DIY cleaning robot

Posted by ispas on September 24, 2009, 12:56 pm

See my recent home made robot with cleaning capabilities using computer
vision for obstacle avoidance and more:

http://gidesa.altervista.org/robot/robot1.php

Posted by l o on September 25, 2009, 2:30 am

> See my recent home made robot with cleaning capabilities...
> http://gidesa.altervista.org/robot/robot1.php

It took several attempts to get your web site to come up, but
eventually I got through. Thanks for sharing your project.

You didn't elaborate on your video processing you use to navigate. How
is that working out? I know that is a very big challenge to
accomplish.

I am speculating that the Microsoft Natal 3d imaging camera will make
robot navigation significantly easier. I remember reading some specs
about the 3DV camera, as it was known before Microsoft took it over.
My memory is vague, but I think it had a depth resolution of 6".
While laser scanners might give better depth resolution, I think the
Natal camera is such an improvement in its ability to depth-map an
entire image, the weakness in depth resolution won't really matter.

I just tried to get more technical information from 3dvsystems.com
(the originator of the technology) but they were not to be found.
Perhaps they are letting Microsoft take the lead on this.

Joe Dunfee

Posted by ispas on September 28, 2009, 1:55 pm

l o wrote:

> It took several attempts to get your web site to come up, but
> eventually I got through. Thanks for sharing your project.

The provider was changing servers just a few days ago, when I wrote my
message!

> You didn't elaborate on your video processing you use to navigate. How
> is that working out? I know that is a very big challenge to
> accomplish.

I use the well known optical flow algorithm, included in OpenCV, too.
It work well with my robot. Consider that it has to move in very
limited spaces in my flat, around furnitures of every type. And the
lighting varies very much, from a room with windows exposed to direct
sun, to a corridor that is much more dark. It's very very different
from that "simple" laboratory rooms that I see in experimental works on
Internet. The spaces often are only a little more wide than the robot
diameter (around 40 cm). So it's much more difficult, in my opinion.
There are also problems with bright light spots. For example, because
the floor is clear, not opaque, and the direct sun light creates very
bright spots. Other problems are with monochrome walls (ex. white).
They don't create a good optical flow. I think that a single camera
alone cannot resolve all that.
Anyway the results are good, especially using a wide angle lens on
camera. But sonar and a mechanical bumper are again a need if want to
avoid obstacles "hard" for optical flow.
Regards

Giandomenico De Sanctis

Posted by John Nagle on September 29, 2009, 12:25 pm

ispas wrote:

> l o wrote:
>

>> It took several attempts to get your web site to come up, but
>> eventually I got through. Thanks for sharing your project.

>
> The provider was changing servers just a few days ago, when I wrote my
> message!
>

>> You didn't elaborate on your video processing you use to navigate. How
>> is that working out? I know that is a very big challenge to
>> accomplish.

>
> I use the well known optical flow algorithm, included in OpenCV, too.

Optical flow generally doesn't work very well. Try the LK tracker
in OpenCV. That looks for corners and locks onto them. Good in
indoor spaces.

John Nagle

Posted by ispas on September 29, 2009, 12:39 pm

John Nagle wrote:

> Optical flow generally doesn't work very well. Try the LK tracker
> in OpenCV. That looks for corners and locks onto them. Good in
> indoor spaces.

LK tracker can be useful to drive the robot to a known position.
But it seems to me difficult to use it for obstacle avoidance, in an
very cluttered indoor space, just full of unknown/casual objects.
The lock to selected corners is difficult in (relative) fast movement,
for example if the robot turn on place.
And LK method depends anyway from image details, as optical flow. So
the problems with bright reflections, monochrome walls, etc. are the
same.
What you suggest? I could recreate the set of selected corners every 2
frames, as I do with optical flow.
Anyway I have good result with optical flow.

Page 1 of 2 1 2 > last >>

Subject
Posted In
Date

http://robot.pro <- domain name "robot.pro" for sale
Industrial Control Group
2007-11-13

http://robot.pro <- domain name "robot.pro" for sale
General Robotics Forum
2007-11-13

http://robot.pro <- Domain Name for Auction
General Robotics Forum
2007-10-14

Anyone used a Gameboy Camera on thier robot ?
General Robotics Forum
2005-12-16

Advances in robot programming with Lego Mindstorms
General Robotics Forum
2012-04-13