[Dprglist] Seeing the RoboColumbus cones

[David Anderson]

John,

Interesting post.   Also great to see you out and about at the last 
competition.   And dinner.   Robots + Mexican food == DoublePlusGoodness.

I will start by saying that you, being by nature an inquisitive and 
research-oriented kinda guy, should certainly be encouraged to delve 
into this and find us a superior method of finding colored objects, be 
they fluorescent orange or any other color.   By all means.

But I think perhaps there is a misunderstanding here.  You wrote:

"I think we have a conceit that orange is unique enough because it looks 
dramatic to our eyes."

Instead, fluorescent orange is used precisely because there is otherwise 
so little of it naturally in the environment.  The software can be tuned 
to recognize any arbitrary color.   But most of those colors are present 
in copious amounts in nature.  Which makes the recognition task much 
harder.  Not much fluorescent orange in the average outdoor scene.

The digital cameras we use are in fact most sensitive to light in the 
infrared, because of the nature of the CCD tech.  And everything 
outdoors has a significant infrared signature because of the nature of 
the illumination, i.e., the sun.   As a result all modern digital 
cameras, including the one in your phone as well as those used by 
professional photographers, have an infrared filter on the front end.   
Otherwise the infrared swamps out everything else.   (In fact if you 
want a camera that can see into the infrared, to see heat and such, all 
you need to do is remove that filter.   There are hacks on the internet 
to do that very thing.)

The reason that fluorescent orange works so well with these cameras is 
not because of human conceit, but rather because there is so little of 
it naturally in the environment to confuse the recognition algorithm.   
So the signal to noise ratio is very high, and the likelihood of false 
detections is commiserately low.

In fact, this is true for all the fluorescent colors, not just orange.

The cameras and recognition algorithms we use work equally well with 
fluorescent blue and fluorescent green.   Again, because there is so 
little else in the environment that are those colors.   ( Just kind of 
hard to find fluorescent blue and fluorescent green traffic cones, 
though we could paint some those colors if so desired.)

In some early experiments I added a 1 million candle-power flashlight to 
the robot in order to try to smooth out the wide range of lighting 
differences in outdoor situations.   But I found after a lot of 
experimentation that it wasn't really needed.

As an aside, a biology professor once told me that the reason that our 
eyes are optimized for "visible" light which, as you point out, is less 
than an "ocatve" of the electromagnetic spectrum, is because eyeballs 
evolved underwater, and indeed are full of water, and water is 
transparent to those frequencies.  Birds and insects can also see into 
the ultraviolet, but ultraviolet causes cancers, and only critters with 
short life spans have that capability: they don't live long enough for 
the cancers to be a problem.

But as I say, if you can come up with a different type of sensor or 
sensor plus illuminator that works better, bring it on!   We're all 
interested.

cheers,
dpa







On 12/05/2016 06:07 PM, John Swindle wrote:
> Dave Anderson and others,
>
> Some time back I asked why visible light cameras are used to locate 
> the traffic cones. I think the answers were that such cameras are 
> cheap, and something to the effect that the color of the cone is notable.
>
> I don't think I made my point: As with sonar operating at a frequency 
> that we think is noise-free because we don't hear all the interfering 
> noise that exists above our hearing, I think we have a conceit that 
> orange is unique enough because it looks dramatic to our eyes. But 
> visible light is such a tiny part of the usable spectrum. What I was 
> suggesting is that there might be a characteristic of the cones that 
> could be exploited to give fantastic distinction to the cone's 
> location, even if the exploit is unique to the cones that are being 
> used that day. The exploit might require the robot to illuminate the 
> cone. (Or, shall I say, irradiate the cone? The group might find that 
> disturbing.)
>
> My comments are motivated by several things:
>
>  . The problems of using visible light cameras to guide a self-driving 
> vehicle, where radar, LIDAR, and other approaches are much more foolproof.
>
>  . "Alternate light sources" cited in the forensic TV shows. A fancy 
> name for colored light and some beyond-visible stuff.
>
>  . Tricorders and ship's sensors.
>
> Later,
> John Swindle
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dprg.org/pipermail/dprglist-dprg.org/attachments/20161205/b5703813/attachment.htm>