feed the
JN
f surface
aditional
n into a
eoscopic
t under-
ce point.
rlapping
nt in the
A recent
93). The
tially in
natching
/e seen a
speed of
of radio-
e pull-in
n if the
re large,
ions, the
rarchical
ures, e.g.
e match
d radio-
eduction
straining
jed into
bout the
en metric
imeras in
3-dimen-
lements;
ated by
an urban
vith geo-
as (radar
orted as
ar image
-infested
| sharply
serve as
uminated
2.3 Verification and Quality Control
The most significant limitation to successful stereo-
scopic machine vision is one's ability to reliably
verify and accept a surface point. The effort to deter-
mine that a match point is consistent with the expec-
tations and with neighboring match-points often ex-
ceeds by far the effort to obtain a match at all. The
success in identifying match-points is traditionally
being improved by looking for matches at selected
image locations obtained through interest operators.
Yet, these concepts have not been able to produce
working stereoscopic machine vision systems that
will operate as reliably as a human operator can, except
for very well constrained scenarios.
3 OBJECTS IN SINGLE IMAGES
3.1 An Applications Scenario
The advent of computer graphics and digital enter-
tainment as well as a refined need for disaster pre-
paredness, training and simulation, planning and citi-
zen participation is creating a market for fully 3-
dimensional geographic information ^ systems of
urban areas. The ,CyberCity“ needs tools to recon-
struct roofs and buildings (Gruber et al, 1995). It is
not necessary to rely on stereoscopic machine vision
to reconstruct buildings and building boxes from
photographs. Braun (1993) has shown that the
knowledge that a building's walls are vertical and that
rooflines typically are horizontal can be exploited
when using a single aerial photograph. This process of
identifying building boxes and attaching roof shapes
from single images revives traditional geometric con-
cepts that were originally developed at the turn of the
century. At the current time, however, buildings can be
extracted automatically from single aerial photo-
graphs or under minimal manual control.
3.2 GIS and Aerial Images
In industrial urban areas one can depend on the
availability of a 2-dimensional geographic informa-
tion system in which the third dimension may be
available as an attribute. Therefore one will know the
footprint of buildings and one will have an elevation
measure attaclied to each building. This can be used to
project the footprint and a prediction of the roofline of
a building into an exterially oriented aerial photo-
graph. The photograph itself then serves to first verify
the geometric information in the GIS and second to
improve the geometry and detail of its information.
Gruber et al. (1995) have shown a process by which a
2-dimensional GIS and single images can serve to
develop building boxes in a process that is similar to
the work of Braun (1993), but includes the informa-
tion of the GIS and is automated.
3.3 Structured Light
Of course the application scenario with ,,CyberCity" is
not the only one in which constraints about the ob-
jects permit one to rely on single images. They can
also be used when combined with a proper illumi-
nation system to reconstruct surface shapes. This
classical approach in machine vision is used in in-
dustrial applications. Structured light may employ a
sequence of parallel planes of light and dark that are
projected onto a surface. If it is plane then the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
intersection with the planes of light result in straight
lines. Deviations from straight lines are a measure of a
deviation of the surface from a plane. Successful
applications of structured light sometimes also com-
bine it with stereoscopy. This may on one hand help in
resolving ambiguities, and it may on the other hand
improve the accuracy. Structured light's limitations
may be the difficulty of obtaining well-defined transi-
tions between light and dark.
4 RANGING
4.1 Laser Ranging and Radar Altimetry
Ranging is applicable in the robotics environment
when the distance to an object needs to be known. The
planetary guidance of a landing spacecraft may use
ranging to determine the absence of obstacles. An
automated vehicle may determine its distance from an
object by laser ranging. Systems exist that will use a
single point source and scan the object along profiles.
Multiple sources may be used to scan multiple
profiles. The concept of laser radar“ measures the run-
time from the source to the object and the direction in
to which the energy was transmitted and from which
echoes were received. These ideas are also the topic of
NASA-concepts for a satellite-based topographic
mapping system. The result of a laser range finder is a
collection of surface points that are irregularly spaced
and potentially noisy. Ranging of course is also the
subject of altimetry. This is the traditional planetary
tool to determine the shape of a surface. On Earth
altimetry addresses the shape of the water bodies. The
footprint of a radar altimeter is fairly large and one
typically assumes that the first echo to arrive at the
antenna is from the nadir. This assumption is only
correct if the surface is smooth. In accentuated terrain
the first echo may be from a point off to the side.
Planetary altimetry suffers from ambiguities such as
those obtained on Venus with the Pioneer and
Magellan missions, when observations are made in
mountainous terrain.
4.2 From Points to Surfaces
Ranging as a data collection mechanism needs to be
complemented by a resampling technique that con-
verts the surface points to a continuous surface. Two
issues exist: first, that the noise in the data be filtered
and that a smooth surface results from a rough point
cloud; second, that a data structure comes into
existence that actually represents the surface as
opposed to an unstructured collection of points. The
transition from points to surfaces is trivial if the
underlying object is smooth and if the object re-
presents a 2.5-dimensional situation. The problem
becomes difficult when a fully 3-dimensional situa-
tion exists. We will discuss this further in Section 9.
5 SHAPE FROM SHADING
5.1 Basic Idea
Shape-from-shading is a traditional technique of ma-
chine vision to obtain information about the local
surface shape given the variations in image bright-
ness. It has traditionally been applied in a controlled
illumination environment, e.g. when a robot has to
select among various objects presented in a bin.
