buildings, or trees, to the scene description and thus determine what

is there. The model objects would be described in memory as having

certain characteristics, and the program would attempt to match these

against various parts of the image. Scene understanding involves a

more general recognition problem describing physical objects in a

scene based on images. For example, a scene may be divided into

regions that match various objects stored in memory such as a house,

tree, and road. Once the scene is divided into known regions, the

interrelationship between these regions provides information about the

scene as a whole.

 When it is necessary to recognize specific objects, correlation

techniques are often used.2 A reference image of the desired object

is stored and compared to the test image electronically. When the

correlation coefficient is over a specified threshold, the computer

interprets the image as containing the object. The correlation

procedure may also provide the location of the object in the scene and

enable tracking. The correlation coefficient may be used in decision

making to determine robot action. Because even a single object may

present itself in many ways, correlation procedures are complicated by

the immense reference file that must be maintained.3 Special

correlation techniques may provide invariance to specific changes, but

a wide range of object conditions (i.e., temperature, color, shape,

etc.) make correlation recognition a complicated computer task.4 The

best computer vision systems now available have very primitive

capabilities. Vision is difficult for a computer for a number of

reasons. The images received by a sensing device do not contain

sufficient information to construct an unambiguous description of the