Processing digital video in software can be a cumbersome task that can eat
up many of the resources of a computer. Even the simplest of
algorithms, when done on several million pixels per second, can
cause major slowdowns. Real-time video processing is extremely
important in applications such as robotics where processor power
is limited or devoted to many tasks. I propose to implement simple
algorithms, including edge detection, motion detection, and some
simple filters. These processes are characterized by their simple
mathematics, lack of branches to slow processing speed, and identical
actions on every pixel. These properties allow data to be segmented
and sent through many different chips at the same time. The parallel
processing allows the chips to operate at slow speeds, while
still accomplishing the same task as a single, much more powerful
and expensive chip.
One disadvantage of using hardware over software is that hardware
lacks the versatility of software. By using FPGA’s, the chips
can be dynamically reprogrammed as needed. Algorithms can be
stored in ROM on the board, and loaded into the chips as needed.
Each chip is identical, so a relatively small amount of memory
is needed. RAM’s can also be used so that the user can write
(using VHDL, or some equivalent) and load in their algorithms.
I/O presents the largest challenge for logic using multiple parallel
programmable chips (LUMPP chips). The slow speeds mean that the
data must be brought onto and sent off of the chip in large segments.
This requires a large buffering and immense bandwidth between
the buffer and the chips.
After
discussing the project with Bruce Maxwell, I believe this project
can process full sized video at a good frame rate, at a relatively
low cost, and can be useful in freeing up resources on his robots.
