We frequently encounter situations
where features appear and disappear from the image sequence. This problem
is known as
occlusion. The factorization method described in the
previous section cannot be applied,
per se, to solve this problem.
The image sequences yield a matrix W which has some unknowns values
(i.e., it is partially filled). Fortunately, we can explore
some information in the sequences that can allow us to find the unknown
entries of W. This is done by doing projections of the feature coordinates
onto camera positions.
Tomasi and Kanade [1], established a condition for reconstructing an unknown image point pair (ufpvfp ) in frame f as follows: we can reconstruct such point pair if for 3 given frames f1 ,f2 ,f3 (not necessarily consecutive) point p is visible and if we can find at least 3 more points p1 ,p2 ,p3 (again, not necessarily consecutive), which are visible in all the four frames (f, f1 ,f2 , f3). We will illustrate the methodology as follows. Suppose W = [U;V] is a 10 x 4 matrix, with U5 x 4 and V5 x 4 . This matrix has 2 unknown values denoted by X.
[u11
u12 u13 u14]
[u21
u22 u23 u24]
[u31
u32 u33 u34]
[u41
u42 u43
X ]
[u51
u52 u53 u54]
[v11
v52 v13 v14]
[v21
v22 v23 v24]
[v31
v32 v33 v34]
[v41
v42 v43
X ]
[v51
v52 v53 v54]
The first step is to use the factorization method over W8 x 4 . This matrix is obtained by eliminating the rows in W whose values are unknown in U and V, i.e.,
[u11
u12 u13 u14]
[u21
u22 u23 u24]
[u31
u32 u33 u34]
[u51
u52 u53 u54]
[v11
v52 v13 v14]
[v21
v22 v23 v24]
[v31
v32 v33 v34]
[v51
v52 v53 v54]
After factoring W8 x 4 , we get:
R 8x 3 = [i1T i2T i3T i5T j1T j2T j3 T j5T ]T
S = [s1 s2 s3 s4]
, which respectively represent the translation, rotation and shape generated from the submatrix W8 x 4 .
Therefore, by the factorization method, the above can be expressed as: W8 x 3 = R8 x 3S + t8 x 1e4T , where e4 = [1,1,1,1].
In order to have the full R (rotation matrix0, we need to compute i4 and j4, which are unknown. We first need to make the origins of i4 and j4 coincide by referring to the centroid c = 1/3 (s1 + s2 + s3) , where the indexes in s denote the points visible in all the four frames. In frame f , the projection of the centroid c (i.e., in vector t) has its coordinates:
a4 = 1/3 (u41 + u42 + u43) and b4 = 1/3 (v41 + v42 + v43) .
As can be seen, with these coordinates, we get the full vector t.
Now, we define Sp'
,U4p' and V4p' for p=1,2,3 (in this example) by subtracting
S,
U4p
and V4p (respectively) by their coordinates with
respect of their centroid (i.e., c, a4 and
b4, respectively). We find i4 and
j4
by
solving:
i4T [
s'1
s'2 s'3 ] = [
u'41 u'42 u'43
]
j4T [
s'1
s'2 s'3 ] = [
v'41 v'42 v'43
]
From the factorization method, we have that:
u44 = i4T
s'4+
a4
v44 = j4T
s'4+
b4
We now have all the missing information.
This method of reconstructing a point is
called row wise extension, since we propagated S over the
frames. There is another method called
column wise extension,
in which we propagate the feature points instead.
Next:Results and Discussion Previous: Perspective Approximation Contents:Shape and Motion from Image Streams