00100	Our study of curved and complex objects  uses  depth  information  as
00200	stereo  is used in human vision.  Depth data is obtained from a laser
00300	triangulation system, implemented by Gerald Agin.  The system casts a
00400	plane  of  light  on  the  scene;  illuminated points are viewed by a
00500	television camera.  The laser functions as a  bright  light;  because
00600	the  light  is  monochromatic,  we  can use an interference filter to
00700	operate in roomlight.  Because the beam is parallel,  we  have  large
00800	depth of field.  The laser beam is not powerful; sunlight is 40 times
00900	brighter.
01000	
01100	Here we see  the  deflection  assembly  and  one  of  our  television
01200	cameras;  the  laser itself is under the table.  The periscope brings
01300	the light up above the table to a mirror which bends  the  beam  into
01400	the  deflection  assembly.   The  beam  is  approximately  two  mm in
01500	diameter; it appears much larger on the  film.   A  cylindrical  lens
01600	diverges  the  beam  into  a  plane,  which  is then deflected by the
01700	rotating mirror.
01800	
01900	The plane intersects objects along some plane curve,  viewed  by  the
02000	television  camera.   Depth  discontinuities  in space, correspond to
02100	breaks in the image curve. Slope discontinuities in space, correspond
02200	to slope discontinuities in the image.  Since we know the angles of
02300	both the beam and the camera, we know the distance of the object.
02400	We know the relative depth to an accuracy of 1mm, limited by the
02500	camera resolution and the apparent width of the beam.
02600	
02700	Now we see a sequence of laser lines as seen by the camera and
02800	displayed on a graphic console.  The centers of the lines are
02900	found and smoothed, and then fit with straight line or ellipse
03000	segments.  The plane sweeps across the scene in two orientations.
03100	We see the sequence for a snake-like object.
03200	
03300	Now we see the same process performed on a baby doll.
03400	We can show the three-dimensional figure as it would appear rotated
03500	in space.  Since the camera can see only the front half of the
03600	object, the figure looks like a hollow shell.
03700	
03800	Our aim is a perceptual system which makes its own models of objects
03900	in a symbolic representation.  The representation depends on a part/whole
03925	segmentation.
03950	Our representation is three-dimensional
04000	and based on volume primitives, rather than surface primitives.  That is,
04100	our primitives are arm-like parts, rather than surface patches.  The
04200	power of the representation is the part/whole decomposition; the choice
04300	of primitives is important to implement a good choice of parts.  We have
04400	used primitives called "generalized cylinders" which are defined by
04500	cross-sections defined along a space curve, called the axis.  These
04600	primitives include locally conical, locally helical and locally toroidal
04700	shapes.  
04800	
04900	A preliminary segmentation of objects chains together laser traces which
05000	are adjacent, parallel, and have similar lengths.  These indicate
05100	that the object has a smoothly changing cross section.
05200	
05300	The midpoints of traces provide an initial guess for the axis.  
05400	An iterative procedure interpolates circular cross sections normal to the
05500	approximate axis, then determines a better guess for the axis
05600	and repeats.  In the process, the generalized cylinder is extended
05700	to segment the object at places where the diameter of the cross
05800	section changes dramatically.  
05900	
06000	We see the iterative cylinder tracing applied to the snake
06100	and the doll.  The doll is described in terms of six cylinders,
06200	one each for the body and head, and two each for the arms and legs.
06300	
06400	Work is underway to deal with more general cross sections
06500	than circular; the program is being extended to link together
06600	primitives into a complete skeleton.  Object recognition will use
06700	these descriptions.