nbul 2004
e time we
image for
candidate
han some
. The size
nowledge
periment).
a certain
mage 2 is
ate match
pport
pair in T
ach other
ity" and
continuity
xisting in
refers to
d pair. Or
candidate
matching
e are less
lefined as
candidate
upport is,
points
For each
wo set of
distance
shift), If
2 equals
3s Q, )
og with
ult, given
5)
atch, P
7
1 P and
easure of
2
7)
distance
distance
? average
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
dist(P. P0 Q0. [d P. P) *d(Q,,Q,)]/2
—óg (h,k)/ €,
(6, (hk) = {°
If ( 9, , P ) isa candidate match and 5, (A, k) «&,
Otherwise
Where £, is the threshold of relative distance difference, its
empiric value is applied in calculation and 0.3 is applied in
paper. Since the candidate matching points Q of P. are more
than one, the numbers of d | Ó. j (A, " ) values are more
than one; only the m is applied
as support of point Pp anf "its M Dn match point
pair (P. ) In actual calculation, the points in the neighbour
domain of P are more than one, if N( P.) is used to express
the point set (excluding p ) in the neighbour domain of
point P, calculates the support of point pair (P, Q ) i
N (P ) one by one, then the average after accumulative serves
as generally initial sert
SUP O, ye vai @(|0,(h,k)|) (8)
m hei kr)
Where m is the number of points in N (P )-
When calculating S^? (P. o ) > treat each point pair D Q, )
initially and equally because there is no priori knowledge at
beginning. But after the r time iterant (r > 0) , the support of
(P. O0 on (2.0% relies not only on the difference of
location between P and 0 ; but also on their
sr (PO ) value, namely allowing feedback of local
support. These two factors can be combined together as a
different way; the minimum value is selected, therefore:
2.0 )= Dax mint IS (P. O Y oo. (^, k)D1(9)
7
the iterative until expect for p the measure of support for less
then known threshold value.
3.4 Sequence, stereo double matching restraint
Stereo — correspondence and sequence correspondence
simultaneously exist in the three-dimensional feature
correspondence movement analysis. The method and goal of
these two matches are identical from the aspects of image
processing. During the process of object three-dimensional
feature point’s correspondence, both stereo-sequence match and
sequence-stereo match can be applied. However, the different
matching order will has different matching effect to the final
moving object three-dimensional feature point correspondence.
In actual operation, the adjacent images took by one camera are
similar, because the time interval between two adjacent images
is very short. Thus, the sequence image match from same
camera is easy, but because its base line between viewpoints is
relatively short, three-dimensional reconstruction is difficult.
Therefore, the estimated depth is not precise in the situation
When noise exists (it is even impossible when baseline is fairly
Short). With this correspondence, there is usually a certain
baseline between different cameras, the three-dimensional
reconstruction precision is fairly high among stereovision
because the distance between viewpoints is large, but stereo
matching is difficult, especially when huge disparity and image
distortion exist. The double match restraint namely first extracts
random feature points on moving objects from different-time
but same-sequence images, determining the corresponding
relationship of feature points among the same sequence images,
establish the sequence match of binocular image sequences. If
the image sequence sample density is appropriate, the reliability
of the feature match in sequence image can be guaranteed. Then
according to match corresponding point coordinates which
obtained from the sequence match result matching with
corresponding images of same-time different sequence (left and
right images stereo match). Therefore, the difficulty of random
stereo match can be decreased to a great level through this
double match restraint. As a result, the whole correspondence of
moving object random three-dimensional feature points can be
obtained preferably.
Double match restraint process is shown as Figurel. Where the
left and right image stereo vision system at time /; are Z and
I ; respectively, the corresponding feature points on image
I. and i are M, and M, , the three-dimensional feature
points on Ts object corresponding to M; and M, is
M, A HJ? b is rotational matrix and translation vector of
moving objects between time f; to time 7, , Roslin are
rotational matrix and translation vector between left and right
stereo vision camera, the match between feature points
m, and M, is stereo match (correspondence); the match
e een feature point ni, and m, or Hl, and m, 4 is
sequential (moving) match (correspondence) ; the
correspondence between feature points M ; and M id 15
correspondence of moving objects random three-dimensional
feature points. / — 3 in Figure 1.
In practical realization, when carrying on the preceding stage
match, the feature points are extracted from two images (front
and rear) simutaneously, and original matching table is created
between the features of two images, the possible candidate
match points of a feature is found in the other image. But when
carrying on the latter match, corresponding feature point is
searched on the other image for matching according to the
preceding match result (coordinates) in order to enhance the
computation speed and the match precision. Before the
movement sequence match, the image difference can be used to
examine the dynamic moving object, and limit the matched
object on moving object (adandoning the static background of
moving object).
Through the sequence-match and stereo-match, after obtaining
the pixel coordinates of corresponding binocular sequence
image feature points of moving object at different time, and
using the transformation relation of image and object
coordinates obtained from calibration, the object coordinates of
corresponding feature points can be obtained from formula (2).
Using these points sequences in the three-dimensional space,
the object parameters of movement are estimated.
