Home

Basic Contents

Computer Vision ... (Published 1992)

Simulated Human Vision..... Ian Overington

Location: Eastbourne. UK
ianoverington@simulatedvision.co.uk ............ www.simulatedvision.co.uk

local features such as single edges and corners. It has been shown by Hildreth that, in general, one
can
never determine the true flow of local portions of straight edges, owing to the fact that, if the
ends of the straight portion are outside the bounds of the local receptive field for analysis, there is no
way of deriving the tangential component of local motion (the so-called aperture problem). By
analogy, the same applies to measurement of stereo disparity, other than from knowledge of the
stereo baseline and trigonometric computations. On the other hand, if one has a local fragment of
an edge which is strongly curved then, provided that the curvature or orientation difference
information is known, it
is theoretically possible to derive the full flow or stereo disparity directly and
absolutely. Obviously the stronger the local curvature, or the larger the length of edge over which
an analysis is carried out (i.e. the larger the 'aperture'), the more reliable will be the measurement.
In the limit, the ideal features for such measurements must be corners. For such a form of practical
analysis one must therefore first address the supplementary question of how to
sense corners. It so
happens that the work described in Chapter 12, concerned with feature analysis on single frames,
provides a direct and simple method of such analysis of fragmentary edge data extracted from
VISIVE-like preprocessing. This analysis can code each pixel on an edge in terms of not only its
strength, orientation and sub-pixel position, but also the local edge curvature (in the input optical
image) and the locations of corners. The process can be instructed, amongst other things, either to
flag only corners or to flag all points where the local curvature is greater than a given number of
degrees/pixel. A similar form of analysis can be applied to the output from the paired-frame,
gradient-based analysis described in Chapter 5 for extraction of optical flow orthogonal to
fragmentary edges. From this, one can readily select just the points in a scene which are amenable
to direct true flow derivation. Equally, such a form of analysis can be applied to outputs of paired
frame analysis carried out on
stereo pairs.



Continued