00100					1   August  1969
00200	
00300		CART VISION PROJECT PROPOSAL
00400	
00500	Dear John,
00600			At last week's cart-meeting you seemed to feel that there
00700	was a lack-of-visible-progress with respect to the cart project.
00800	You seemed to want promises to be made in order that progress can be
00900	measured.  Further on, you attempted to derive a strategy for the project
01000	to follow from gedanken experiments  -  strategies such as "simulation";
01100	thought-experiments such as "car-passing-on-hill".  And finally, as an
01200	off-hand remark, you asked that if there were twenty people (who you
01300	would have hired had you known what to tell them to do) then what would they
01400	be doing.
01500	
01600	
01700	PROMISES
01800	
01900	I promise:
02000		I.   To use realistic road data. Film, cart and a camera on a cable;
02100			I will collect road data thru color filters.
02200		II.  To write an image enhancer program  -  to increase contrast
02300			using JPL type operators, to focus by stealing Sobel & Tenebaum's
02400			routine, and to average over a number of frames.
02500		III. To write an image reduction program  -  reduces an image to a set
02600			of possibly overlapping blobs each circumscribed by a polygon
02700			and characterized by various properties such as area, average
02800			intensity, moment, convexity, perimeter-length, homogeneity,
02900			and the like  -  by extending my FLIP program somewhat.
03000		IV. To write an image extrapolator program - merely extends polygon-blobs
03100			beyond edges of image & guesses crude depths from focus setting.
03200		V. To write a 2D to 3D converter - Take extrapolated image-polygon & depth
03300			guess thru an inverse perspective transform with respect to lens-angle,
03400			pan,tilt and camera location.
03500		VI. To write a camera controller or use Rod's cart controller
03600			or to merely move the camera (say down the road) or to change the slide.
03700			and to issue commands or advice to the Cart controller with respect
03800			to goals or detection of obstacles.
03900		VII. To write an image predictor that generates a 2D perspective 
04000			image of the road with intensities derived for a single 
04100			light source such as the sun or a headlight & the polgon-blob
04200			properties.  I have already written and have working a hidden line
04300			eliminator.
04400		VIII.  To write a verifier-corrector - the perdicted image is
04500			checked against the previous and extrapolations are altered
04600			accordingly.
04700	
     

04800	STRATEGY
04900	
05000	I believe that attacking problems such as style-of-driving
05100	or what happens if we are passing a truck on a hill and ...   are premature
05200	when the robot-driver is as to date still blind.   Concentrating on vision
05300	with respect to roads - there are two
05400	problems The Image-Correlation-Problem and The World-Modeling-problem.
05500	The image-correlation-problem comes in several flavors: tracking motion of moving 
05600	objects (correlating the object in several images), correlating the Views from a moving vehicle,
05700	Windowing - that is patching together images collected from a pan of the camera,
05800	kinetic depth perception, and parallax.  The world-modeling-problem involves 
05900	a strategy of handling most everything in a similair manner as oppose
06000	to Feldman's world-model of handling afew  well know objects.
06100	I believe that the image-predictor-corrector is a realistic way to handle
06200	correlation type problems  -  and that the polygon-blob-property world
06300	model is a kludge that is used for want of a better ultimate high-level general
06400	representation  -  I really consider
06500	blob-polygons as an intermediate stage between visual data
06600	and semantic data, and would like to see a mapping or recognition step
06650	which converts polyblobs into nouns and adjectives.
06700	
06800	
06900	IF I HAD TWENTY PEOPLE or the money equivalent:
07000	
07100		I might be able to define & coordinate specific tasks with respect to:
07300			Image-enhancement
07400			Computer-Graphics (for image predicting and for debugging)
07500			Illumination Calculation
07600			World-Models (hand written world models, and 
07700				world model generators)
07800		With respect to hardware shopping:
07900			Radio Controlled Car
08000			Color Vision
08100			Video Image Output  -  for seeing results of digital averaging
08200					digital contrast, and illumination predictions.
08300			Hardcopy Graphical device faster than the plotter more resolution
08400					than the line printer.
08500			Ambulatory Robot
08600			Robot with an Eye-stalk (camera that can translate as well as pan/tilt).
08700