A project for Oak? « LeafLabs Garden

LeafLabs Garden » Community » The Lounge

A project for Oak?

(16 posts) (6 voices)

Started 5 years ago by gbulmer
Latest reply from ridgebackred

12 Next »

gbulmer
Moderator

Folks have probably seen the Ultra Low-Cost Laser Rangefinders by Neato Robotics

For those who haven't seen it, it is worth a look. Essentially it uses a low-cost laser and a low-cost digital image sensor (mobile phone?) to triangulate the position of the laser 'dot', and calculate a distance. It scans 360 degrees to cover a room.

A sneaky part is it is only 10cm wide, and all the parts are fixed to the same base, allowing the thing to be calibrated once at manufacture. The entire optical system is rotated 360 degrees at several revs/second to create a local 'obstacle' map.

To get good resolution, they are doing sub-pixel interpolation using the image data. They use a DSP for this part.

This sounds like something that might be worth a go with Oak.

Maybe LeafLabs can think of a way of making/obtaining the optical system. It would make a great tool for low-cost (autonomous) roving robots.

Posted 5 years ago #
josheeg
Member

I agree a smarter camera that does the difference of every pixcel or detects distances by a projector sending out lines like in the makerbot cyclopse is a good aplication then the robot can avoid obsticles and maby identify ones it can work with.

Posted 5 years ago #
gbulmer
Moderator

I think the makerbot cyclops is one of the family of low-cost 3D scanners which work upto a couple of feet away, and take upto several minutes to scan an object. AFAIK, it doesn't really measure absolute distance, but gets a displacement relative to the surface it is scanning from each scan line, and builds up a 'cloud' of data from many scan lines. A piece of post-processing stitches together the neighbouring points in the cloud to yield a 3D 'mesh'.

The Neato Laser Rangefinder measures absolute distance (relative to the robot, of course). It claims to scan 360degrees, at 10Hz and examines the scene at approximatly one degree increments, and so reads a subframe from the camera at 3.6k frames/second (they say almost 4k frames/secobd).

It triangulates the single laser dot reflected back onto the camera sensor to calciulate distance. Extracting the appropriate subimage (so that it can scan at 4k frames/second) and doing the subpixel iterpolation looks a bit beyond a Cortex-M3, but maybe 'easy' with Oak's FPGA.

So they would both be great to have, and an LDR would be great for a mobile/roving robot.

Posted 5 years ago #
okie
Administrator

awesome project idea!

Posted 5 years ago #
gbulmer
Moderator

Okie - Any ideas about how we could get hold of, or make something like that Laser Rangefinder sensor?

Being able to scan 360 degrees of the environment at 10Hz would be a fabulous sensor for any sort of mobile robot. It won't win a DARPA challenge, but it might be able to dodge a rampaging VW Touareg!-)

Posted 5 years ago #
poslathian
LeafLabs

I usually take laser rangefinders as evidence that were all still "doing it wrong" - I get around just fine with just a couple well placed image sensors in my face.

BUT - neat project idea. Im not sure I fully understand the "low cost" image sensor part, being that - as was pointed out - a 3.6fps refresh rate doesnt come standard on most CMOS sensors. I suppose you really only need to grab a single horizontal line, not the whole frame - perhaps someone can find a sensor that supports this type of configuration.

Other than that, FPGA's+Cameras = awesome!

Posted 5 years ago #
gbulmer
Moderator

I usually take laser rangefinders as evidence that were all still "doing it wrong" - I get around just fine with just a couple well placed image sensors in my face.

I see your point, but how many yotta-ops are you applying to the problem?-)
IIRC, about 10% of the brain is associated with vision, so about 10^99 neurons?
A lot more than one yotta-ops.

I have no problem using man-made physical phenomina, like coherent light from a LASER.
We didn't have the chance to evolve the capabilty :-)
(but then, I believe evolution is the best explanation we've got, unlike most of the folks on your side of 'the pond' :-)
(I should not start this, but I'm in a playful mood :-)

Anyway, the reason I posted is to stimulate folks to consider the power of FPGA+high-volume sensors, and image sensors are a great example. If the FPGA can be tied directly to the image sensor, we're cooking!

Many modern camera sensors can return subimages (window) much more quickly than an entire image. A friend has a video camera which, I think, but I might be muddled, can do over 800 frames/second for about 10% of the image.

