.
The
proposed
tracking
algorithm
were
e
v
aluated
in
tw
o
e
xperiments.
First,
the
e
xperiments
were
conducted
for
separate
conditions
for
intersection
and
closed-pedestrian.
F
or
intersection
scenarios,
a
series
of
repeated
e
xperiments
were
arranged
to
produce
pedestrian
intersections.
T
o
encourage
frequent
track
crossings,
the
pathw
ay
for
tar
gets
were
defined
which
in
v
olv
e
pedestrians
v
elocity
state
components
which
slightly
bias
to
w
ard
the
each
of
others.
Each
scenario
w
as
simulated
and
e
xperi-
mented
in
three
repeated
tests.
It
w
as
e
xpected
that
the
algorithm
w
as
able
to
track
all
pedestrian
in
all
intersection
scenarios.
Then,
the
e
xperiment
for
detect
ion
for
closed-pedestrians
(multi-pedestrian
w
alking
side
by
side)
were
conducted
in
which
more
complicated
compared
to
intersection
cases
due
to
uncertain
occlusion
interv
al
depending
on
pedestrian
w
alking
pattern
and
direction.
Figure
3.
Scenario
1:
Pedestrian
Detection
during
Intersection
Cases.
Figure
3
sho
ws
the
detection
results
for
mo
ving
pedestrian
during
intersection.
The
simulation
data
for
this
scenario
is
represented
by
’x’
mark
er
for
benchmarking
purposes.
It
is
observ
ed
that
at
certai
n
parts
of
the
detection,
the
implementation
using
3
LFRs
w
as
able
to
produce
more
observ
ations
result
based
the
plotted
point
for
Point
Extraction
(PE)
compared
to
2
LRFs
approach.
It
is
supported
by
the
a
v
erage
detection
rate
which
obtained
dur
-
ing
intersection
scenarios
using
the
proposed
approach
which
w
as
98.9%
compared
to
92.6%
using
the
benchmark
approach.
T
otal
of
4.3%
reduction
of
f
alse
al
arm
w
as
achie
v
ed
for
the
proposed
approach.
In
general,
impro
v
ed
detection
results
were
observ
ed
compared
to
benchmark
approach
which
pr
o
duce
d
lo
wer
f
alse
alarm
due
to
imple-
P
edestrian
Detection
using
T
riple
Laser
Rang
e
F
inder
s
(A.H.A.
Rahman)
Evaluation Warning : The document was created with Spire.PDF for Python.
3042
ISSN:
2088-8708
mented
occupanc
y
grid
and
multi
clustering
during
fusion
process.
The
intersection
detection
result
which
described
the
situation
is
highlighted
in
circle
labelled
A
in
Figure
3.
It
sho
ws
tw
o
pedestrians
on
the
right
w
alk
ed
across
each
other
which
caused
intersection
to
happen.
Before
intersection
happened,
both
approaches
were
able
to
detect
both
pedestrian.
The
front
pedestrian
w
as
well
spotted
with
dif
ferent
point
e
xtraction
while
the
second
pedestrian
only
detected
by
top
laser
.
During
intersection,
the
proposed
approach
w
as
able
to
detect
the
occluded
pedestrian
as
opposed
to
benchmark
method.
It
remained
for
a
fe
w
iterations
before
the
benchmark
approach
able
to
redetect
the
occluded
objects.
The
recorded
pedestrian
data
for
2LRF
approach
w
as
less
accurate
to
be
classify
as
pedestrian
during
PD
A
process,
since
the
occluded
pedestrian
w
as
only
detected
after
a
fe
w
iterations
when
partial
occlusion
decreased.
Meanwhile,
the
proposed
approach
w
as
successfully
deal
with
this
problem
where
the
generated
pedestrian
data
obtained
with
the
additional
LRF
allo
wed
a
better
detection
results.
The
second
e
xperiment
e
v
aluated
the
capability
of
the
proposed
method
to
deal
with
group
pedestrians
w
alking
closed
to
each
other
which
produced
a
comple
x
occlusion
scenario.
Figure
4
sho
ws
the
detection
results
of
fi
v
e
consistent
pedestrians
for
closed
scenario
where
pedestrian
mo
ving
closed
to
each
other
which
led
to
occlusion.
It
is
observ
ed
that
in
2LRFs
approach,
Pedest
rians
with
ID
#4
and
#5
suf
fered
from
occlusion
at
the
end
of
the
pathw
ay
.
