![](bg3.png)
4) Update the vanishing point estimation:
4.a) Regularly sample some points (with a
four-pixel step) on the first road border denoted
p
s
.
4.b) Through each point of p
s
, respec-
tively construct a set of 29 evenly distributed rays
(spaced by 5
◦
, whose orientation from the hori-
zon is more than 20
◦
, and less than 180
◦
) for each
point of p
s
, denoted as L
s
4.c) From each L
s
, find a subset of n rays
such that their OCRs rank top n among the 29
rays.
4.d) The new vanishing point vp
1
is se-
lected from p
s
as the one which maximizes the
sum of the top n OCRs
1
.
5) Starting from vp
1
, detect the second road
border in a similar way as the first border, with a
constraint that the angle between the road borders
is larger than 20
◦
.
3 OUR APPROACH
Our interest in LASV (Kong et al., 2010) is motivated
by the possibility of using it as a pre-segmentation for
our application. Our implementation aims to work on
a vehicle platform equipped with a 30 FPS camera.
This is why we needed to accelerate its computation
to meet the real time requirements.
We tested this algorithm on several datasets (Geiger
et al., 2012; Kong et al., 2010; Leskovec et al.,
2008), in our case this algorithm will be used with
continuous capture. This is why we have tried to
improve this method on datasets using successive
frames instead of random environment pictures
datasets like the ones used by Kong.
The original code was written in MATLAB by Kong.
During the initial tests, the original implementation
required a computation time of 18 seconds per image.
Our optimisation is mainly concentrated on step
I.3) of the original algorithm. Our improvements
were done in the same environment in several steps.
We have benchmarked each evolution to justify its
purpose.
3.1 Methods
For all methods, we have reduced image size down
to 240 pixels height and 180 pixels width, just like
the original algorithm does. The (Fig.3) illustrates
subsequent methods (it shows an asphalt road, indeed
this algorithm worked on unmarked and marked
road).
Method I Temporal Dependence
a) For first image, computation of vanishing point
research is done on the complete image.
b) For subsequent images, the best vanishing point
candidate is only selected from a 10% by 10% image
size zone around the coordinates of the vanishing
point found previously.
The 10% by 10% range is explained by the fact that
the image will not move much between two frames,
considering camera preset of 30 FPS. The length of
10% image size has been found empirically.
Method II Reset
a) The temporal method is used but at every 10
frames we do a complete image computation.
Method III Size Reduction
a) Image size is decreased by half before doing
the complete image computation. This operation is
repeated on every images.
Method IV Temporal + Size Reduction + Reset
a) Image size is decreased by half before computing
the vanishing point research.
b) For first image, computation of vanishing point
research is done on the complete image.
c) For subsequent images, the best vanishing point
candidate is only selected from a 10% by 10% image
size zone around the coordinates of the vanishing
point previously found in order to refine the vanishing
point position.
d) Every 10 images, the computation is done on the
complete image.
Method V Downsampling
a) Image size is decreased by half before computing
the vanishing point research.
b) Image size is then increased and the best vanishing
point candidate is only selected from a 10% by 10%
image size zone around the coordinates of the vanish-
ing point found previously on the same but smaller
image in order to refine the vanishing point position.
c) These operations are done on every image.
Method VI Temporal + Downsampling + Reset
a) Image size is decreased by half before computing
the vanishing point research.
b) Image size is then increased and the best vanishing
point candidate is only selected from a 10% by 10%
image size zone around the coordinates of the vanish-
ing point previously found on the same but smaller
image in order to refine the vanishing point position.
c) The following image size is decreased by a ratio
GeneralRoadDetectionAlgorithm-AComputationalImprovement
827