COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	COMPUTER VISION AND ITS RELEVANCE TO THE DEFENSE DEPARTMENT
C00021 ENDMK
C⊗;
COMPUTER VISION AND ITS RELEVANCE TO THE DEFENSE DEPARTMENT


	It  is convenient  to  divide work  in  computer vision  into
scientific  work and work  in basic  technology.  Scientific  work is
aimed at getting certain  knowledge, and basic technological work  is
aimed at  getting certain  kinds of  capability independently  of how
this capability  may be applied.  In the field of computer vision, as
in artificial intelligence generally, these efforts  are intertwined,
and  it is  often not clear  which goal  dominates in  the work  of a
particular group.


KNOWLEDGE

	In computer  vision, the  main scientific  questions are  the
following:

	1. What  kinds of information  about the world  are available
through vision?   Clearly only partial information  about a person is
available from vision when  you see him seated  behind a desk.   When
you see a machine at a distance even less information is available.

	2.   What  information about  the  world can  be assumed  and
combined  with  direct vision  information  to get  information about
three  dimensional  objects?    What  rules  determine  what  can  be
legitimately inferred or conjectured?

	3. What is  it reasonable to try to know  about an object and
how can this information be represented in the memory of a  computer?
Clearly this is different for a machine  part whose complete shape is
often determinable and  necessary and a tree or head of hair where it
is impossible  and  unnecessary to  determine  the location  of  each
individual hair  or leaf.   But  note that a  barber or  tree surgeon
while  not keeping  in mind  the location  of all hairs  or branches,
requires temporary information about particular hairs  or branches in
order to do the next step of his work.

	4. What  information about a  scene comes from  the different
visual characteristics?   What redundancy  is there?   By what  rules
can the different  cues be combined legitimately to  give information
about  a scene?  The  kinds of information that  have been considered
include brightness  and  brightness  edges, color  and  color  edges,
texture and regions  grown by combining small  regions of homogeneous
texture,    the  distance as  measured  by range  finders  or  by the
parallax of two views  together with parallax edges.   Recently there
as been  a lot of work  on so-called top-down approaches  where a lot
of information about what  objects may be presumed  to be is used  to
interpret lower level  information.  Thus  if an object  was presumed
to be a human, two and not more arms must be accounted for.


CAPABILITY

	The  main capabilities  computer  vision research  is working
towards are the following:

	1. The ability  to find objects such  as persons,   vehicles,
or  machine parts  in  a  complex environment  and  to determine  the
attitudes  of the objects found.   Thus we not only  need to find the
vehicles in  a scene  but we  also need to  know which  way they  are
going.   Machine parts must be located  oriented so that manipulators
can pick them up.

	2. The  ability to  give a  description of  a scene  that  is
complete with regard to those aspects relevant to a certain task.


APPLICATIONS

	Computer  vision has  many  potential applications  including
some of  special interest to the Department  of Defense. Some of them
concern remote control and some concern automation of production.

I. REMOTE CONTROL

	Before going into  detail,   it is necessary  to mention  one
major embarassment for the computer  vision enthusiast.  Namely,  the
situation   is  somewhat  analogous  to   early  discussions  of  the
employment of computers  for business and  inventory purposes.   Much
more elementary  things than the employment of  computers remained to
be done.

	For example, at an early conference on possible  applications
of computers  in  libraries, it  was pointed  out that  a very  large
improvement  could be made  without computers by  adopting a proposal
made by  Joseph  Henry in  1859  for  a uniform  national  cataloging
system.

	An analogous  role with regard  to computer vision  is played
by  remote operation  using television.   There are  still very large
payoffs  for  ordinary  remote  operation  not  involving  computers.
However, just  as with business and  with libraries, there  is also a
great potential payoff arising from the  use of computers, and it  is
not necessary to wait for all potential  remote control payoffs to be
realized  before  beginning the  study  and even  the  application of
computer vision.

	a.  Remotely operated aircraft.

	There are  two reasons  for wanting  the  ability to  control
aircraft by  computer as  well as  the ability  to pilot  it remotely
using  television.  First, communication cannot  be guaranteed if the
aircraft is flying low  or if there is  jamming.  To the  extent that
we rely on remotely  controlled aircraft, potential enemies will work
on jamming techniques.

	Second,   it  may  be  possible  to  accomplish  things  with
computer vision  that cannot  be done  by humans,   specifically  the
computer is potentially much faster than a person.

	We  see  several  tasks for  vision  in  remote operation  of
aircraft.  The first is  simply navigation.  In particular, it  seems
possible to write programs that will  navigate by pilotage,  i.e.  by
comparing what  it sees with a map.  This would permit flight at very
low altitude  through  valleys  and around  buildings,   etc.    This
capability should  be contrasted with that provided  much more simply
by terrain  avoidance radar  in which  the aircraft  follows a  fixed
ground  track   and  avoids  hitting   obstacles  by   climbing  when
necessary.    It is  not a  priori  clear how  much  application this
ability would have.   It  might be quite  important if  anti-aircraft
defenses develop in certain ways.