The
proposed
configuration
w
as
able
to
reco
v
er
more
detections
for
Pedestrians
#5.
The
detection
and
f
alse
alarm
rate
for
this
scenario
obtained
a
higher
a
v
erage
detection
rate
at
89.6%
w
as
achie
v
ed
in
closed
scenarios
using
the
proposed
approach
compared
to
78.7%
using
the
benchmark
approach,
2LRFs.
Both
approaches
(3
LRFs
and
2
LRFs)
produced
the
same
f
alse
alarm
rate
at
4.2%.
The
pre-post
detection
results
for
closed-pedestrian
scenario
were
as
labelled
B
in
Figure
4.
It
sho
ws
tw
o
pedestrians
on
the
right
w
alk
ed
across
each
other
which
caused
occlusion
to
happen.
Before
occlusion
happened,
both
approaches
were
able
to
detect
all
pedestrians.
The
front
pedestrian
w
as
spotted
with
dif
ferent
point
e
xtraction
while
the
second
pedestrian
only
detected
by
top
laser
.
During,
the
proposed
approach
w
as
not
able
to
detect
the
occluded
pedestrian
similar
to
benchmark
method.
It
lasted
for
a
fe
w
iterations
before
the
benchmark
approach
able
to
redetect
the
tw
o
occluded
objects
b
ut
ne
v
er
reco
v
ered
the
third
occluded
pedestrian.
Figure
4.
Scenario
2:
Pedestrian
Detection
during
Closed-pedestrian
Cases
The
detection
results
using
fusion
of
LRFs
for
combine
case
scenarios
are
s
h
o
wn
in
Figure
5
where
tw
o
closed
pedestrian
scenarios
(2
pedestrians
each)
and
2
intersection
cases.
The
detection
of
fi
v
e
pedestrians
were
found
correctly
with
some
occlusions
remain
appeared
in
tracking
process.
The
a
v
erage
detection
rate
at
93.5%
w
as
achie
v
ed
in
this
scenario
using
the
proposed
approach
compared
to
78.2%
using
the
benchmark
approach.
The
f
alse
alarm
achie
v
ed
for
both
approaches
(3
LRFs
and
2
LRFs)
were
at
5.1%.
3.2.
Effect
of
Laser
Fusion
in
P
edestrian
Detection
The
approach
presented
in
this
research
described
a
multi-part
person
detection
based
on
multiple
2D
LRF
scans.
The
first
highlight
in
pedestrians
detection
using
laser
scanners
is
the
position
of
the
lasers.
F
or
the
e
xperiments
presented
in
this
research,
the
lasers
were
placed
at
tw
o
dif
ferent
fix
ed
heights.
These
heights
were
selected
to
co
v
er
IJECE
V
ol.
7,
No.
6,
December
2017:
3037
–
3045
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISSN:
2088-8708
3043
Figure
5.
Scenario
3:
Pedestrian
detection
during
Combine
Intersection
and
Closed-pedestrian
a
range
of
400
mm
for
feet
and
1200
mm
for
w
aist.
Ho
we
v
er
,
in
this
research,
it
has
been
find
out
that
2
layers
with
e
xtra
LRF
unit
for
bottom
layer
to
c
o
v
er
feet
le
v
el
were
enough
to
impro
vise
the
pedestrian
detection.
The
detection
results
produced
in
all
four
e
xperiments
pro
v
ed
that
better
detections
accuracies
were
achie
v
ed.
This
outcome
has
been
supported
by
a
research
conducted
by
Carballo
et
al.
[14]
and
Mozos
et
al.
[17]
who
ha
v
e
performed
a
fusion
of
multiple
layers
by
se
gmentation
of
fused
scan
data,
geometrical
features
e
xtraction
and
association
for
e
v
ery
detected
person
to
allo
w
good
position
estimation
and
prediction
pedestrians
direction.
The
combination
of
both
areas
creates
a
3D
v
olume
which
helps
locating
the
position
of
the
person
more
closely
related
to
the
center
of
det
ected
pedestrian.
The
e
xperimental
results
found
that
the
proposed
LRF
configurations
were
able
to
increase
the
detection
rate
and
lo
wer
or
same
f
alse
alarm
rate
in
all
gi
v
en
scena
rios.
The
Pedestrian
Data
Analysis
(PD
A)
w
as
applied
to
solv
e
the
misclassification
rates
thus
achie
ving
lo
wer
f
alse
alarm
rate.
In
comparison,
Carballo
et
al.
[13]
found
that
laser
intensity
w
as
able
to
impro
v
e
the
detection
results
in
the
single-layer
system.
