BUEHLER%KOBOT@VENUS.YCC.YALE.EDU ("buehler%kobot.decnet@venus.ycc.yale.edu") (09/25/89)
David Lowe writes:
// I am looking for information from anyone who has experience interfacing
// a transputer network to some type of image digitizer for near real-time
// computation.
//
// The problem, of course, is that your average TV camera outputs data
// at the rate of about 8 megabytes/sec when digitized. Here at the
// Laboratory for Computational Vision, we already have Datacube image
// processing equipment that can digitize and do local processing on
// this data at video rates. However, for higher-level processing the
// data must be transferred over a VME bus at far slower rates. Has
// anyone looked at parallel input to transputers from a MAXBUS or
// other video data source? Are there general high-speed IO solutions
// for transputers? Any information appreciated.
We need real time vision in our robotics work in the Yale Robotics Lab.
About a year ago we evaluated numerous commercial vision systems. None
of the systems excited us enough to buy it (cost, performance, interfacing,
flexibility, complexity, etc.) so we built a simple but powerful distributed
vision system based on Transputers/A110 ourselves.
Below you find a short abstract-description as it
was sent to the Natug89 meeting. A more detailed description
can be found in the Natug-1 proceedings (Natug-1 Salt Lake City, UT,
April 5-6, 1989, to obtain ($32)from G.S.Stiles, Dept. of El.Eng., Utah State
University, Logan, UT, 84322, stiles@cc.USU.EDU), or else I can email you the
latex version of the paper.
At this opportunity I also would like to ask if anybody else built a
"real-time" vision system,
succeeded to interface cleanly to a commercial system or has any thoughts about
real-time vision hardware in general.
I would be very interested in seeing some comments on this net.
martin (buehler%kobot.decnet@venus.ycc.yale.edu)
=============================================================================
The Cyclops Vision System (ABSTRACT)
Abstract submitted for the
North American Transputer Users Group Meeting.
April 5/6, 1989
Salt Lake City, Utah
M. Buehler, C. J. Taylor, N. Vlamis, and A. Ganz
Center for Systems Science
Yale University, Department of Electrical Engineering
We describe a real-time distributed
vision system frontend for transputer based systems.
The need for real time vision arose out of our robotics research with
dynamical tasks,
specifically tracking balls for
robot juggling: The system must be able to track at least two objects
in 3D (stereo-vision) at field rates (60 Hz).
For computation and robot control we are using networks of
our Yale transputer based distributed control nodes, the XP/DCS
(Proc. 1988 National OCCAM Users
Group Meeting). The motivation to develop our own vision system
as opposed to buying some commercial unit is similar to the reasons why we
chose to build our
own distributed control node two years ago:
* networking via the transputer links
achieves flexibility and virtually unlimited expandability
* system is built to satisfy all our requirements
* by using OCCAM and the standard development environment we avoid the
development of system software
* high performance and simplicity is achieved
by employing state of the art devices (Transputer,A110,AD9502,FastSram)
* low cost (entry boards cost: 10k for stereo 60Hz, without A110 Filter Bd.)
As Cyclops incorporates the transputer based XP/DCS, it ideally matches
our XP/DCS based control system for interconnection
via communication links.
Furthermore a distributed vision system
architecture connecting to a distributed control topology has another
advantage for many applications:
Results that are computed in one part of
the vision system can be provided directly to the appropriate part of the
control topology.
Cyclops consists of five main elements: A camera, a digitizer board,
a (optional) filter board and one or many parallel video memory boards,
each of which connects to a XP/DCS CPU board.
The functions are straightforeward --- the camera provides interlaced fields
of video data to the digitizer, which converts the analog signal into an
eight bit pixel stream, accompanied with some synchronization signals.
The video bus containing the pixel stream is brought to the filter board.
Here any two dimensional filter or convolution with a 6x14 pixel size window
can be applied to the data in real time. Finally the video data is written
into one or several video memory boards,
where it can be accessed by its XP/DCS transputer.
All memory boards attach to the video bus in parallel.
Only its drive and noise limitations
restricts the maximum number of video memory boards.
Even though Cyclops satisfies a specific need, it is very flexible in
its operation so as to allow its application to most vision tasks.
In order to illustrate its generality, we will describe three different ways
to distribute the video fields to the video memory boards
and the implications for update rate and latency.
By update rate we mean the rate at which results
are computed by any of the processing boards on the video bus whereas
latency refers to the total processing time from raw video data to end result.
**Group Mode: the video fields are loaded into all memory boards
simultaneously.
Depending on the application, each processor then picks
a dynamic or static partition of the full field for processing.
A new field is loaded after all processors have finished processing --- thus
latency is equal to update rate.
**Scan Mode: for many applications, the video data is not easily
separable into
partitions without requiring extensive inter-processor communication
and decrease performance. In this case one can achieve an
increased update rate while keeping the same latency as with only one memory
board by sequencially scanning successive fields into different memory boards.
Maximally eight memory boards can receive different fields.
**Mixed Mode: this mode combines scan and group mode.
Successive fields are scanned into groups of boards as opposed to into single
board as in scan mode. This reduces
latency when compared to the scan mode.
A monocular Cyclops (without the filter board, which is currently
under construction) was built and tested
by Taylor and Vlamis as a semester project for a computer
architecture class offered during fall 1988 by the last author.
It consists of one digitizer board and one video memory board, and tracked
two ping pong balls at 30 Hz. (Using very unsophisticated initial software and
dark background. With two video memory boards the update rate is 60 Hz, the
latency worst case is 1/30 sec). So we have not pushed the performance yet.
The critical ingredients that made this project a
successful one were the existing design experience with the transputer
in the Yale robotics lab, the availability of the XP/DCS CPU board,
high speed CCD cameras with
electronic shutter, single chip eight bit video digitizers, INMOS 2d transform
integrated circuits (A110), and affordable high speed static memories.
===============================================================================