<s>
In	O
computer	B-Application
vision	I-Application
,	O
a	O
stixel	B-General_Concept
(	O
portmanteau	O
of	O
"	O
stick	O
"	O
and	O
"	O
pixel	B-Algorithm
"	O
)	O
is	O
a	O
superpixel	O
representation	O
of	O
depth	O
information	O
in	O
an	O
image	O
,	O
in	O
the	O
form	O
of	O
a	O
vertical	O
stick	O
that	O
approximates	O
the	O
closest	O
obstacles	O
within	O
a	O
certain	O
vertical	O
slice	O
of	O
the	O
scene	O
.	O
</s>
<s>
Introduced	O
in	O
2009	O
,	O
stixels	B-General_Concept
have	O
applications	O
in	O
robotic	B-General_Concept
navigation	I-General_Concept
and	O
advanced	O
driver-assistance	O
systems	O
,	O
where	O
they	O
can	O
be	O
used	O
to	O
define	O
a	O
representation	O
of	O
robotic	O
environments	O
and	O
traffic	O
scenes	O
with	O
a	O
medium	O
level	O
of	O
abstraction	O
.	O
</s>
<s>
One	O
of	O
the	O
problems	O
of	O
scene	O
understanding	O
in	O
computer	B-Application
vision	I-Application
is	O
to	O
determine	O
horizontal	O
freespace	O
around	O
the	O
camera	O
,	O
where	O
the	O
agent	O
can	O
move	O
,	O
and	O
the	O
vertical	O
obstacles	O
delimiting	O
it	O
.	O
</s>
<s>
One	O
drawback	O
of	O
dense	O
reconstruction	O
is	O
the	O
large	O
amount	O
of	O
data	O
involved	O
,	O
since	O
each	O
pixel	B-Algorithm
in	O
the	O
image	O
is	O
mapped	O
to	O
an	O
element	O
of	O
a	O
point	O
cloud	O
.	O
</s>
<s>
Vision	O
problems	O
characterised	O
by	O
planar	O
freespace	O
delimited	O
by	O
mostly	O
vertical	O
obstacles	O
,	O
such	O
as	O
traffic	O
scenes	O
or	O
robotic	B-General_Concept
navigation	I-General_Concept
,	O
can	O
benefit	O
from	O
a	O
condensed	O
representation	O
that	O
allows	O
to	O
save	O
memory	O
and	O
processing	O
time	O
.	O
</s>
<s>
Stixels	B-General_Concept
are	O
thin	O
vertical	O
rectangles	O
representing	O
a	O
slice	O
of	O
a	O
vertical	O
surface	O
belonging	O
to	O
the	O
closest	O
obstacle	O
in	O
the	O
observed	O
scene	O
.	O
</s>
<s>
A	O
stixel	B-General_Concept
is	O
characterised	O
by	O
three	O
parameters	O
:	O
vertical	O
coordinate	O
of	O
the	O
bottom	O
,	O
height	O
of	O
the	O
stick	O
,	O
and	O
depth	O
.	O
</s>
<s>
Stixels	B-General_Concept
have	O
fixed	O
width	O
,	O
with	O
each	O
stixel	B-General_Concept
spanning	O
over	O
a	O
certain	O
number	O
of	O
image	O
columns	O
,	O
allowing	O
downsampling	O
of	O
the	O
horizontal	O
image	O
resolution	O
.	O
</s>
<s>
In	O
the	O
original	O
formulation	O
,	O
each	O
column	O
of	O
the	O
image	O
would	O
contain	O
at	O
most	O
one	O
stixel	B-General_Concept
,	O
and	O
later	O
extensions	O
were	O
developed	O
to	O
allow	O
multiple	O
stixels	B-General_Concept
on	O
each	O
column	O
,	O
allowing	O
to	O
represent	O
multiple	O
objects	O
at	O
different	O
distances	O
.	O
</s>
<s>
The	O
input	O
to	O
stixel	B-General_Concept
estimation	O
is	O
a	O
dense	O
depth	O
map	O
,	O
that	O
can	O
be	O
computed	O
from	O
stereo	O
disparity	O
or	O
other	O
means	O
.	O
</s>
<s>
The	O
original	O
approach	O
computes	O
an	O
occupancy	B-General_Concept
grid	I-General_Concept
that	O
can	O
be	O
segmented	O
to	O
estimate	O
the	O
freespace	O
,	O
with	O
dynamic	B-Algorithm
programming	I-Algorithm
providing	O
an	O
efficient	O
method	O
to	O
find	O
an	O
optimal	O
segmentation	O
.	O
</s>
<s>
Alternative	O
approaches	O
can	O
be	O
used	O
instead	O
of	O
occupancy	B-General_Concept
grid	I-General_Concept
mapping	I-General_Concept
,	O
such	O
as	O
manifold-based	O
methods	O
.	O
</s>
<s>
A	O
membership	O
function	O
over	O
the	O
pixels	B-Algorithm
can	O
be	O
defined	O
based	O
on	O
the	O
depth	O
value	O
,	O
where	O
the	O
membership	O
represents	O
the	O
confidence	O
of	O
a	O
pixel	B-Algorithm
belonging	O
to	O
the	O
closest	O
vertical	O
obstacle	O
or	O
to	O
the	O
background	O
,	O
and	O
a	O
cut	O
separating	O
the	O
obstacles	O
from	O
the	O
background	O
can	O
again	O
be	O
computed	O
effectively	O
with	O
dynamic	B-Algorithm
programming	I-Algorithm
.	O
</s>
<s>
Once	O
both	O
the	O
freespace	O
and	O
the	O
obstacle	O
height	O
are	O
known	O
,	O
the	O
stixels	B-General_Concept
can	O
be	O
estimated	O
by	O
fusing	O
the	O
information	O
over	O
the	O
columns	O
spanned	O
by	O
each	O
stixel	B-General_Concept
,	O
and	O
finally	O
a	O
refined	O
depth	O
of	O
the	O
stixel	B-General_Concept
can	O
be	O
estimated	O
via	O
model	O
fitting	O
over	O
the	O
depth	O
of	O
the	O
pixels	B-Algorithm
covered	O
by	O
the	O
stixel	B-General_Concept
,	O
possibly	O
paired	O
with	O
confidence	O
information	O
(	O
e.g.	O
</s>
<s>
disparity	O
confidence	O
produced	O
by	O
methods	O
such	O
as	O
semi-global	B-General_Concept
matching	I-General_Concept
)	O
.	O
</s>
