We recently had some positive experiences using lines instead of points
for the relative orientation of a stereo model. Contours of features were
digitized monocularly on both diapositives. Since the digitized points are
not conjugent points, contours of features were used to orient the two
photographs. The method developed allows US to find the conjugent feature
contours within a set of candidates. The two conjugent lines do not
necessarily begin and end at the same location in both photographs; the
best matching part has to be found that represents the con jugent line pair.
By reducing the line pair to the center of gravity and by using invariant
moments, the orientation can be computed very accurately and reliably.
The matching algorithm used is independent of scale and rotation.
Presently, we are applying the same technique to match segments of zero
crossings as opposed to matching single zero crossings in a scan line only.
After convolving, the Image zero crossings are computed to subpixel
accuracy. The convolution values are considered as discrete points of a
surface. Hence, the zero crossings result from intersecting this surface
with a horizontal plane at a convolution value of zero. As demonstrated by
Hildreth /4/, the linear interpolation between convolution values of
opposite sign lead to subpixel precision, That probably accounts for the
hyperacuity of the human visual System. The resolution of a foveal
receptor is approximately 0.4 minutes of arc; on the other hand, stereo
acuity can be as precise as S seconds of arc.
The objective is to find automatically a sufficient number of points for
orienting digital images as obtained, for example with a 3 line strip
camera of MBB (see /5/) without the need for the approximate locations of
the conjugent points. The method as outlined above will render con jugent
points with subpixel accuracy, obtained in an iterative manner from coarse
to fine channel zero Crossings. Since a large number of points can be
determined, additional parameters in the orientation procedure will
compensate for the model deficiencies caused by the atmosphere, lens
distortions and imperfections of the imaging process.
- 606 -