[FoRK] coffee drone

Jim Whitehead ejw at cs.ucsc.edu
Sat Jun 23 12:01:37 PDT 2012


Kinect is an inexpensive depth sensor.

Jim
-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity.

Damien Morton <dmorton at bitfurnace.com> wrote:

On Fri, Jun 22, 2012 at 3:22 PM, Stephen D. Williams <sdw at lig.net> wrote:

> On 6/22/12 12:12 PM, Damien Morton wrote:
>
>> On Fri, Jun 22, 2012 at 3:01 PM, Stephen D. Williams<sdw at lig.net> wrote:
>>
>> On 6/22/12 11:32 AM, Damien Morton wrote:
>>>
>>> I've been thinking about an autonomous personal cameraman.
>>>>
>>>> ...
>>>>
>>>
>>>
>>>
>>> Machine vision isn't quite reliable enough for this, and the beacons
>> that
>> exist require expensive IR vision gear to see.
>>
>
> If you were to wear a 1 ft. square target on your back and a similar one
> on your head, it would be trivial to track visually. Should be doable
> without that, but trivial with it. Code's been out there for recognition /
> tracking for at least 4 years. You'd probably want a zoom camera paired
> with wide angle to get good distance, but pretty trivial. Tracking
> stability would be based on visual flow, corner search, and movement
> prediction / motion estimation.
>

You're talking about the markers used for AR systems - these are going out
of style now that OpenCV has markerless tracking

https://www.youtube.com/watch?v=ZnDN76GNlA8
https://www.youtube.com/watch?v=3Uwd0165k0Q&feature=watch_response

Here's an example of a marker based autonomous tracking drone
http://www.theregister.co.uk/2012/06/18/bot_jogging_buddy/

You have to wear the right t-shirt though.


> Get me a quad or hex copter that can carry a good smart phone, and I'll
> implement something. In a few months when I have time. But they'll
> probably be selling by then. Already the open source autopilot is rivaling
> defense projects in some sense.
>
>
> Tracking in the open isn't too hard. The difficult level is environment
> understanding and planning so that it wouldn't run into trees, buildings,
> tunnels, other quadracopters, etc. I.e. a full self-driving skycopter /
> skybot.
>
>
So, there are solutions that build up point clouds from stereo vision
systems,

http://www.youtube.com/watch?v=9Y4RQVpp-BY&feature=related

And there are solution that work using Structure from Motion:

http://www.youtube.com/watch?v=AdmTGAo1Jvg

There are also solutions that use LIDAR, expensive at $5K for a lidar unit

http://www.youtube.com/watch?v=WoMrPKCpWXw

My guess is that you'd want your drone to first survey the area it was
going to be operating in, and use some offline process to build up a map of
the area.

Or, if you felt your drone had enough energy budget to carry something like
a beaglebone around with it, you might be able to do the mapping on-board.

What we really need is an inexpensive low-power per-pixel depth sensor.
_____________________________________________

FoRK mailing list
http://xent.com/mailman/listinfo/fork



More information about the FoRK mailing list