Ho
we
v
er
only
minimal
f
alse
alarm
rate
reduction
w
as
achie
v
ed,
b
ut
highlighting
about
detection
rates
with
smaller
misclassification
rates.
Another
issue
highlighted
by
w
as
the
dra
wback
to
get
simultaneously
range
and
intensity
from
multi-sensors,
where
higher
angular
resolution
w
as
used,
contrib
uting
to
50%
reduction
of
total
scan
points.
T
o
solv
e
the
proble
m,
an
optimised
parallel
processing
w
as
implemented
for
fusion
of
all
laser
data
with
full
angle
resolution.
The
e
xperiments
were
conducted
to
solv
e
se
v
eral
critical
scenarios.
It
in
v
olv
ed
people
w
alk
across,
the
trajectory
of
each
person
intersect
each
other
which
causing
detections
fragmented
into
se
v
eral
parts.
In
the
scenario
where
laser
tracking
f
ails
due
to
the
data
confusion
when
people
w
alk
close
together
,
that
situation
is
v
ery
dif
ficult
to
deal
with.
The
situation
is
comple
x
where
the
laser
data
of
the
bottom
person
is
lost
due
to
occ
lusion
by
other
people.
The
tracking
process
depended
on
the
direction
and
speed
of
upper
person
which
treated
as
group
tracking
due
to
the
high
confidence
in
similarity
of
mo
v
ement
and
pattern.
Furthermore,
combine
case
scenarios
were
conducted
for
performance
e
v
aluation.
The
presented
fusion
of
LRFs
w
ork
ed
f
airly
well
with
much
better
performance
when
the
v
ehicle
w
as
static
than
when
the
v
ehicle
mo
v
ed
as
seen
in
the
e
xperimental
results.
The
use
of
multiple
LRFs
im
pro
v
ed
the
performance
of
the
detection
algorithms
due
to
the
increased
amount
of
data
and
made
object
tracking
more
rob
ust
ag
ainst
occlusion.
Ho
we
v
er
,
the
scheme
w
as
highly
dependent
on
the
correct
ali
gnment
of
the
tw
o
bottom
LRFs
and
could
cause
system
f
ailure
if
the
misalignment
w
as
lar
ge
enough.
F
or
that,
in
this
research,
a
series
of
v
erification
and
calibration
of
the
fusion
result
were
done
before
running
the
e
xperiments.
The
a
v
erage
detection
rate
for
all
e
xperiments
for
the
proposed
fusion
method
w
as
recorded
at
92.5%
which
is
an
increment
of
9.9%
from
the
benchmark
approach.
The
a
v
erage
f
alse
alarm
rate
for
both
implementations
were
5%
and
5.9%
which
represents
(0.9%)
of
reduction.
4.
CONCLUSION
This
paper
presented
the
e
xperiment
al
results,
analysis
and
discussion
for
the
proposed
configurati
on
for
fusion
of
three
LRFs
using
pedestrian
data
analysis.
It
is
sho
wn
that
it
w
as
able
to
achie
v
e
better
detection
results
and
assure
detection
of
static
objects.
The
e
xperimental
results
in
dif
ferent
outdoor
s
cenarios
sho
wed
an
increment
in
pedestrian
detection
accurac
y
compared
to
implementation
using
double
layer
of
tw
o
LRFs.
P
edestrian
Detection
using
T
riple
Laser
Rang
e
F
inder
s
(A.H.A.
Rahman)
Evaluation Warning : The document was created with Spire.PDF for Python.
3044
ISSN:
2088-8708
A
CKNO
WLEDGEMENT
The
authors
w
ould
lik
e
to
thank
the
Malaysia-Japan
Internati
onal
Institute
of
T
echnology
(MJIIT)
in
Uni-
v
ersiti
T
eknologi
Malaysia
for
the
support
and
funding
of
this
research.
REFERENCES
[1]
X.
Shao,
H.
Zhao,
K.
Nakamura,
K.
Katabira,
R.
Shibasaki,
and
Y
.
Nakag
a
w
a,
“Detection
and
T
racking
of
Multiple
Pedestrians
by
Using
Laser
Range
Scanners,
”
in
IEEE/RSJ
International
Confer
ence
on
Intellig
ent
Robots
and
Systems
,
no.
1.
IEEE,
2007,
pp.
2174–2179.
[2]
A.
Carballo,
A.
Oh
ya,
and
S.
