International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Lidar data
Intensity data Height data Optical imagery
Segmentation Segmentation
v Y
Road areas and open areas Grass land,
tree areas
Iterative Hough Morphologic
transform operation
Candidate Candidate
road parking «-
stripes areas
I I
* Y
Verified Verified
road stripes parking 4
areas Vehicle detection
Topology detection
Road
network
Figure 2. Workflow of integrated processing
for road extraction from urban areas
3. INTEGRATED PROCESSING
3.1 Segmentation of lidar data and high resolution imagery
We separate roads from trees, buildings and grasslands with
minimum misclassification fusing the intensity and height data.
In reflectivity, the spectral signature of asphalt roads
significantly differs from vegetation and most construction
materials. The reflectivity rate of asphalt with pebbles is 17%
for the infrared laser, and no other major materials have a close
reflectivity rate. In height, pavements are attached to the bare
surface and appear as smooth ribbons separating the street
blocks in a city.
It can be easily found that integrating intensity and height data
may produce reliable road detection results. On the one hand,
the intensity provides the spectral reflectivity, which can help
identify most roads even if the objects coated by the same
material are also included. On the other hand, the height data
can help identify most non-building and non-forest areas even if
those low open areas such as grasslands are also included.
Using height information, the built-up areas with higher
elevations than their surroundings will be safely removed; while
using the (first-return) intensity information, the vegetated areas
are easily removed. In detail, compared to roads, grasslands
have different intensity although they have low elevation, trees
have different values in both intensity and height, and buildings
have high structures with elevation jumps although they may be
coated rainproof asphalt.
After segmentation of the lidar data, the possible road areas and
other areas are converted to a binary image. Figure 3 shows the
segmented data. Parking lots are kept because of same
reflectance and low heights as roads, and bridges and viaducts
are removed because of their large heights.
From the true colour high resolution imagery, the grass lands
and trec areas can be separated from the open areas. First,
because the roads and parking areas are covered and coated by
concrete or rainproof asphalt, the saturation of the pixels of the
areas is low while in the grass lands and tree areas it is high and
the hue tends to be *green'. So using a threshold the grass lands
and tree areas can be separated from the low saturation areas.
Subtracting the grass lands and tree areas, we can obtain the
areas containing candidate road stripes and parking areas.
- =.
=z
Figure 3. Extracted open areas (white)
containing road stripes and other areas
3.2 Extract Road Stripes by Iterative Hough Transform
The streets demonstrate ribbon features in geometry. We used a
modified Hough transfer method to directly detect the candidate
stripes of the streets from the segmented lidar data — the binary
image. Hough transformation is frequently used for extracting
straight lines. When we treat a ribbon as a straight line with the
width of the street, traditional Hough transfer can be used for
the detection of the streets. Figure 4 shows the Hough space
after once transfer. The space is formed using the straight line as
given by:
p = xcos0 + ysin O (1)
where Ô is the angle of the line’s normal with the x-axis; 0 is
the algebraic distance from the origin to the line.
Instead of detecting the peak points in the transfer space, we
detect the ‘maximal bars’ as pointed out in Figure 4. To detect
all possible ribbons, first step is to determine the primary
direction of the street grid. The parallel ribbons and ribbons
with right angle crossing to them are also extracted. The
extraction is conducted directly from the segmented binary
image on contrast to extraction from ‘thinned’ ribbon, and the
width can be estimated roughly by the bar width (the difference
of ?). We iteratively carry out the Hough transform. In each step
of transform, we only detect a maxima response in the Hough
space, and then the extracted stripe pixels are removed from the
binary image. This will reduce the influence of multiple peaks
in the transform space. The iteration will be terminated by the
trigger criteria of the maxima that indicates the length of the
stripe.
Interna
———— ———
—
(b
3.3 Vei
The de
Streets ;
real str
some v
lines o
segmen
simply
segmen
in the s
building
We jud
areas ai
high po
we emp
extracte
of vehi
used fo
possible
shown
study, t
à regio
possibil
applied
Figure |
than ro:
the ana
the opti
being a
of bein
which ¢
