7A-4-5
information on the multiple image matching method, interested
readers can refer to Tao et al., (1997) and Tao (1997).
In the current implementation, the search window size is set as
25H x 17V pixels, considering the georeferencing errors. The
correlation coefficient threshold is set as 0.70. If the number of
match points is more than half of the total points applied, the
corresponding line segment exists. Since the match points may
not be distributed along a perfect line, a line approximation
algorithm using the least median square criterion (Rousseeuw,
1987) is used to filter out the spurious points and determine the
best line.
The above method is also applied to the other consecutive
images (forward and backward images) to find
correspondences. Finally, a straight-line constrained
photogrammetric triangulation can be employed to obtain a
more reliable and accurate result of 3-D line reconstruction.
This feature triangulation algorithm has been described in a
number of publications (Mikhail, 1994; Wilkin, 1996).
Current image pair
Figure 7. Multinocular line matching
5. TEST RESULTS AND EVALUATION
A batch processing method was implemented to detect the
transportation objects based on a map database. A stereo image
sequence, consisting of 729 images was used for the test. The
images were collected by the VISAT system in the city of
Laval. Totally, 56 desired vertical objects, mainly including
light poles and traffic signs were chosen and projected onto the
image sequence. After the processing of verification, 52 objects
were verified to exist using the method described above, and
four objects were not able to be detected. Based on the results of
manual check, one object was actually missing and the rest
three were blocked by other objects (moving vehicles). This test
demonstrates that the developed method is able to achieve a
successful rate of 94.6% (=53/56).
A typical example of the whole process is given in Figure 3a -
3e. The Figure 3a is a stereo image pair. There are four vertical
objects of interest (marked in the image). Object No.l is to be
examined. Figure 3b is the output of edge detection using the
six directional masks. All the possible vertical edges have been
detected. Figure 3c shows the effect of edge thinning. It can be
seen that a very large number of undesired edges have been
filtered out, while the vertical edges have been kept. A refining
process is conducted using the Sobel gradient operator. More
accurate direction values of existing edges are calculated and
only those edges whose directions are vertical with a tolerance
of 15° remain. As we shall see in Figure 3d, the edge points of
road boundary in the left image are mostly eliminated since
these edges are not vertically oriented. The effects of line
formation are shown in Figure 3e. Since we are interested in the
distinct features, only those line segments longer than 50 pixels
are recorded in the line file and displayed. This result
demonstrates that the new line grouping algorithm functions
very well, and is able to bridge the gaps to form a line segment
from edges. In this example, the left-bottom line segment was
chosen to perform feature correspondence and multinocular line
reconstruction. The position information of this line segment is
used to update its database.
Other examples are given in Figure 8 and Figure 9. The desired
features were detected successfully. The processing windows as
well as extracted line features have been overlapped in these
images. For a comparison, in Figure 9, the entire images were
processed using the line detection method described above. It
can be seen that the use of predicted windows generated using
the existing map databases will largely eliminate the
ambiguities in the process of line detection, i.e., line features
existing outside of the processing window will not affect the
final results.
6. CONCLUDING REMARKS
A new approach for the verification and updating of transportation
objects with vertical line features using road image sequences is
proposed. Based on the evaluation of practical testing results, the
following conclusions can be drawn:
• Using multiple constraints developed from the mobile mapping
system, the approach is able to detect and reconstruct the
transportation objects in a highly automatic and fairly reliable
manner.
• The prior position information of the desired objects derived
from the existing databases greatly simplifies the detection of
objects and makes the approach much reliable.
• Abundant image information of the objects offered by the
georeferenced multiple image sequences allows a robust
implementation of the object extraction and reconstruction
algorithms.
• It has been demonstrated that the developed algorithms of line
feature extraction, line grouping, line feature correspondence,
multiple image matching, and multinocular line reconstruction
functioned successfully.
This approach has gained a great attention of the user community.
Requests for a further test and licensing of the software is being
processed. A future research will be aiming at the detection and
recognition of arbitrarily shaped objects along road corridors. In this
case, color image information as well as the prior knowledge of the
desired objects froman existing database will be utilized.
REFERENCE
Chen, H. H., and T. S. Huang, 1990. Matching 3-D Line
Segments with Applications to Multiple Object Motion