004
by ROMA uses a simplified geometrical model, i.e. a surface To improve the number of matched points, an iterative method
r or mesh, image correlation and oriented photographs to determine that applies this principle has been implemented. Indeed, corner
ural 3D points visible on photographs and included in the mesh. detector can classify the points according to their degree of
tant interest. Usually only the best points are conserved. It would
ing The point p3 is the correlation The point P2 is used as not be judicious to take a larger part of points because it would
xt. felt es used as he Pl hunglogoys approximate value for raise the number of wrong matching. However we observed that
point for a computation of the 3D point IT1 vitae : e
rate ihe conelarion process if a point is repeated, his degree of interest has approximatively
iust the same rank in the other image. Then the algorithm can match
feature points progressively by rank.
ces] [EE 3.3 Combined algorithms
and
any Even though the feature-based matching is more robust than the
are area-based matching, it depends entirely on the repeatability of
her the detector. Another idea is to combine the two principles to
be ; avoid this drawback. For each triangle of the network of
es en. measures chosen by the user, a reference image is determined
We scan the current triangle. X - . J . z
ons The point IT is Ihe UR SZ (5) and the feature points are calculated in this area. Then
of in the current triangle. supposing that those points lie on the 3D-triangle, their
PI is the TI projection on to the Mis projected as I} homologous points are found by correlation in the other images
photo 1 on to the image 2 through the I-MAGE process.
If this method provides many successful matching, the
The | — - repartition of the points can be inhomogeneous on some scenes
lata Figure I. Principles of ROMA if interest points are clustered.
een We use four steps in this Semi-automated Primitive A fifth algorithm is based on the same structure as the previous
can Measurement Method, considering that a mesh has been 9ne but it forces the points to be regularly disposed. As in the
our measured and computed from a set of 3-D points visible on at first method, a regular set of points is created on the surface of
een least two images: the triangles. Then those points are projected on a reference
cts image. But instead of matching them directly, best feature
can Y For each triangle of the mesh we scan triangle and get points are calculated in their neighbouring and the homologous
ust point []. Each point [[ is projected as pl on to the ^ points of these new points are searched in other images.
and photograph 1;
e Y []is projected as p2 onto the second image; 34 3D Viewing
Point p2 is used as an approximate position to initiate the In order to visualize the reconstructed surfaces of the scenes, a
ave area based correlation process with pl; VRML export file can be generated. Thus, a triangulation of the
ard > P : fhe Comelation: nl. and. ; 3D-points including the initial network and the collection of
Sus Point p3 is the result of the correlation; pl and its now measures is necessary. Since some points are too close to
on homologous p3 are used for the computation ofthe 3-D co- each other for a good viewing, a part of them is eliminated. A
Be ordinates of [ [1 local optimization deletes the points which disappearance
This first implemented algorithm generates automatically Would modify the Structure of the surface, according to the
regularly 3D-points through area-based matching. Initially a set Chosen criteria presented m (Schroeder et al., 1992). Then
S of points was measured by the user. This collection of points is relevant images are projected on the triangles of the grid.
i triangulated and the regular scanning of the 3D-surface of each
oi triangle provides theoretical 3D-points which are projected ona 3-5 Results
Ose reference image. The semi-automated Primitive Measurement ; ;
the process called I-MAGE supplies measured 3D-points thanks to The five implemented-algorithms have been evaluated and
re- ar ; " compared using different photographed scenes. Some objects
automatic correlation with other images. ; : =,
the As a geometric-construction-based method, it gives a regular with a well-known simple geometry visible on the photography
ges m tn Shame n a ; were used to estimate the accuracy of the methods.
g grid but the correlation process tends to fail when it works on
sed low-textured image zones.
ing 32 Other Simple alçorithins Methods Advantages Drawbacks
hat Area-based Many points Wrong matching on
[in On the contrary, the second algorithm uses the featured parts of matching Good repartition of | low-textured images
images. Firstly, interest points are extracted by Harris’ feature points and problematic
improved algorithm (Harris & Stephens, 1988) considered as images
ent the most efficient according to (Schmid et al. 2000). Then the : :
homologous points of the reference image interest points are Feature-based | Good precision, Very few points
searched among other images. But the correlation process is matching especially with low-
the limited by two geometric conditions. On the one hand, out of textured images
the continuity constraint, when a point belongs to the projection of
yer, a triangle from the initial network, his homologous point must Iterative Good precision, More points than in
ga be in the projection of the same triangle in the other image. feature-based | especially with low- | the previous method
ese Moreover, each homologous point is checked by exchanging matching textured images
reference and search matrix.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B6. Istanbul 2004
125
LE 5
