)079, China
mapping
'he camera used
'omputer vision.
moved. But, the
nsists of relative
g technique and
est the proposed
uipment or have
computer vision
wlo carried out
building using
| accurate point
jer information.
| some of which
lon, automation
ch need to solve
1 which a robust
Qu goal is to
ct with image
s process, only
> and all other
until 3D model
econstruction of
' or indoor. It is
Is described in
IBRATION
ocess is image
king photos and
a scene picture.
itomatic image
age station, that
aseline is short,
is not a serious
computers. As
F a coal pile, 83
n neighbouring
jects or scenes,
ced to retrieve
le not same size
is needed and a
netric camera is
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
used in this method and its interior parameters are known. And
the whole process of obtaining photos' exterior parameter is the
same as traditional aerial photogrammetry, including the
automatic obtaining of orientation points and tie points (link
point), relative orientation in succession, model link,
construction of free flight strip and free network adjustment
with bundle method. Because we take photos with short
baseline, orientation points and tie points can be obtained from
automatic matching, and detailed algorithms and formula of
model connection and strip construction can be seen in
photogrammetry manual. After the process above, photo
parameters are obtained in self-define coordinate system. Fig. 1
is the sketch map for all image stations automatically drawn
with photo parameters by program after adjusting. The photos
just form a circle.
Fig. 1 Orientation of Coal Pile Photos
Objects N
RA hin
V nox
[mage station
— md V e
Base line
Fig. 2 Images used for matching
3 MULTI-BASELINE STEREO MATCHING AND
SURFACE RECONSTRUCTION
3.1 Multi-stereo matching
Image matching is a basic and crucial process for automatic 3D
reconstruction. But to get reliable and robust matching results is
still very difficult because of following problems existing in
images: (1) Radiometric problems: resolution, reflectance,
illumination, lab processing noise, digital camera noise; (2)
Geometric problems: relief displacement and occluded areas,
projective deformation, scale variation; (3) Textural problems:
featureless surface, repetitive texture, ambiguous levels such as
tree top and ground below them, thin objects.
In this paper, an image matching method based short-baseline
and multi-photo has been developed, as shown in Fig. 2.
Obviously, the geometry distortion of the objects in images with
short-baseline is little. But it is known that the intersection
accuracy is low when the baseline is short. So we use
multi-photo intersection to maintain the accuracy as shown in
Fig. 2.
Because geometry distortion in close-range photography is
relatively large, traditional single-stereo matching which uses
only two images is very difficult to meet the demand of
matching in reliability and accuracy. The multi-stereo matching
method which uses multi images and combines with short
baseline and multi-photo perfectly solves the image matching
and intersection accuracy problem at same time. This method
has following characters: on one hand because the baseline
between the neighboring photos is relatively short, the geometry
distortion of images is relatively little, thus help automatic
matching; on the other hand, because baseline is short and multi
photos are used, overlap between the neighboring photos is
normally very large, we can obtain the corresponding points
with multi overlap by matching transit using corresponding
points in neighbouring photos. The corresponding points pass
constantly through neighboring photos until they can not match,
thus each 3D point have multi corresponding 2D image points
as shown in Fig. 2. In Fig. 2, it also can be seen that the further
the object to be measured, the small the intersection angle, so at
this time we use more images to intersect when calculating the
3D space coordinates. The farther the object, the more images
arc used. Obviously, there are lots of redundant measurement
for each group of corresponding points, if obtaining the weights
of measurement by iteration method with variable weights and
calculating using bundle adjusting, the reliability and accuracy
of the coordinates of the model points will improve
significantly.
3.2 Surface reconstruction
After image matching, 3D points can be calculated with image
matching results and image parameters. Then these 3D points
are used to construct triangular network, DEM and contour,
then sometimes make epipolar image and orthophoto, finally
produce 3D landscape map. Here some questions need us pay
more attention when making orthophoto. Because the DEM
involves a series of photos, we must correctly choose and
resample the appropriate photos to make sure that the
orthophoto keep unanimous on color tone.
4 RESULT OF THE EXPERIMENTS
We carried out many experiments to test our 3D reconstruction
algorithm. Some test results are shown in Fig. 3. In Fig. 3, (a) is
a photo of a plaster statue and (b) (c) are its 3D reconstruction
results; (d) a sculpture, (e)(f) its 3D reconstruction results; () a
photo of a relief and its 3D reconstruction results in (h); (i) a
photo of a coal pile and its 3D reconstruction results in (j) . The
plaster statue is relatively small, we took 5 photos; the sculpture
is 3 meters wide, 5 meters high, we took 6 photos; the relief is
12 meters wide, 6 meters high, we took 12 photos; the diameter
