New project - feasable?

[smitjs]

Hello everybody.
I have had long embedded experience, mostly with 8051, but none with Arm.
Would this be possible with Teensy 3.1?
The project will be a camera with software to identify the game animal specie, and do some action (like darting for injections) based on species.
Therefore pan and tilt outputs would be required, but can be from a seperate microcontroller on a link like UART or SPI.
The camera module could be OV7670 or similar. These all have at least an 8-bit bus and possibly six extra control lines.
It would be nice if photos could be stored on SD cards, in order to check the functioning.
The hardware boot I understand, but what about debugging? I am used to the Jtag on 8051. Is similar possible with Teensy 3.1?
I would not like to start the project and find that it is not feasable.
Would the GCC toolchain be feasable for use with Teensy 3.1?
The OpenCV or SimpleCV library will be a great help, but they use Python. Can Micropython be run in Teensy?
Any advice greatly appreciated.
Thank you
smitjs
South Africa

[PaulStoffregen]

I'm pretty sure Teensy 3.1 is too small to run OpenCV.

[smitjs]

Thank you Paul

[Roland]

Hi Paul,
as far as I know you plan to builld a teensy++3 (or higher) with a much faster processor.
Could you imagine to see a chance to run opencv on this kind of teensy?
When (not the exact date, but the year) can we expect to get the new teensy?
If the future teensy is no option for opencv I will buy an odroid for that purpose,
but I rather stick to teensy as I make great experiences with the teensy 3.1 just now.
Greetings
Roland

[WMXZ]

Hi Paul,
as far as I know you plan to builld a teensy++3 (or higher) with a much faster processor.
Could you imagine to see a chance to run opencv on this kind of teensy?
When (not the exact date, but the year) can we expect to get the new teensy?
If the future teensy is no option for opencv I will buy an odroid for that purpose,
but I rather stick to teensy as I make great experiences with the teensy 3.1 just now.
Greetings
Roland

Roland, slow down. Paul concentrates at the moment Teensy LC, Arduino1.6, Teensy1.2, and the forum, all at the same time, and then you ask for Teesny 3++/4 !
OK, in another thread he said, later this year he may develop such board.

[PaulStoffregen]

LOL, yes, I am buried in coding and testing for Teensy-LC and Arduino 1.0.6, but I do try to answer questions too.

There already is a lengthy thread about future high-end boards. Probably best to start there.

https://forum.pjrc.com/threads/24633...f-a-Teensy-3-1

I'm not really familiar with OpenCV, but from what little I do know, it seems unlikely any single-chip microcontroller will be up to the task within the next year or two. Well, except maybe Intel Edison & Curie? Intel has something (almost) everyone else doesn't: 22 nm transistors.

If you go to ARM's website and look for pretty marketing info on Cortex-M7, you'll find a chart about M7 scaling to smaller transistors. The chart makes it seem like 45 nm might be soon, and perhaps it could be for M7 cores on the same die as other stuff in cell phones. But for microcontrollers, on-chip analog circuits and flash memory dramatically change the way the wafers are fabricated. This is a huge topic, and again, I'm not really an expert. But I do know just enough to say pretty confidently that we're probably not going to see Cortex-M7 scale up to 500+ MHz speeds on single-chip microcontrollers within the next year.

For OpenCV, I'd go for a single-board computer that's capable of running it well.

[Roland]

Ok, Thanks for the answer and the advice!

[GremlinWrangler]

Hi Roland, off the Teensy topic but be careful with ambitions with CV since it's still a 'hard' task. Fundamentally there is a need to examine every pixel and compare it to adjacent pixels. All the software vision magic in the world doesn't help if your hardware is not fast enough to visit every pixel in the image and do that first step of edge detection.

So either you use a serious CPU capable of hitting X by Y pixels at Z FPS getting them all into RAM and then completing say 50 instructions at each one. Or you cheat and use custom hardware tuned to extract specific geometry and pass basic yes/no maps to the micro controller. Which is why we have got used to face detection on mobile devices while forgetting that detection of arbitrary objects against a none static background is going to be interesting.

Unsure how far you are on the CV end but would be suggesting getting a classification package running on a desktop and then wind down the resources it is allowed to access there to determine a scope of your needed hardware in terms of RAM and CPU cycles.