<s>
Features	B-Algorithm
from	I-Algorithm
accelerated	I-Algorithm
segment	I-Algorithm
test	I-Algorithm
(	O
FAST	O
)	O
is	O
a	O
corner	B-Algorithm
detection	I-Algorithm
method	O
,	O
which	O
could	O
be	O
used	O
to	O
extract	O
feature	B-Algorithm
points	O
and	O
later	O
used	O
to	O
track	O
and	O
map	O
objects	O
in	O
many	O
computer	B-Application
vision	I-Application
tasks	O
.	O
</s>
<s>
The	O
FAST	O
corner	B-Algorithm
detector	I-Algorithm
was	O
originally	O
developed	O
by	O
Edward	O
Rosten	O
and	O
Tom	O
Drummond	O
,	O
and	O
was	O
published	O
in	O
2006	O
.	O
</s>
<s>
The	O
most	O
promising	O
advantage	O
of	O
the	O
FAST	O
corner	B-Algorithm
detector	I-Algorithm
is	O
its	O
computational	O
efficiency	O
.	O
</s>
<s>
Referring	O
to	O
its	O
name	O
,	O
it	O
is	O
indeed	O
faster	O
than	O
many	O
other	O
well-known	O
feature	B-Algorithm
extraction	O
methods	O
,	O
such	O
as	O
difference	B-Algorithm
of	I-Algorithm
Gaussians	I-Algorithm
(	O
DoG	O
)	O
used	O
by	O
the	O
SIFT	B-Algorithm
,	O
SUSAN	O
and	O
Harris	B-Algorithm
detectors	O
.	O
</s>
<s>
The	O
FAST	O
corner	B-Algorithm
detector	I-Algorithm
is	O
very	O
suitable	O
for	O
real-time	O
video	O
processing	O
application	O
because	O
of	O
this	O
high-speed	O
performance	O
.	O
</s>
<s>
FAST	O
corner	B-Algorithm
detector	I-Algorithm
uses	O
a	O
circle	O
of	O
16	O
pixels	O
(	O
a	O
Bresenham	B-Algorithm
circle	I-Algorithm
of	O
radius	O
3	O
)	O
to	O
classify	O
whether	O
a	O
candidate	O
point	O
p	O
is	O
actually	O
a	O
corner	O
.	O
</s>
<s>
Firstly	O
,	O
corner	B-Algorithm
detection	I-Algorithm
with	O
a	O
given	O
N	O
is	O
processed	O
on	O
a	O
set	O
of	O
training	O
images	O
which	O
are	O
preferable	O
from	O
the	O
target	O
application	O
domain	O
.	O
</s>
<s>
Secondly	O
,	O
a	O
decision	B-Application
tree	I-Application
algorithm	O
,	O
the	O
ID3	B-Algorithm
algorithm	I-Algorithm
is	O
applied	O
to	O
the	O
16	O
locations	O
in	O
order	O
to	O
achieve	O
the	O
maximum	O
information	B-Algorithm
gain	I-Algorithm
.	O
</s>
<s>
The	O
information	B-Algorithm
gain	I-Algorithm
can	O
then	O
be	O
represented	O
as	O
:	O
</s>
<s>
A	O
recursive	O
process	O
is	O
applied	O
to	O
each	O
subsets	O
in	O
order	O
to	O
select	O
each	O
x	O
that	O
could	O
maximize	O
the	O
information	B-Algorithm
gain	I-Algorithm
.	O
</s>
<s>
For	O
example	O
,	O
at	O
first	O
an	O
x	O
is	O
selected	O
to	O
partition	O
P	O
into	O
Pd	O
,	O
Ps	O
,	O
Pb	O
with	O
the	O
most	O
information	O
;	O
then	O
for	O
each	O
subset	O
Pd	O
,	O
Ps	O
,	O
Pb	O
,	O
another	O
y	O
is	O
selected	O
to	O
yield	O
the	O
most	O
information	B-Algorithm
gain	I-Algorithm
(	O
notice	O
that	O
the	O
y	O
could	O
be	O
the	O
same	O
as	O
x	O
)	O
.	O
</s>
<s>
This	O
generated	O
decision	B-Application
tree	I-Application
can	O
then	O
be	O
converted	O
into	O
programming	O
code	O
,	O
such	O
as	O
C	B-Language
and	O
C++	B-Language
,	O
which	O
is	O
just	O
a	O
bunch	O
of	O
nested	O
if-else	O
statements	O
.	O
</s>
<s>
The	O
complied	O
code	O
is	O
used	O
as	O
corner	B-Algorithm
detector	I-Algorithm
later	O
for	O
other	O
images	O
.	O
</s>
<s>
Notice	O
that	O
the	O
corners	O
detected	O
using	O
this	O
decision	B-Application
tree	I-Application
algorithm	O
should	O
be	O
slightly	O
different	O
from	O
the	O
results	O
using	O
segment	O
test	O
detector	O
.	O
</s>
<s>
This	O
is	O
because	O
that	O
decision	B-Application
tree	I-Application
model	I-Application
depends	O
on	O
the	O
training	O
data	O
,	O
which	O
could	O
not	O
cover	O
all	O
possible	O
corners	O
.	O
</s>
<s>
So	O
each	O
time	O
a	O
different	O
t	O
is	O
set	O
for	O
the	O
decision	B-Application
tree	I-Application
algorithm	O
.	O
</s>
<s>
FAST-ER	O
detector	O
is	O
an	O
improvement	O
of	O
the	O
FAST	O
detector	O
using	O
a	O
metaheuristic	B-Algorithm
algorithm	O
,	O
in	O
this	O
case	O
simulated	B-Algorithm
annealing	I-Algorithm
.	O
</s>
<s>
So	O
that	O
after	O
the	O
optimization	O
,	O
the	O
structure	O
of	O
the	O
decision	B-Application
tree	I-Application
would	O
be	O
optimized	O
and	O
suitable	O
for	O
points	O
with	O
high	O
repeatability	O
.	O
</s>
<s>
However	O
,	O
since	O
simulated	B-Algorithm
annealing	I-Algorithm
is	O
a	O
metaheurisic	O
algorithm	O
,	O
each	O
time	O
the	O
algorithm	O
would	O
generate	O
a	O
different	O
optimized	O
decision	B-Application
tree	I-Application
.	O
</s>
<s>
According	O
to	O
Rosten	O
,	O
it	O
takes	O
about	O
200	O
hours	O
on	O
a	O
Pentium	B-General_Concept
4	I-General_Concept
at	O
3GHz	O
which	O
is	O
100	O
repeats	O
of	O
100,000	O
iterations	O
to	O
optimize	O
the	O
FAST	O
detector	O
.	O
</s>
<s>
In	O
Rosten	O
's	O
research	O
,	O
FAST	O
and	O
FAST-ER	O
detector	O
are	O
evaluated	O
on	O
several	O
different	O
datasets	O
and	O
compared	O
with	O
the	O
DoG	O
,	O
Harris	B-Algorithm
,	O
Harris-Laplace	O
,	O
Shi-Tomasi	O
,	O
and	O
SUSAN	O
corner	B-Algorithm
detectors	I-Algorithm
.	O
</s>
<s>
Speed	O
tests	O
were	O
performed	O
on	O
a	O
3.0GHz	O
Pentium	B-Device
4-D	I-Device
computer	O
.	O
</s>
