May
2006, Issue 190
Image
Processing for Robots
Renesas
M16C Platform Design Contest 2005
SEARCH
& TRACK OBJECTS
Image
processing isn’t a feature normally associated with
inexpensive low-power robotics platforms. I designed
the UniRoP because I wanted to learn more about simple
picture processing. Some of the M30624’s features led
me to expect that my experiments would be successful.
The M30624 has approximately 31 KB of RAM, and it can
be programmed in direct memory access (DMA) mode. The
videoa1.c software module on the Circuit Cellar FTP
site includes all the necessary functions for simple
object detection.
To
test image quality, simple test software can read an
image via an RS-232 cable to the PC. The serialcam.exe
PC tool and a picture of the test scenery are posted
on the Circuit Cellar FTP site. To minimize RAM usage
and computational power, the image’s size is limited
to approximately 64 × 120 black and white pixels. The
system’s picture quality is poor because of the low
resolution. The image-processing software’s task is
to detect and display the largest object.
The
main problem for the search algorithm is the distinction
between the object and the background. In my first experiment,
object and background pictures were separated by the
intensities of their images. The picture is described
by a matrix of the image pixel IM,N. You now need an
intensity level (S) operator to separate the image areas
belonging to the object (O) from background (B). The
probability (P) of the intensity is:
IO
is the number of object pixels. IB is the number of
background pixels.
The
intensity distribution is:
The
overall error (E) is easy to calculate:
An
optimized level operator is defined as:
This
equation separates the object from the background. If
the object looks like a Lamberthian radiator (an object
with equal light intensity), s2
is less than (µO – µB). Thus, you can ignore the second
part of the equation.
Listing
2 gives you an idea of how to calculate the value
of the level operator S. With the bLevel value, the
object and background are differentiated. All values
larger than bLevel can be part of the object. Smaller
values are part of the background.
The
equations look impressive, but how does the algorithm
work? Photo 2 shows some results. Photo 2a is the x-coordinate
with a fixed intensity level in the search algorithm.
Parts of the background are marked as a part of the
object. The result isn’t useful.
|
a)
b)
c)
(Click
here to enlarge)
|
Photo
2—I sampled various methods to improve the photo
quality. The shots are grainy at this size, but
you get the idea. a—Here you see the image without
a color filter. b—I used a red filter for this image.
c—This time the object is a Lambertian radiator. |
A
red color filter was put on the lens for Photo 2b. Unfortunately,
some pixels weren’t detected. It’s better, but not good.
In
Photo 2c, however, the x-coordinate is set to the correct
value. The searching algorithm doesn’t use a fixed level
for separating the object and the background. The value
has to be adapted by dependencies of actual light conditions
(the red filter also helps).
Assuming
the object emits Lambertian light rays, the luminance
density is the same on every point of the object surface,
the algorithm separates the object much better from
the background. A low-power red LED illuminates the
ball to give a Lambertian radiator. You may download
a short video clip illustrating the results and showing
the UniRoP in action (as well as a couple of other robots)
from the Circuit Cellar FTP site.
