COMMENT ⊗   VALID 00009 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	file system structured by programs
C00004 00003			hand/eye personnel
C00006 00004		total personnel
C00007 00005			done
C00008 00006			todo
C00009 00007			todo
C00012 00008			todo
C00016 00009	How can we improve the tv input chain?  
C00024 ENDMK
C⊗;
file system structured by programs
  modify ETV to keep track
  a program to make a file out of ETV header pages

keep John in touch with what we are doing

ERTS imagery
Biomedical imagery
papers to Sylvia Weir
sensor study
standard ccd camera
tube for Cohu
Geomed as programming language
Spine to polyhedra representation
FRANK HONEY
MARTY KNUTSON
415-965-5340
AIRBORNE SCIENCE OFFICE
	↓
DON MULHOLLAND
	↓
MIKE BADER-Div Chief

		hand/eye personnel

			Staff

MANIPULATION		VISION			STRATEGIES	MISC
     			Binford	


			STUDENTS
Scheinman
Finkel (HAL)		Bolles
Shimano (force)		       			Taylor
Takeda (touch)		Ganapathy			
Silva			Schukin

			Karl Kortkamp(hardware)


ARPA proposal
Bo
Gafford
Hieronymus
Quam
Bill Clancy--Don Gennery
Ganapathy
Moravec
Scheinman
Schukin
Taylor
Wilson
Zingheim

not on any proposal
Dan Silva
Dave Grossman
student for Liu
Liu free


NSF proposal
Binford
Taylor $14000
Pingle 1/2 $8670
Gafford 1/2 $7200
Scheinman student
Shimano student
open student

requests for funding
Jo Mei Chang (semi-promised)
Kicha
Cathy Heising


	total personnel
total proposed
Hieronymus			staff
Bill Clancy--Don Gennery	student
Moravec				student
Scheinman			student
Bo				overhead
Gafford				overhead
Wilson				overhead
Zingheim			overhead
Binford				staff
Taylor				student until January
Shimano 			student
*Quam				Grossman
*open student			Silva
*Ganapathy			open
*Schukin			open
*Taylor $14000			open
*Pingle 1/2 $8670		open
*Gafford 1/2 $7200		open duplicate
*Scheinman student		open duplicate
total open $45000
		done
		progress in hand/eye
March
  Takeda touch sensor analysis scheme
  Miyamoto display
  Bolles prediction for visual feedback
  Chris Jacko touch sensor research results (NSF)
  Takeda measured sensitivity of touch sensor
        1 gm gave reading difference of 84+-9.
        sensitivity about .1 gm
  TOB study CCD signal processors
		todo
1.  VISION LANGUAGE
2.  SYSTEMS
	Reddy-like
	Yoram
3.  SITE VISIT
5.  DISPLAY OBJECTS
		todo
  		HARDWARE
-1  HARDWARE PEAK DETECTOR FOR LASER EXPT.  TWO SETS OF 3 BUFFERS.
0  DRAWINGS ARM,TTL I/O BUS,DAC TARGET CONTROL, TURNTABLE CONTROL
	TURRET CONTROL,LASER MIRROR CONTROLLER
	GET 1 MONTH OF ELECTRONIC DRAFTSMAN FROM JOB SHOP
1	TV
2	FORCE-SENSING
3	MOD FOR LASER EXPT
4	MOVE LASER AND BEAM SWITCHYARD
5	NEW ARM ASSEMBLY
6	NEW ARM SWITCH
7	SEPARATE ELECTRONICS NEW ARM
	CLOSED-LOOP, GAIN AND VELOCITY SPECIFIED BY COMPUTER
	VELOCITY CONTROL LOOP
	MAKE SURE THAT THIS CAN BENEFIT COLLISION REFLEX
8	DEBUG A/D AND POTS FOR ARM
	SEPARATE 14 BIT A/D?
9	DEBUG TV DIGITIZER AGAIN
10	NEW HAND W/ REDUNDANT MOTIONS
	IMPROVED VERSION
11	SHAFT ENCODERS
12	NEW TOUCH SENSORS
13	TOUCH INTERRUPT
14	5 FINGER HAND 
15	PDP-6 TTL I/O BUS
	PDP-6 INTERRUPTS
16	UNDERFLOW MOD FOR PDP-10
17	TOOL HOLDER
18	TACHS ON JOINTS
19	FORCE SENSING ON OTHER JOINTS
21	NEW A/D FOR TV

22	QUAM ARITHMETIC BOX
	NOT PROPERLY IN OUR DOMAIN
23	panel of motor drivers of different power
	as standard interface with mechanical devices.


1.  construction algorithms (simple)
	as Hitachi
    arches and hard things
2.  assemblies
3.  touch and force
    hardware queue
4.  color constancy
5.  laser for hand/eye
6.  world model for curved objectx
	manipulation
7. avoidance
8. retrieval of visual memory