An extreme example is http://www.laboratorytalk.com/news/phj/phj108.html which can do 5,000 one megapixel images/second, but a jaw dropping 150,000 sub-images/second. This isn't something like STEAM, but an image sensor.

So, I strongly agree, FPGA+Camera = Awesome!
So when will Oak be ready?-)

Posted 5 years ago #
okie
Administrator

you meant 10^9 neurons?

Posted 5 years ago #
poslathian
LeafLabs

These high speed cameras are simply amazing, but nearly as expensive as youd expect (10-200k$). But i think youre right about windowed output on cheaper consumer image sensors. Lets try and find one that does what we want. I wouldnt be too surprised if someone makes an inexpensive sensor that has single-scan-line output mode for just this sort of thing.

Of course, you could also try and get time of flight working ;)

Posted 5 years ago #
gbulmer
Moderator

Okie -

you meant 10^9 neurons?

I am really, really sorry about this, ...
no, 10^99 neurons for me, you can speak for yourself (urggh)

It's about 100,000,000,000, i.e. 10^11 neurons, so about 10^10 for vision, with a 'fan-out' of 7,000?

An average neuron can fire in 5ms, so an MIT neuron, is probably quicker, say about 2ms?
Flicker stops being a big problem arond 50Hz, so 10 sequential neuron events?
In theory one neuron might trigger 7000^9 other neurons in that time, or 4x10^34 triggers, so all 10^10 neurons could get involved, doing 4x10^24 ops = 4 yotta-ops!

A friend of a friend has got something like time of flight working
Real time low cost pseudo-random noise continuous wave lidar

Posted 5 years ago #
poslathian
LeafLabs

I think MIT neurons are actually a bit slower, due to the inevitable burn out effect.

Posted 5 years ago #
gbulmer
Moderator

Well you folks seem to manage okay, but maybe it's those Border Cafe Fajitas, and Hornitos Margarita's that keep them neuron's poppin' and jivin' !-)

(while as I was doing the estimate, I realised I needed to raise the 7000 an extra few powers of 10 to get to yotta-ops, so that's why your visual processing neurons are so fast :-)

I was a bit surprised at 5ms/neuron-firing, how (relatively) little processing power was available for 'real-time' vision processing. Still a lot, but not an amazing amount (any more).

Posted 5 years ago #
gbulmer
Moderator

Anyway, back to the Laser rangefinder.

Apitina and Omnivision make sensors at a reasonable price (sub $30 at one off prices), which support windowing.

At 48MHz pixel transfer rate (e.g. one of the less capable Aptina's), 3600 frames/second (10 frames at 360 positions) would be a 115 x 115 pixel square, which should be big enough to encompass the laser dot if it is keeping track of where things are going.

Some of those image sensors support 96MHz clock rate, so 115x115pixels should be a reasonably conservative estimate.

Push the data straight into an FPGA for initial processing, and it should give reasonable distance feedback 360 degrees around, at 10 sweeps/second.

Do any of LeafLabs have contacts in an engineering department? They could probably make a reasonable prototype mechanism for you to experiment with.

Some of the papers suggest that integrating this sort of 3D distance data with traditiomal image/vision data gives very high quality object recognition and navigation. With a few Oaks, and some low-cost cameras, you might have a pretty good low-speed DARPA-challenge-style sensor system for well under $1000, maybe well under $300.

Just a thought ;-)

Posted 5 years ago #
poslathian
LeafLabs

oh yeah! Will have to check in to those sensors, thanks!

Posted 5 years ago #
brijesh
Member

Hello Guys,

I have had same thoughts about low cost laser range finder. I went through the same steps of trying to find a suitable camera and camera interface.

Finally ended up trying with Taos Inc's 128*1 linear array sensor.
http://www.taosinc.com/ProductDetails.aspx?id=3

Initially I tried to hack something together on a bread board. That did not work as the laser and image sensor have to be rigidly mounted, small disturbances result in large errors.

So ended up designing and building a custom board with lens and laser mounting. I just got the boards fabricated and will be firing them up soon. By the way I am also using STM32 processor on the board, but it is lower end one STM32F100C4.

If you guys are still interested in the this project(after 7 months), I will keep you guys posted. If someone wants to work on it, just let me know.

Cheers

Posted 4 years ago #

RSS feed for this topic

12 Next »

Reply »

You must log in to post.