; : I, X ed x! C edo va o ; I, PA x) Y S - y.)
1,3 DON, x) vit y j! i,j*1 Ex ey TL el^ y ji
I, (x, = x.) Waist Ye} I, (x, - x! ly, - y.)
Ded jx. e xol yu yu DM, Xue X, Myr my)
‘3 iu xy US "3 ie ed rr
and
Ap, ..v.:80:; ; = the corrections to the unknown parallaxes
1,3 1+1,3+1
When solving this system with the least square method we will get a
banded matrix. If we also add radiometric parameters it will become
banded-bordered.
Ordinary matching methods do not take the relation between the points
into consideration. With the multi-point approach we are capable of
strengthening the connection between the points by constraints formula-
ted as ficticious observations. The use of the discrete second deriva-
tive in the two directions will minimize the curvature
2p - D.
i4 909570 "n
i-1.3
And analogous in the y-direction. The structure of the normal equation
system is not affected by this constraint. Other constraints are pos-
sible. It is also possible to extend this method by using breaklines
and discontinuites where the position of these could be used as un-
knowns (Rosenholm, 1986).
DATA SNOOPING
The data snooping technique according to Baarda has been used both in
photogrammetry and in geodesy for the exclusion of gross errors from
adjustments. The technique is a statistical test of the size of the re-
sidual in the adjustments. With the least square matching technique it
becomes possible to use the data snooping in automatic point measure-
ments to exclude pixels with large radiometric differences between the
images. This was mentioned by A.W. Gruen (1985). The normalized resi-
dual is computed as
wd eoa, Ma
where
€ = the residual
o, = the standard error of unit weight
Quy = the weight coefficient of the residual
w = the residual normalised to unit standard deviation
When the absolute value of the normalised residual w is too large the
observation is rejected. When the standard error of unit weight is a
known quantity the normalised residual is considered to be normally
distributed. If the standard error of unit weight is independently
computed in another adjustment the corresponding distribution of the
normalised residuals is the t-distribution. When the standard error is
*1578 -
