perm filename NOTE.Q[LSP,BGB] blob
sn#010895 filedate 1972-11-09 generic text, type T, neo UTF8
00100 Cart Vision - November 1971 to November 1972 to November 1973.
00200
00300 B. g. Baumgart
00400
00500 My cart Vision work consists of a Geometric World Modeling
00600 program, GEOMED; and a television picture anaylsis program, CAREYE.
00700 Many essential smaller programs, such as camera locus solving, are
00800 embedded in these two large programs. My approach to visual
00900 perception may be characterized as "extreme mechanical deteminism",
01000 and is based on three assumptions:
01100
01200 i. I assume that given the location of the TV camera and the sun and
01300 given a faceted (not necessarily planar) polyhedron model of the
01400 world a computer can be made to compute a television image very
01500 nearly like the one produced by a real digital TV camera looking at a
01600 real world under a real sun.
01700
01800 ii. I assume that given two similar TV pictures (or region-edge
01900 pictures) a computer can be made to find the substantial similarities
02000 and the slight differences. And that the slight differences in a
02100 single image can be used to correct the camera and world models, and
02200 that slight differences accumulated over many images will add
02300 substantially to the world model and indeed may be characterized as
02400 visual perception.
02500
02600 iii. I assume that the `tabula rasa', world model initialization, can
02700 be done with human assisted programs. The process of constructing a
02800 computer model of the mundane physics and geometry of the real world
02900 is is not fully understood; and inspite of recent successes in
03000 automatic strategy planning it is difficult to automate a poorly
03100 understood process.
00100 Over the past year CAREYE has been recoded twice and GEOMED
00200 has been recoded four times. In GEOMED a new hidden line eliminator
00300 and geometric data structure have been developed; while in CAREYE,
00400 the design strategy which was to convert TV image data directly into
00500 region-edge data has evolved a middle step of converting TV images
00600 first into video intensity contours, similiar to those of krakauer,
00700 and then into region-edge data structures. Also this year I wrote an
00800 alternate reference manual for Micro Planner and assisted Richard
00900 Orban in installing my hidden line eliminator, OCCULT, into the
01000 Hand-Eye system. The present cart vehicle is in working order and has
01100 been exercised from time to time under manual remote radio control.
01200
01300 During the next year, I anticipate no new ideas and no
01400 recoding. I will implement my present ideas within existing system
01500 designs. In the coming year I will get my hidden line eliminator that
01600 predicts shadow edges and regions to work, and I will get my body
01700 locus solver that constructs 3D polyhedron models out of 2D image
01800 regions from several cameras to work, and I will get my two camera
01900 locus solvers to work (one based on region matching the other based
02000 on landmark-feature point matching).