Page 153 - Introduction to Autonomous Mobile Robots

Chapter 4
138
match quality for each pixel. This is valuable because such additional information can be
used over time to eliminate spurious, incorrect stereo matches that have poor match quality.
The performance of SVM provides a good representative of the state of the art in stereo
ranging today. The SVM consists of sensor hardware, including two CMOS cameras and
DSP (Digital Signal Processor) hardware. In addition, the SVM includes stereo vision soft-
ware that makes use of a standard computer (e.g., a Pentium processor). On a 320 x 240
pixel image pair, the SVM assigns one of thirty-two discrete levels of disparity (i.e., depth)
to every pixel at a rate of twelve frames per second (based on the speed of a 233 MHz Pen-
tium II). This compares favorably to both laser rangefinding and ultrasonics, particularly
when one appreciates that ranging information with stereo is being computed for not just
one target point, but all target points in the image.
It is important to note that the SVM uses CMOS chips rather than CCD chips, demon-
strating that resolution sufficient for stereo vision algorithms is readily available using the
less expensive, power efficient CMOS technology.
The resolution of a vision-based ranging system will depend upon the range to the
object, as we have stated before. It is instructive to observe the published resolution values
for the SVM sensor. Although highly dependent on the camera optics, using a standard
6 mm focal length lens pair, the SVM claims a resolution of 10 mm at 3 m range, and a res-
olution of 60 mm at 10 m range. These values are based on ideal circumstances, but never-
theless exemplify the rapid loss in resolution that will accompany vision-based ranging.

4.1.8.3 Motion and optical flow
A great deal of information can be recovered by recording time-varying images from a
fixed (or moving) camera. First, we distinguish between the motion field and optical flow:
• Motion field: this assigns a velocity vector to every point in an image. If a point in the
environment moves with velocity v , then this induces a velocity in the image plane.
v
0 i
It is possible to determine mathematically the relationship between and v .
v
i 0
• Optical flow: it can also be true that brightness patterns in the image move as the object
that causes them moves (light source). Optical flow is the apparent motion of these
brightness patterns.
In our analysis here we assume that the optical flow pattern will correspond to the
motion field, although this is not always true in practice. This is illustrated in figure 4.26a
where a sphere exhibits spatial variation of brightness, or shading, in the image of the
sphere since its surface is curved. If the surface moves, however, this shading pattern will
not move hence the optical flow is zero everywhere even though the motion field is not
zero. In figure 4.26b, the opposite occurs. Here we have a fixed sphere with a moving light
source. The shading in the image will change as the source moves. In this case the optical