|
[DPRG] Are representations what makes Can Can hard?
Subject: [DPRG] Are representations what makes Can Can hard?
From: Chris Jang
cjang at ix.netcom.com
Date: Wed Dec 6 22:00:59 CST 2006
> What makes Can-Can so difficult is that it moves beyond the realm of
> sensing and into semantics. There is not, and will probably never be,
> a "can sensor" that will produce a signal when a soda can is present.
> Instead, a robot must deduce, infer, or guess whether the sensor
> readings represent a can, which should be retrieved, or an obstacle,
> which should be avoided.
For computer vision, object detection and recognition are difficult
problems. In my layperson's understanding, all approaches are inevitably
statistical at some level. Even worse, we don't usually know how to
characterize what is in the sensor data that indicates the presence of an
object let alone how to recognize what that object is.
Systems often require extensive training. A machine learning algorithm may
or may not converge to a good solution. I haven't implemented any of these
ideas yet so can't say how it really works. I'm just relating what I've
read in research papers and am planning to do (which often does not match
up with the reality of actually doing it).
> Obstacles (walls) and cans all look very similar to the sensors that a
> robot has. In our contest, they are both:
> - white,
> - vertical, and
> - approximately the same height
I took two soda cans out on the deck at work and took some video to see
how (very) primitive box feature thresholding works at object detection.
http://golem5.org/embedcv/video/cancan.mpg (358400 bytes, MPEG-1)
It probably works well enough for detecting cans. They provide the
strongest stable features in the video except for the electrical boxes on
the wall. The electrical boxes are well above ground level and easily
rejected based on position in the image frame.
The corner formed by the strong shadow can also be rejected. It is formed
by elongated box feature blob outlines. Those blobs are too large to fit
inside a small can-sized bounding box.
The ground is made of tiles with strong edges that sometimes register as
features. But those features are highly directional so will flicker in and
out as the camera moves. They aren't stable so can be rejected.
That leaves the cans.
More information about the DPRG mailing list
|