Y
uta,
“Fusion
of
double
layered
multiple
laser
range
finders
for
people
detection
from
a
mobile
robot.
”
IEEE,
Aug.
2008,
pp.
677–682.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=4648023
[3]
Y
.
Y
akiyama,
N.
Thepvilojanapong,
M.
Iw
ai,
O.
Mihirogi,
K.
Umeda,
and
Y
.
T
obe,
AIPSJ
Online
T
r
ansactions
,
v
ol.
4,
pp.
515–528,
2009.
[4]
L.
Adia
viak
o
ye,
P
.
P
atrick,
B.
Marc,
and
J.
M.
Auberlet,
“T
racking
of
multiple
people
in
cro
wds
using
laser
range
scanners,
”
in
2014
IEEE
Ninth
International
Confer
ence
on
Intellig
ent
Sensor
s,
Sensor
Networks
and
Information
Pr
ocessing
(ISSNIP)
,
April
2014,
pp.
1–6.
[5]
K.
O.
Arras,
O.
M.
Mozos,
and
W
.
Bur
g
ard,
“Using
Boosted
Features
for
the
Detection
of
People
in
2D
Range
Data,
”
in
Pr
oceedings
2007
IEEE
International
Confer
ence
on
Robotics
and
A
utomation
.
IEEE,
Apr
.
2007,
pp.
3402–3407.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=4209616
[6]
T
.
Zhao,
R.
Ne
v
atia,
and
B.
W
u,
Se
gmentation
and
tr
ac
king
of
multiple
humans
in
cr
owded
en
vir
onments,
IEEE
T
r
ansactions
on
P
attern
Analysis
and
Mac
hine
Intellig
ence
,
v
ol.
30,
pp.
1198–211,
2008.
[7]
C.
Thorpe
and
A.
Suppe,
“LAD
AR-based
detection
and
tracking
of
mo
ving
object
s
from
a
ground
v
ehicle
at
high
speeds,
”
in
IEEE
IV2003
Intellig
ent
V
ehicles
Symposium.
Pr
oceedings
(Cat.
No.03TH8683)
.
IEEE,
2003,
pp.
416–421.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=4209616
[8]
C.-C.
W
ang,
“Simultaneous
Localization,
Mapping
And
Mo
ving
Object
T
racking,
”
Ph.D.
dissertation,
Carne
gie
Mellon
Uni
v
ersity
,
2004.
[9]
M.
Hashimoto,
Y
.
Matsui,
and
K.
T
akahashi,
“Mo
ving-object
tracking
with
multi-laser
range
sensors
for
mobile
robot
na
vig
ation,
”
in
2007
IEEE
International
Confer
ence
on
Robotics
and
Biomimetics
(R
OBIO)
.
IEEE,
2007,
pp.
399–404.
[10]
S.
Sato,
M.
Hashimoto,
M.
T
akita,
K.
T
akagi,
and
T
.
Og
a
w
a,
“Multilayer
lidar
-based
pedestrian
tracking
in
urban
en
vironments,
”
in
2010
IEEE
Intellig
ent
V
ehicles
Symposium
.
IEEE,
Jun.
2010,
pp.
849–854.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=5548135
[11]
A.
Carballo,
A.
Oh
ya,
and
S.
Y
uta,
“People
detection
using
range
and
intensity
data
from
multi-layered
Laser
Range
Finders,
”
in
2010
IEEE/RSJ
International
Confer
ence
on
Intellig
ent
Robots
and
Systems
.
IEEE,
Oct.
2010,
pp.
5849–5854.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=5649769
[12]
——,
“Laser
reflection
intensity
and
multi-layered
Laser
Range
Finders
for
people
detection,
”
in
19th
International
Symposium
in
Robot
and
Human
Inter
active
Communication
.
IEEE,
2010,
pp.
379–384.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=5649769
[13]
——,
Reliable
P
eople
Detection
Using
Rang
e
and
Intensity
Data
fr
om
Multiple
Layer
s
of
Laser
Rang
e
F
inder
s
on
a
Mobile
Robot,
International
J
ournal
of
Social
Robotics
,
v
ol.
3,
pp.
167–186,
2011.
[14]
N.
McKinle
y
,
“Simultaneous
Localization,
Mappings
and
Object
T
racking
in
an
Urban
En
vironment
using
Multiple
2D
Laser
Scanners,
”
Ph.D.
dissertation,
2010.
[15]
B.
Kim,
B.
Choi,
M.
Y
oo,
H.
Kim,
and
E.
