Determining the angle

Maybe a circle of lines radiating out at different angles.
All lines being equal in length when viewed frontal no
rotation?

A crude ascii version  :)


    \  |  /
 --   +    --
    /  |  \

Yeah, I was thinking of some very simple versions but it seems like my
resolution will be kinda low so there won't be any room for fine
calculations (length, distances and so forth). But maybe it works... Gotta
think about it...

For the manipulators used in space, they use a 3" long black rod with a
white tip, centered in a 1" dia white circle on the rod's platform.  Viewed
straight on, the white tip is centered in the circle, but even small amounts
of pitch or yaw misalignment can be detected as it makes the white tip shift
within the circle.  Horizontal reference stripes give roll misalignment.

Mike Ross

Ok. If I got it right, it's a rod fixed in the centre of the circle, and we
have the top-view (we see the full circle and just the tip of the rod - when
it's straight). Right?

Speculating, not speaking from direct experience, I'd expect that a
distinctively colored circle would work best, especially if low
resolution and processing power are important, the visual target is
reasonably large (at least 5 or 10 pixels high or wide), and high
precision is not required.

Speaking from experience, use HSV (*1) or a similar color scheme to
detect the circle.  Threshold (separate object pixels from background
pixels) based on a formula like object=[abs(hue-orange)<5 AND
saturation>0.7 AND value>0.4].   Hunter's orange is easily identified in
the average indoor setting.  For microprocessor efficiency, rescale the
HSV equations to values of 0-255.

If such a circle is viewed "full frontal", then you will see a circle;
if viewed at an angle, you will see an ellipse.  You can easily
calculate this ellipse using the mean and moments of the pixel
coordinates (*2).  Based on these axes, you can determine the axis of
tilt (major axis of the ellipse).

If you also need orientation, then a second or third uniquely colored
circle can be used.  Looking at the vectors between the centers of these
circles will yield orientation information, in addition to the
previously obtained tilt.

I've seen other algorithms which rely on a checker-board pattern.  They
use edge detection to fit lines across the board.  These lines are then
used to calculate the board's position and orientation.  These also
seemed more processor intensive.

Have fun,
Daniel

Quickly searched links:
*1: http://en.wikipedia.org/wiki/HSV_color_space
*2:
http://palantir.swarthmore.edu/maxwell/classes/e27/F03/E27_F03_Lectures.pdf

    There's source code for a checkerboard-based aligner in OpenCV on
SourceForge. The math needs some work; it sometimes reports totally
bogus alignments.

                John Nagle

In terms of resolution...

There are 2x as many green pixels as red or blue in a bayer color pattern.
The blue are noisy as well.

RGRGRGRGRGRGRG
GBGBGBGBGBGBGB



http://palantir.swarthmore.edu/maxwell/classes/e27/F03/E27_F03_Lectures.pdf

The solution to that is to not look at the Bayer pattern directly, but
convert it to an image format you know how to use. And to severely
oversimplify the noise and contrast problem, one of the best ways to clean
up the data set is to simply drop the entire set of "damaged" information.
The problem is identifying that the blue channel, for example, has more
noise than useful signal. Image processing is a science, and already
addressed every one of the issues we're likely to bring up. The work now is
to make put all those DSP mips to work and extract a useful image in
realtime. Useful is the key word; it doesn't have to resemble what we
consider a good photograph to be useful for machine vision.

It should be pointed out that emulating the human eye and visual perception
is a much more complicated problem than the one we're actually trying to
solve. To be precise, the problem is identifying our relationship to objects
in the immediate environment. Vision is so immediately obvious as the "best"
way to do this -- by virtue of our possessing a very well developed sense of
sight -- that we ignore other possible solutions, conveniently forgetting
that we don't fully understand how the brain processes what we "see". To a
blind person, vision is a damned poor way to find his way around. Rather
than reinvent eyes, it might do us better to find usable substitutes. My
favorite alternative reality is that side-scanning Doppler ultrasound has
promise. At the very least, I don't have to deal with a noisy blue channel
and Bayer patterns.