		todo
1. A is above B if vertically above in the image.
	indicates horiz surface
2. Symmetry of non-vertical supports (legs)
	equal length of non-vertical legs
3. 2↑n sums on tangents
4. in a net, only link edges and regions with certain relations
	concurrency, T's (in front of) etc
5. propagation techniques in curves and edge operators
6.  Check vs threshold in multi-entry at each entry.  Mark those
	above threshold in a list. Use a 3-d for circles.  We don't
	know enough to do so.
7. Biomedical cognitive memory
8. Use Buchanan-Luckham system for vision
10. Use vision to set up an assembly.  Minimize setup time,
	able to use big machine.
11. ARC examples for inspection language
12. Type checking records
13. motion project
14. self-determination of torque terms for arm
	torque=gravity torque+angular accel+friction
15. stereo calibration
16. shape repr
	use negative radius for negative volumes
	calc volume moments for spine-cross section
18. Barnea and Silverman algorithm shows considerable speedup
	for machine with those as hardware (add and test operation)
19. fast numerical solution to retinex
20. region samples based on an area representation
21. line scan sampling-tracking between
22. coarse cell pointers-proximity
	separate storage of lines and proximity of lines
23. shadows as depth cues
24. fast stereo as a proximity sensor
26. fourier focussing as proximity sensor for linear diode array
27. redundant, feature-oriented descriptions
28. represent fabrication operations
29. represent assembly operations
30. protocols for vision strategies
31. protocols for assembly
32. stereo simulation system for robot assembly programming
33. linear vs non-linear, sequential decision
34. given output of two Hueckel disks, combine info into a single
	curve...hierarchical combination!!!
35. hand held camera
36. Cal Quaite  Hansen Labs  ultrasonic imaging 
	may get around staining problems and allow imaging techniques
	in biology
37. formalization of cylindrical transform
38. spline fit to local cones from projection
39. continue main piece at joint

How can we improve the tv input chain?  
	input more bits
		video buffer
		scan at half the rate,same bandwidth
		digitize 7 bits, take lower and upper 4 bits in two passes
	improve tv signal
		higher signal/noise
			improve preamp
			improve tube current
			filter
		use available signal/noise over better range
Any reasonable way to get  in  more  bits  requires  a  7  or  8  bit
digitizer.  A buffer memory is a convenient thing to use.  I would be
happy to slow down the camera by a  factor  of  two  or  four  (where
resolution  is needed). The primary reason for wanting a video buffer
is for 3 color display, which  is  not  of  pressing  importance  for
vision.   For  display, we could again slow down the scan by a factor
of 3 to 6, provided persistence were adequate.

What intensity resolution can we expect from the camera?  We can  get
about  1000na  signal/2  na  noise.   Thus we have about 500/1 range.
These are distributed in a linear scale.  At high light  values,  the
s/n ratio is quite high.  We don't need all 500 levels.  To quantize,
we would like a linear scale at low light values (up to 32/1)  and  a
log  scale  beyond that. That is about 32xlog(17)+32=160 levels.  The
mods suggested below should improve signal/noise by about a factor  7
(now about 70/1).  The signal/noise ratio is low at low light levels;
we have suggested controlling light levels so that when  we  want  to
peer  into  the  dark, we can let in more light.  The signal/noise in
the dim areas will then be adequate, but the  bright  areas  will  be
driven  far beyond saturation.  Thus, we need a tube which can handle
large light overloads  without  affecting  the  tube  or  the  image.
Blooming  is  the  usual  symptom  even  for tubes which don't damage
easily.  One technique to control blooming is the anti-comet-tail gun
(Philips).   This  requires  mods  to  change  gun  focus on flyback.
Silicon vidicons  with  controlled  low  blooming  are  available  as
special items.   We have been told we can get one at a small premium.
A little blooming is probably tolerable.  This is because we need  to
increase light only for dark areas.  If they are next to a highlight,
the contrast is adequate without increase light.  So a small blooming
of the highlight does not hurt.

The silicon vidicon has a possible problem whose magnitude we have no
data on.  A few defects can be corrected for.  Any smooth shading can
be  corrected  for.   Pattern  noise  which  is  not smooth (salt and
pepper) must be less than a few percent.  We are  investigating  this
in contact with Philips Labs.

Besides changing the tube, we can improve the preamp and improve
the filtering.  Phase compensated filters have better transient
response (less ringing) than uncompensated filters.  These are
the easiest solution.  A delay line with multiple taps into an
integrator with t/n integration time is the most simple of the
time domain schemes we have considered.  Clearly, the time domain
and not the frequency domain is of interest to us.  Most of the
standard filter discussions center on frequency domain and introduce
bad things into the time domain.

Selection of an a/d is not complete.  We need 100 nsec sample time
if we are to speed up to 500 samples/line.  There is an advantage
in getting all the samples on one sweep, in that it requires less
precision in the timing for interleaving; the requirement now is that
the timing must be good to 1/500 to interleave.  The full horizontal
resolution is about 533 and vertical about 400 with 50% MTF.  Clearly
we want 7 or 8 bits.  We also want a reliable device.  The Computer
Labs a/d has been around and we have very good references on it.  The
Intercomputer Electronics a/d we have some doubts about because of
their performance with the d/a which we bought from them.  The other
two devices were not ready when we asked last, and we would like to
get detailed user references on them before we would be content.

camera interface
	a/d 						$3k to $6k
		50usec/100nsec/sample = 500 samples/line
		Phoenix $3800 8 bits 10 Mhz
			$2500 6 bits 10 Mhz
		
		Datel	$2500 8 bits 10 Mhz
		Computer Labs $5600 8 bits 10 Mhz
		Intercomputer Electronics  $7350 7 bits 10 Mhz
	buffer memory and display 8x256x333x3=2.05Mbits
						.8cents	$32k
	filters				$.5k
		delay line, multi-tap, 10 ns integrator
	  or 	phase-compensated filter

	log amplifier			$.2k
	silicon vidicon for Cohu	$1.5k
	update Cohu			$1k
		stabilize deflection system
		new preamp		$100?
		brakes
		iris control