Ul 
In view of vector analysis of the two criterions, the obvi- 
ous disparity betwen them can be used for rejecting some 
unreliable results. Hence, the two criterions are combined 
as a complete algorithm: 
1 
p,z SA ;—B;|=mim 
sv M 
Gam S YA; + B, max (25) 
i1 
where "AND" means both of the two conditions must be satis- 
fíed. In practice, tolerating computation rounding errors, 
the parallaxes obtained from each criterion differing within 
one pixel size will be incorporated into success and take 
the everage. If a gross displacement is fund from the two 
criterions, then thise target point should be treated as a 
failure point. The parallax of the failure point will be in- 
terpolated from the nearest successful points. Though the 
interpolated value has lower precision, that is much better 
than the wrong value. 
In considering of the poor-information area, it is.better 
to give an aditional condition: 
O4Bz7U 15 (26) 
or to rewrite (25) into 
i=1 
p,=D}|A;—B,|=min 
evo 
dA . B; 2 max20,, (27) 
i=1 
whereJtn is a given threthold, depending on the noise level. 
The computation load of (27) is heavier than (13) or (25), 
but less than correlation coefficient algorithm(15). Anyhow 
the reliability of such dual criterion is better than any of 
the three. 
(b) Sum of absolute differences using both the original and 
the first derivated signals, e.g. 
9,7 321A, — B,| ^ min 
i=1 
n-i 
AND {4 D,= A;— {=m 
1 =a i AB, | min (28) 
"»-1 
D,z 50144412014 
$41 
Comparing with (27), this algorithm make further reduce of 
computation load, because there are pluse and minus calcula- 
tions only. 
(c) Triplex criterion 
If necessary, triplex criterions can be used to further 
- 421 -