aditya_mangalampalli
Posts: 8
Joined: Mon Apr 03, 2017 2:00 am

Object and Feature Recognition

Mon Jun 26, 2017 5:53 pm

Is there an object recognition software for the Raspberry Pi that available for everybody to use that ends up being free? Whenever I search online, I always find OpenCV and facial recognition software instead of feature or object recognition. Any insights, thoughts, or ideas anyone?

amcdonley
Posts: 194
Joined: Mon Jan 26, 2015 5:56 pm
Location: Florida, USA

Re: Object and Feature Recognition

Thu Jul 20, 2017 4:17 pm

The problem space is huge, and OpenCV is *THE* solution. We need to grow the number of people that understand how to apply OpenCV functions to basic vision problems.

There are no vision "components" yet, only tools for folks to build with.

cpunk
Posts: 85
Joined: Thu Jun 29, 2017 12:39 pm

Re: Object and Feature Recognition

Thu Jul 20, 2017 8:41 pm

OpenCV is neat, and I've worked with it, though not on a Pi, but a generic Debian system with and Atom CPU.

To say the least, it was slow... and it's no secret that vision needs high-end processors, which Atom is definitely not... and neither is Raspberry Pi's tiny CPU (although 4 cores is nice).

Thus comes the question... of why am I working with a Pi currently? Only because of the GPU + software combination that makes it possible to leverage the H264 encoder. :) Some folks have been nice enough to discover that the GPU, whilst encoding video into H264, makes it possible to grab the leftovers of that encoding process (macroblock motion vectors) and use them for very interesting stuff. :) That's much less data, and much more concentrated useful information, than actual video frames, so it's massively faster to work with...

...and it doesn't close off the possibility of consulting real video frames for color data, selectively, when it's needed.

I'm not currently using OpenCV, though. OpenCV is a huge code base. Now that I'm building flight control software for a drone/robot which is liable to break stuff if it malfunctions... I can't really afford to say that I don't know how a critical component works. Thys I currently avoid OpenCV. In an ideal world, I would know it well enough to use it for every project, and in an ideal world, I would feed the motion vectors obtained from the GPU also into OpenCV...

...but right now, it's easier for me to DIY it all.

Return to “Camera board”