DPRG
DPRG List  



[DPRG] stella driving down a path

Subject: [DPRG] stella driving down a path
From: Chris Jang cjang at ix.netcom.com
Date: Sat May 12 16:40:43 CDT 2007

Hi Kenneth, it is a good suggestion. Unfortunately, there are
device and driver issues.

The only stable configuration I found is capturing at the native
640x480 resolution of the webcam. This image is immediately
subsampled during decoding down to 320x240. All subsequent
processing works at 320x240. So it is a little faster but not by
much.

I had to make an economic decision - spend time working on the
device driver and possibly acquire a new camera - or - go with
what I have now and is working. I elected to proceed as-is. It is
good enough.

It is kind of shocking to see that very slow frame rates from an
overexposed camera are still good enough to drive a robot car at
relatively high speeds.

Also, I have also noticed that the feature based algorithms I am
using degrade as resolution lowers. The filters become unstable.
So lowering resolution may improve framerate at the cost of vision
performance.

Chris


-----Original Message-----
>From: Kenneth Maxon <kmaxon at qwest.net>
>Sent: May 12, 2007 1:39 PM
>To: dprglist at dprg.org
>Subject: RE: [DPRG] stella driving down a path
>
>
>
>Hi Chris, just for kicks and giggles, have you considered lowering resolution as a means of increasing frame rate?  So many of the algorithm layers in image processing are x*numpix, with a few being numpix^2 or combinations there of, this may have already been discussed, and I know you have the long term solution to move to 10x horse power in the future, just thought it might be an interesting short term idea.
>
>-Kenneth
>
>-----Original Message-----
>From: dprglist-bounces at dprg.org [mailto:dprglist-bounces at dprg.org]On
>Behalf Of Chris Jang
>Sent: Saturday, May 12, 2007 8:40 AM
>To: Charles Thurston; dprglist at dprg.org
>Subject: Re: [DPRG] stella driving down a path
>
>
>Thanks - my mom asked me why Stella drove off the path. It is
>a good question and deserving of an answer. I think the answer
>is interesting as it is visually obvious.
>
>http://golem5.org/video/mvi4027_crash.jpg
>
>Improvements to any one of the following will prevent this failure.
>
>1. more reliable and higher framerate camera system
>2. more sophisticated computer vision algorithms
>3. a real control system with a model of the environment
>
>Number 1 is constrained by the hardware. That is for the next robot.
>
>Number 2 is applying what I already know. Like Stanford's robot, I
>think Stella will have to use statistical methods. Even with a
>multi-million dollar robot like Stanley, there was a point beyond
>which improving the hardware to eliminate noise was too expensive.
>So it had to be done in software.
>
>Number 3 is stuff I need to learn. Stella needs to have an adaptive
>control strategy. This also implies a re-architecture of the
>software into a modular form. At this point, I may move the higher
>level code into C++. However, the low-level code will remain in C.
>And the user interface code for the LCD, pushbutton and radio is in
>the Lua scripting language. I am trying to use the optimal level of
>abstraction for different problems.
>
>
>-----Original Message-----
>>From: Charles Thurston <cthurston at sc.rr.com>
>>Sent: May 12, 2007 7:47 AM
>>To: dprglist at dprg.org
>>Subject: Re: [DPRG] stella driving down a path
>>
>>Hello Chris,
>>
>>Impressive
>>
>>Friday, May 11, 2007, 9:35:17 PM, you wrote:
>>
>>> http://golem5.org/video/mvi4027.mpg
>>
>>> Hello Everyone,
>>
>>> Here is Stella again. There are no changes to the simple
>>> obstacle avoidance algorithm from the last movie in a
>>> parking lot. What's going on is purely reactive behavior.
>>> There is no path planning or kinematic model.
>>
>>> Note that Stella has no sensors other than a USB webcam
>>> running at around 1.3 frames per second.
>>
>>> It looks interesting but really there is a lot of work to
>>> do. The computer vision is the simplest possible - look
>>> for a clear patch of ground immediately ahead - and
>>> ignores a lot that is useful. There is also no control
>>> model. I just picked constant thresholds and duty cycles
>>> for sensing obstacles and driving the motors. It is
>>> entirely ad hoc and does not adapt at all. That's why
>>> Stella over-corrects. A real control system is necessary.
>>
>>> By the end of the year, it will probably be necessary to
>>> move on to another robot. That will probably be a hobby-
>>> grade RC vehicle with at least 10x the computer power.
>>
>>> Chris
>>> _______________________________________________
>>> DPRGlist mailing list
>>> DPRGlist at dprg.org
>>> http://list.dprg.org/mailman/listinfo/dprglist
>>
>>
>>-- 
>>Best regards,
>> Charles                            mailto:cthurston at sc.rr.com
>>
>>
>>_______________________________________________
>>DPRGlist mailing list
>>DPRGlist at dprg.org
>>http://list.dprg.org/mailman/listinfo/dprglist
>
>_______________________________________________
>DPRGlist mailing list
>DPRGlist at dprg.org
>http://list.dprg.org/mailman/listinfo/dprglist
>
>
>
>
>_______________________________________________
>DPRGlist mailing list
>DPRGlist at dprg.org
>http://list.dprg.org/mailman/listinfo/dprglist

More information about the DPRG mailing list