|
[DPRG] webcam robot vision limitations
Subject: [DPRG] webcam robot vision limitations
From: Chris Jang
cjang at ix.netcom.com
Date: Wed Jun 27 20:42:56 CDT 2007
>> So a word for others who may try this - mass storage DMA bandwidth is the most critical feature to look for. CPU speed, RAM, device I/O bus speeds are not as important as fast disk or flash storage. If you see a SBC that uses SD/MMC or CF flash, make sure DMA is used.
>
>Do you have any estimation of the amount of processor you'd need to
>run the algorithms you want to run (feature based techniques /
>convolution of reasonable resolution)?
A straight answer is about 10x more CPU performance. So I think
something like a 1 GHz VIA mini-ITX board would work. Instruction
set optimization (e.g. MMX on x86) could allow for a slower CPU.
With the 200 MHz ARM9, it takes several hundred milliseconds to
transform a 640x480 grayscale image into an integral image and then
convolve for edges. Several hundred must become tens of milliseconds
to fit within the 60 to 70 ms per frame at 15 fps. Just decoding the
webcam JPEG output at full resolution is around 70 ms per image.
I know it would be better to use a camera that outputs uncompressed
images. But then the USB must be fast enough to keep up - which is
why most webcams output JPEGs anyway.
I've had the notion of using the webcam as a line scan imager. Is the
entire image of each frame necessary? Perhaps not, only decode parts
of it. Throw away the bulk of each image early in the processing
pipeline. But as I've not read about anything like this, it may be
impractical.
More information about the DPRG mailing list
|
