Graphics Reference
In-Depth Information
feature extraction step that can cater to such contextual information is worthy of
further study.
It can be seen that
E
+
have nonlane features also. We now propose shift
and match technique to extract lane features and eliminate nonlane features from
each band. In order to do this, we compute the horizontal projection vectors
p
+
and
E
−
and
p
as shown in Fig.
10.3
. Peaks are formed in these projection
vectors where there are clusters of pixels in
E
for
E
and
E
+
−
−
and
E
. Since the dark
→
light and
+
−
light
→
dark transitions in a lane marking are separated by
δ
pixels in the IPM image
I
W
, the peaks corresponding to the lane edges in
p
and
p
are also separated by
+
−
asmall
δ
. In order to capture these pairs of transitions of lanes,
p
is shifted by
δ
+
places to the left and multiplied with
p
−
resulting in the vector
K
B
i
for scan band
B
i
, i.e.,
K
=
(
p
+
δ)
p
(10.1)
−
where
represents pointwise multiplication. Figure
10.3
shows the result of the
shift and match operation performed on
p
for the upper band selected in
Fig.
10.2
. It can be seen that we get peaks in
K
B
i
in Fig.
10.3
at the same locations
as the left edge of each lane marking in the upper band in Fig.
10.2
. The locations of
the peaks in
K
B
i
for each scan band
B
i
are then used along with the road model to
eliminate outliers.
In this paper, for the sake of illustration and simplicity, we limit the discussion to
a simple straight road model in the IPM domain, i.e., we assume the road is straight
(deviated only by a few pixels). Considering that input images are calibrated with
WCS in IPM image, the lane marking positions can be predicted in a deterministic
manner. Let us take the case of ego-lane. After calibration, if
x
L
and
x
R
correspond to
the lane positions of the left and right lanemarkings of the ego-lane, the lanemarkings
are expected to be in the vicinity of these lane positions. The peaks positions in
K
B
i
from each scan band are mapped to the predicted lane markings
x
L
and
x
R
.This
and
p
+
−
Fig. 10.2
Generating steerable filter output from bands