	The second  task is finding targets  and controlling weapons.
This  is an  area where  there is  a potential  very large  gain over
human performance.   Namely,  suppose that  it were possible to  fire
many individually  aimed shots per second.   This would revolutionize
strafing,  because  it would  then be  possible to  use much  smaller
attack aircraft containing a much smaller  amount of ammunition. Thus
there could be  many more of them and the acceptable loss rates would
be much higher and the acceptable exchange ratios with  targets could
be much  lower.   We believe  that it is  possible to  recognize such
targets  as people and vehicles,   and it is  possible to develop the
ability to  decide  in many  cases which  people are  to  be shot  at
better than a person can do it in the heat of battle.

	b. Remotely operated tanks.

	We envisage  these as small,  inexpensive,  and  used in very
large numbers.    The  problem  of  automatically  driving  a  ground
vehicle is  more  difficult than  that of  automatically piloting  an
aircraft, because the  variety of situations is greater.  The problem
of automatically driving under combat conditions is easier  than that
of driving  a car. The  reason is that  almost all of  the artificial
intelligence required  to drive a car will be used to avoid accidents
under rare  conditions.   Combat  vehicles  will suffer  high  losses
anyway so  that a few  percent additional losses  because the driving
programs are not very smart will not be of major importance.

	c. Automatic sentries.

	Here we  get  into  an  area  where  automatic  methods  have
probably got  themselves a bad  name because  of the deficiencies  of
the  devices used in Vietnam.   In our  opinion,  this  is because an
attempt was made to  be too clever, to  use inadequate sensors,   and
to rely  excessively on algorithms  for interpreting output  of these
sensors which  were not really general.  We believe it is possible to
do much better, but we don't claim it is easy.

	The application of computers  to sentry duty arises  from the
fact that  humans have difficulty in paying  continuous attention for
a long time when nothing is  happening.  Computers are quite good  at
this. Moreover, automatic sentries  can be posted quite far  from the
unit  they are protecting,   and high losses will  be acceptable.  In
our view, an automatic sentry  would consist of a computer,   several
low light  and infra red  television cameras, whatever  other sensors
might  prove useful,   communication  back to  base for  pictures and
other data, and  weapons that can either  be fired as aimed  from the
base or automatically according to  the control of the computer.  The
amount of  discretion  allocated  to the  programs  would  depend  on
circumstances.


DEVELOPMENT OF THE MILITARY APPLICATIONS

	The   present  organizational   framework   of   unclassified
research  in universities and  other non-profit  organizations is the
arrangement best capable  of keepin the  U.S.  in  the lead in  basic
research and  basic technology.  Even  though the U.S. work  in these
areas is  unclassified, our lead over other countries, especially the
Soviet Union  has incrased.   Classifiying the  work would so  reduce
the  population available for  thinking about  the problems  that our
lead would be slowed.

	 On  the other  hand,   under  present social  and  political
circumstances, these  organizations are  not suitable for  developing
direct  military applications.  It  will be necessary  to monitor the
basic science and  technology and  decide when military  applications
should be  undertaken. Even  the existence  of these  projects should
probably be classified.


II. NON MILITARY APPLICATIONS OF INTEREST TO THE DEFENSE DEPARTMENT

	Besides   direct  military  applications,    there  are  many
applications  of  computer vision  of  that  will  help  the  Defense
Department  do its  work better.   There  is much  interest  in these
applications  in  the   artificial  intelligence   community,     and
capability  generated  in  developing  these   applications  will  be
available for other uses.

	In  the main, these  concern the need  to be  able to develop
hardware rapidly and to bring small quantities of equipment  for test
into  use as  quickly  as possible.   The  development  of a  general
mechanical  assembly  system  and  the  further  automation  of other
aspects of production such as machining and forming will  be aided by
computer  vision  and  other AI  systems.    Of  course,   not  every
application  requires  the  most  advanced  methods,    and  the   AI
researchers have recognized this.  Indeed  the Stanford and Edinburgh
assembly demonstrations, which  are the most advanced at present, use
few technique  which can  be classified  as artificial  intelligence.
The fact  is that the AI  laboratories are also the  most advanced in
real  time control of experiments in  a time-sharing environment, are
the  most  advanced   in  putting  new  applications   into  existing
time-sharing systems  and are also the most  advanced in conventional
programming techniques.  Besides this,  we are aware of the  boundary
between  AI and  conventional  techniques and  are  less likely  than
others  to get sucked into  trying to do something  the hard way when
experiment with AI  is not  the task  being undertaken.   We are  not
sure how to document this claim  of general competence,  but we think
our visitors come away with that impression.