|
[DPRG] ad hoc filter for can-can video
Subject: [DPRG] ad hoc filter for can-can video
From: Chris Jang
cjang at ix.netcom.com
Date: Tue Jan 2 15:05:21 CST 2007
>Is the video real time or post processed (if that is the correct term)?
>
>http://golem5.org/embedcv/video/cancan2.mpg
Hi David,
It is post processed. There is a lot of extra computation done to help
visualize what is going on. I am still at the stage of struggling to learn
what can work.
However, the algorithms used are all inexpensive. So the while the
results are not generated in real-time, they are realistic for an
on-board system.
>I think you are correct about code library verses application.
>I would think that in primitive biological vision the eye and associated
>intelligence is adapted to the task. Object recognition of all things is not
>a requirement.
I don't know where the transition occurs. I've read that the Difference/Laplacian of Gaussian interest point computation is identified in primate brain structure. It's interesting to see what these early image transforms look like. I wonder how much object recognition is hardwired into our perceptual hardware. So I guess I'm offering the possibility that it is an unclear distinction.
>Applications like line following and can-can are best served being built up
>from many lower level behaviors. Would it be safe to call obstacle
>avoidance, cliff detection, and visual odometry low level behaviors?
I agree. Here is my intuitive view. Anything that doesn't require object recognition is pretty low level. It's possible to solve problems in an environment without necessarily classifying it into distinct objects. I'm curious what the people with SONAR experience have to say about this. I'm starting to believe that vision has similar environment perception problems to SONAR - interpret ambiguous signals.
>Also for the can-can what if you had a primitive that deduced the size of an
>object in the horizontal plane. A wall is too large, a chair leg is too
>small, a can is just right.
Projective geometry would be a good thing (and is probably necessary).
I've avoided this as then a Hough transform is needed to find the extended
features like walls. Everything done so far is very simple with a combination
of pixel and box feature thresholding. I'm not using histograms to identify
shapes.
Chris
More information about the DPRG mailing list
|