Kim,
Rob
ust
Object
Se
gmentation
Using
a
Multi-Layer
Laser
Scanner
,
Sensor
s
,
v
ol.
14(11),
pp.
20
400–20
418,
2014.
IJECE
V
ol.
7,
No.
6,
December
2017:
3037
–
3045
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISSN:
2088-8708
3045
[16]
O.
M.
Mozos,
R.
K
urazume,
and
T
.
Hase
g
a
w
a,
Multi-P
art
P
eople
Detection
Using
2D
Rang
e
Data,
Interna-
tional
J
ournal
of
Social
Robotics
,
v
ol.
2,
pp.
31–40,
2010.
[17]
S.
Gidel,
P
.
Checchin,
C.
Blanc,
T
.
Chateau,
and
L.
T
rassoudaine,
P
edestrian
Detection
and
T
r
ac
king
in
an
Urban
En
vir
onment
Using
a
Multilayer
Laser
Scanner
,
v
ol.
11,
pp.
579–588,
2010.
[18]
A.
H.
A.
Rahman,
H.
Zamzuri,
S.
A.
Mazlan,
and
M.
A.
A.
Rahman,
“Model-Based
Detection
and
T
racking
of
Single
Mo
ving
Object
Using
Laser
Range
Finder,
”
in
2014
5th
International
Confer
ence
on
Intellig
ent
Systems,
Modelling
and
Simulation
.
IEEE,
jan
2014,
pp.
556–561.
[Online].
A
v
ailable:
http://ieee
xplore.ieee.or
g/lpdocs/epic03/wrapper
.htm?arnumber=7280971
[19]
A.
Rahman,
H.
Zamzuri,
and
S.
A.
Mazlan,
Dynamic
T
r
ac
k
Mana
g
eme
n
t
in
MHT
for
P
edestrian
T
r
ac
king
Using
Laser
Rang
e
F
inder
,
Mathematical
Pr
oblems
in
Engineering
,
v
ol.
2015,
pp.
1–9,
2015.
[20]
S.
Gidel,
P
.
Checchin,
C.
Blanc,
T
.
Chateau,
L.
T
rassoudaine,
and
U.
B.
P
ascal,
“Pedestrian
Detection
Method
using
a
Multilayer
Laserscanner
:
Application
in
Urban
En
vironment,
”
in
2008
IEEE/RSJ
International
Con-
fer
ence
on
Intellig
ent
Robots
and
Systems
.
IEEE,
2008,
pp.
22–26.
BIOGRAPHIES
OF
A
UTHORS
Abdul
Hadi
Abd
Rahman
is
a
senior
lecturer
at
Center
for
Artificial
Intelligence
T
echnology
,
F
aculty
of
Information
Science
and
T
echnology
,
Uni
v
ersiti
K
ebangsaan
Malaysia.
He
obtained
Doctor
of
Philosoph
y
from
Uni
v
ersiti
T
eknologi
Malaysia.
His
researches
are
in
fields
of
object
tracking
and
artificial
intelligence.
Further
info
on
his
homepage:
http://www
.ftsm.ukm.my/hadi
Khairul
Akram
Zainol
Ariffin
is
a
senior
lecturer
at
Center
for
Softw
are
T
echnology
and
Man-
agement,
F
a
culty
of
Information
Science
and
T
echnology
,
Uni
v
ersiti
K
ebangsaan
Malaysia.
He
obtained
Doctor
of
Philosoph
y
f
rom
Uni
v
ersiti
T
eknologi
Petronas.
His
researches
are
in
fields
of
netw
orking
and
c
yber
security
.
Nor
Samsiah
Sani
is
a
senior
lecturer
at
Center
f
or
Artificial
Intelligence
T
echnology
,
F
aculty
of
Information
Science
and
T
echnology
,
Uni
v
ersiti
K
ebangsaan
Malaysia.
His
researches
are
in
fields
of
machine
learning
and
artificial
intelligence.
Hairi
Zamzuri
is
an
Associate
Professor
at
Malaysia-Japan
International
Institute
of
T
echnology
(MJIIT),
Uni
v
ersiti
T
eknologi
Malaysia.
His
researches
are
in
fields
of
v
ehicle
dynamic,
rail
w
ay
v
ehicle,
autonomous
v
ehicle,
v
ehicle
suspension
design.
P
edestrian
Detection
using
T
riple
Laser
Rang
e
F
inder
s
(A.H.A.
Rahman)
Evaluation Warning : The document was created with Spire.PDF for Python.