that the second-best grid cells matching enhances the dense of
initial matching results, more abundantly expresses the object
details, and gets the homologous points in two overlapping area
and occlusion area in the multi-view image.
Figure 8. Grid cells of final matching results
3.3 Result evaluation
This paper makes forward intersection with the homologous
points, and obtains the discrete 3D points (Fig. 9). From this
figure it can be seen that there are a few mistakes. As shown in
Fig. 9, the ellipse regions includes the three error matching
points obtained in initially matching of the best grid cells, and
the seven error matching points obtained in matching with the
second-best grid cells. So it needs to adopt a reliable filter to
reject the mismatching results.
Figure 9. Discrete 3D points
3.4 Contrast analysis with different constraint conditions
The existing multi-view image matching object constraint
models, such as GC? algorithm (Zhang, 2005), modified vertical
line locus algorithm (MVLL) (Ji, 2008), constrain the image
space search range by points to be matched moving in the range
of an approximate height value along a certain linear direction
(projection ray locus, vertical line locus) in the object space,
and reduces the search range from two-dimensional to one-
dimensional. The traditional standard normalized cross-
correlation (NCC) is expressed as a function of height value Z,
in the approximate height value ranges
Z e[Z, — AZ, Z, + AZ]. The traditional algorithm calculates
the sum of NCC (SNCC) values iteratively, and selects the
height value Z corresponding to the maximum SNCC value as
the object correct height value. The formulas are,
2,0. G) - 4), G,G) - 1) (4
NCC;(p,,Z)- ET
(2,00 zh 12a SAL)
seW seW
1 n
SNCC(p,,Z) =~ NCC,(p,. Z) 65)
noa
Taking the point P in the image Ll as an example, and
according to the geometric constraint principle in GC? algorithm,
this paper restricts the height search range to 3-10m, and the
search step is 0.1m. The calculated cross-correlation curve is
shown as Fig. 10. The final calculated result shows that the
SNCC < 0.6 and obtains the maximum at the point p which
should be the point P . The reasons include (1) points to be
matched in the image LI are occluded in image L3, and the
NCC, 2 < 0.42 which is too low and (2) the NCC
L1-L2 at
point D and point p' appear peak values because of the
repetitive texture in image L2, which means the local grey
distribution at point P and point p' is similar.
Fig. 11 is the matching result using the MZPC algorithm. In the
matching of the second-best grid cells, it matches the grid cells
with number = 2 on the plane Z 2 7, and selects image L1
and image L2 according to the projection rays passing through
the grid cells, and avoids the occlusion in image L3. The ANCC
value is 0.94825602, and the correct matching result is obtained.
Comparing with the existing multi-view image matching
algorithm based on the object constraints, the MZPC algorithm
limits search distance in the image space under traditional
epipolar constraints or the search distance in object space under
geometric constraints to a grid cell position in the plane, which
avoids the appearance of multiple peaks in the cross-correlation
curve caused by similar texture, and reduces the probability of
mismatches. Simultaneously, the MZPC algorithm carries on
the selective matching for multiple images according to
projection rays in the grid cells to avoid the effects of
occlusions, and thereby improves the reliability of the matching.
n3 OF o: eet X y (ir
s NCC L142 : d R SNCC i W & NCC Li
rt ; “Ÿ
13456789591) 324586785i0 324538675910
Figure 10. Experiment results by projection rays constraint
matching
L1 | L2
Figure 11. Experiment results under moving Z-Plane constraint
Fo
cit
ref
vie
im
ael
mé
alg
(1)
sir
an
(2)
[Ji
the
(3)
of
cr
the
Ce
Re
Ji
au
In
Eı