871
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
cloud can be generated, and stereo image sequences also can be
registered for mapping.
Laser Point Rendering and Image Positioning
For better fusion processing, we should know the corresponding
relationship between any point of laser 3D point cloud and
image sequences, which is called as image linking. Based on
the linking information, it is easy to know which images are
including the appointed point of laser point cloud; and the
position of any place of image by back-projected laser 3D point,
which is called as image positioning. When the linked image
has been calculated, the color information of that point also can
be acquired from that image, which is called as laser point
rendering. The point rendering technology gives color
information to laser 3D point cloud so that laser point cloud
based extraction can be performed well because of additional
color information.
For laser point rendering, it is possible that a laser point is
covered by foreground object in that image so that error color
information is acquired. Our solution is using stereo image
checking. As you know, if a point is covered really by
foreground object, the two projected points in image pair
become non-matchable. Based on this checking, the color
information can be searched in round stereo image pair until the
two projected point can be matched (Detail in Shi 2008).
using shape of laser data. The boundary point can be easily
done if the road have curb. Because of the road curb, if height
of a laser point suddenly changes, the boundary point is
detected. But for these no-curb roads, the next processing-
image and laser fusion —is used for boundary point detection.
The detailed description for automatic extraction of road
boundary and road mark is reported in the reference paper (Shi
2008).
Image and Laser Data Fusion based Final Robust
Extraction
Like road sign, the fusion method can detect it more robustly.
Road signs use particular colors and geometric shapes to attract
driver’s attention, so typical algorithms of image based
detection and recognition use their inherent color, shape and
texture to detect and to extract sign type and legend. On the
other hand, road sign also stand alone in roadside, so it has
inherent features in laser 3D space. Based on these features in
laser 3D space, road sign can be easily detected. This step fuse
laser data and image to detect as successful as possible.
3.3 Success Extraction Ratio from Our Experiments
In 10km experiment data, success extraction of road boundaries
reaches to 95%; but for traffic mark, because of no-capture
image data on curving place of road, the missed ratio reaches to
about 13.4%. For reducing the miss ratio of traffic mark, system
should capture images on curving place as dense as possible.
For seeing effectiveness of fusion method for road sign
recognition, we compare the fusion method with image method.
(1) Fusion method can detect most of road sign (more than 94%)
especially in sun day, in which image method just have about
51% success ratio because of the effect of sun-shining.
(2) In cloud condition, the recognition ratios of image method
and fusion method are more than 90% too, but in sun-shining
condition, fusion method has more recognition ratio than image
method. Although the ratio is 71.4%, it is easy to check its type
by extracted image (its extracted ratio of fusion method reach to
94.3%).
Whether
Type
Existing
RS
Detected
Recognized
Cloud
Image
161
153
(95.03%)
145
(90.06%)
Image&
Laser
161
157
(97.5%)
148
(91.93%)
Sun-
shining
Image
35
18
(51.4%)
16
(45.7%)
Image&
Laser
35
33
(94.3%)
25
(71.4%)
Table 1. Success ratio of road sign extraction
Figure4. Flowchart of Fusion Processing
for Automatic Road Object Extraction
Laser Point Cloud based Candidate Extraction
Rendered Laser 3D Point Cloud not just has color information,
but also it is seamless 3D data. Just because it is seamless, the
extraction of linear object, such as road traffic lane mark,
becomes less costly with “No-mosaic”.
Typically, road surface is flat but pavement surface is higher
than road surface. The feature just makes extraction possible by
4. CONCLUSION
In this paper, we aim to satisfy with the high demands for the
road spatial data. We presents high efficient road mapping
technology by fusing vehicle-based navigation data, stereo
image and laser scanning data for collecting, detecting,
recognizing and positioning road objects, such as road
boundaries, traffic marks, road signs, traffic signal, road guide
fences, electric power poles and many other applications
important to people’s safety and welfare.