DPRG List  


Subject: [DPRG] CMUcam
From: Randy M. Dumse rmd at newmicros.com
Date: Mon Feb 7 17:23:38 CST 2005

> I'm not an expert at image analysis, but am wondering
> about your idea.

I appreciate the reply, none the less. It's an interesting item
to discuss.

> Do you
> think limiting your view to a few scan lines is as
> robust?  What if the course
> is not very clean?

You know, I'm more worried about reflections on the surface of
the course looking like white out areas. Those big plate glass
windows at the Science Place have tricked many robots into very
strange behavior!

> Would a random scuff mark or
> patch of dirt be more likely
> to trick analysis of a few scan lines?  Would
> analysis of the whole quadrant
> help average out the noise in such a scenario?

I plan on putting Ultra Bright white LED's  shining down on the
camera's field of view to try to even out lighting.

The thing about quadrants is you get "bang bang" outputs. That
is, you cant tell left right, up down (if I am understanding the
suggestion correctly).

The thing about using thin horizontal windows is you get center
of mass of the color along a continuum. This should give me up
to 80 positional "outputs" with a horizontal resolution of 80
(Or is it 143?. The X output seems to be 1 to 80 pixels in
width. 80 I think.) So, I hope to use this as a rather fine
error signal, and use a PID loop to compensate, keeping the line
in the middle. The further the MM of the line shows up, left or
right, the greater the need for correction in the steering.

Now my thinking is, if I use, say, half the lines (the upper
half of the picture) rather than just a few scan lines, the MM
will be spread out over a larger area, and so the resolution of
the offset would be decreased. For instance. Let's say I'm right
on top of the line, and its running staright ahead. I'm getting
a MM with an Mx value of 40.

Now a near right angle turn ahead just comes into view, call it
to the left (decreasing pixel positions). If I'm only looking at
the top lines, the Mx value there jumps out left to 20. However
if I'm using 77 lines, top half of all vertical lines, and most
of them are still at 40, and only the top couple are at 20, the
average Mx value will come out like 38. Not much signal change
to know I have to make a radical turn ahead.

Well, that's my thinking, right, wrong, or otherwise. Without
more experimentation, it's just thinking. I did use a white LED
to see if I could track all over the screen, and am pretty
pleased as it looks to capture a tiny dot of the right color
very precisely. So I'm thinking a few line will do.


More information about the DPRG mailing list