54
1
IEEE JOURNAL
OF
ROBOTICS AND AUTOMATION,
VOL
4,
NO
5,
OCTOBER
1988
Structured Light Patterns for Robot Mobility
JACQUELINE
J.
LE MOIGNE
AND
ALLAN M. WAXMAN,
MEMBER,
IEEE
Abstract-In order to assess the feasibility of using a structured-light
range sensor for mobile outdoor and indoor robots,
we
discuss a number
of operational considerations and image processing tools relevant to this
task domain. In particular,
we
address the issues of operating in ambient
lighting, smoothing of range texture, grid pattern selection, albedo
normalization, grid extraction, and coarse registration
of
image to
projected grid. Once a range map of the immediate environment
is
obtained, short-range path planning can be attempted.
I.
INTRODUCTION
GOAL OF computer vision is to endow machines with a
A
sensory capability
so
that they may perform their
assigned tasks with some degree of autonomy. Among the
many visual skills considered desirable, one of the most useful
is the ability to determine the ranges of objects in a scene.
Range information can be exploited
in
a number of different
applications such as object recognition, object acquisition by a
manipulator, and robot mobility. As these tasks are rather
different, one should expect that the type of range data
required will also differ with regard to sampling frequencies
(in
space and time) and resolution of range texture. Moreover,
the different task environments demand different tools in order
to acquire the data. A variety of such ranging techniques have
been described
in
a recent review article
[I].
Most methods are
targeted for implementation on industrial robot arms; for
example, the one
in
operation at the National Bureau of
Standards
[2].
Our work deals with the development of an inexpensive
ranging sensor which could be used
in
the “short-range
navigation” task of a mobile robot. The purpose of the sensor
is to enable the robot to construct a topographic map of its
immediate environment to be used for planning a path over the
terrain while avoiding obstacles. The desire for an “inexpen-
sive” system requires a minimal dependency on sophisticated
hardware: thus our approach is
‘
‘software-intensive.
”
One
method often invoked to accomplish this task is stereo vision,
which is a passive, bi-static triangulation mechanism. By
comparing features
in
two images of the same scene taken by
Manuscript received February 2, 1987; revised February 22, 1988. Part of
the material in this paper was presented at the Seventh International
Conference
on
Pattern Recognition, Montreal, Que., Canada, July 30-August
2, 1984. This work was supported by the Defense Advanced Research
Projects Agency and the
U.
S.
Army Night Vision Laboratory under Contract
DAAK70-83-K-0018 (DARPA Order 3206).
J.
J.
Le Moigne was with the Computer Vision Laboratory, Center
for
Automation Research, University
of
Maryland, College Park, MD 20742. She
is
now
with Martin Marietta Laboratories, Baltimore, MD 21227.
A. M. Waxman was with the Computer Vision Laboratory, Center
for
Automation Research, University of Maryland, College Park, MD 20742. He
is now with the Laboratory for Sensory Robotics, Department
of
Electrical,
Computer and Systems Engineering, Boston University, Boston, MA 022
15.
IEEE Log Number 8822827.
two cameras separated by a known baseline, one may
determine the ranges to these features from their measured
disparities in a simple fashion. The major difficulty encoun-
tered, however, is the so-called “correspondence problem”:
identifying features in the “left image” which correspond to
those
in
the “right image.” Sorting this out computationally
can be extremely time-consuming, rendering
it
unfit
for the
mobility task. One can alleviate this problem to a great extent
by going to an active, bi-static triangulation system based on
the concept of “structured light.” Here, one of the stereo
cameras is replaced by a light source which projects a known
pattern of light on the scene. The remaining camera then
images the illuminated scene from a different vantage point.
The range information manifests itself
in
the apparent distor-
tions of the projected pattern. The vision system must then
extract the pattern from the scene, compare
it
to the known
projected pattern
in
order to assign disparity measures, and
thus recover the range information. Depending on the choice
of projected pattern, one may still have to solve a correspon-
dence problem between the projected and perceived patterns.
If one projects a single spot or line of light onto the scene, then
no correspondence problem arises: however,
it
is then
necessary to scan the projected light over the scene to build up
a range map
[3].
Such scanning devices can make a system
expensive, particularly if they are to be sufficiently rugged and
reliable for the mobility task. The alternative is to project a
grid of points or lines on a scene
in
order to cover the entire
field of view of the camera. One is then faced with a much
simpler correspondence problem to solve, essentially to label
the grid points
in
the imaged pattern according to their
coordinates
in
the projected pattern. A particularly clever
method of doing this for a pattern composed of an array of
points is to encode the labels by modulating the individual
beams over time
[4].
This, however, assumes that the images
remain sufficiently registered over the required timescales. In
our case, we wish to use temporal modulation to enhance the
signal-to-noise ratio (as discussed
in
the following section).
We have chosen to project a grid of horizontal and vertical
lines, along with several dots, onto the scene which is then
imaged by a camera separated from the projector by a vertical
baseline (cf. Fig.
1).
The projection of a grid pattern has been
described previously
[4],
[7] but we added dots to this pattern
which are used as landmarks to initiate the labeling process. In
the following sections we shall describe the considerations
necessary for operating such a system
in
ambient lighting
(indoors and outdoors) and selecting the geometry of the
projected grid. We utilize the stereo ranging formulas derived
in
[7] but extend them to derive a formula relating the
smoothing of range texture to the thickness of the grid lines.
0882-4967/88/1000-0541$01
.OO
O
1988
IEEE