@device(postscript)
@libraryfile(Mathematics10)
@libraryfile(Accents)
@style(fontfamily=timesroman,fontscale=11)
@pagefooting(immediate, left "@c<Technical Report 1996>",
center "@c<Computer Science>",
right "@c<Carnegie Mellon>")
@heading(Projective Geometry and Photometry for Object Detection and Delineation)
@heading(CMU-CS-96-164)
@center(@b(Jefferey A. Shufelt))
@center(July 1996 - Ph.D. Thesis)
@center(FTP: Unavailable)
@blankspace(1)
@begin(text)

Computer vision systems have traditionally performed most effectively in
constrained situations, where limitations on object shape or scene structure
permit reliable image analysis. For example, in model-based recognition
systems, the existence of 3D models of objects of interest allows the
application of geometric constraints to limit the search for interpretations
of low-level image information.

Many problem domains, however, do not have explicit constraints on object
shape or scene content. In aerial image analysis, man-made structures take
on a wide variety of shapes and sizes. Existing techniques for these domains
obtain only partial solutions by the use of simplifying assumptions about
imaging geometry, illumination conditions, and object shape. Few of these
techniques attempt to model perspective or photometric effects, which can be
powerful constraints for object detection and delineation.

The central hypothesis of this work, that rigorous modeling of the image
acquisition process leads to improved detection and delineation of basic
volumetric forms for object recognition, leads to the formulation of a set
of principles for object detection and delineation. In accordance with these
principles, a fully automated monocular image analysis system, PIVOT, was
developed for the task domain of cartographic building extraction from
aerial imagery, using original techniques for vanishing point detection,
intermediate feature generation, hypothesis generation, and building model
verification.

A quantitative comparative evaluation methodology for object detection and
delineation is presented in this work, using unbiased image space and object
space performance metrics on large datasets of imagery. Using this
methodology, PIVOT was compared to three existing building extraction
systems on 83 test images covering a wide variety of geographical areas,
object complexities, and viewing angles. This analysis demonstrates PIVOT's
improved performance from highly oblique viewpoints and on complex manmade
structures, establishing the utility of rigorous camera modeling for object
detection and delineation tasks, and in particular its importance for the
automated population of spatial databases with cartographically accurate
three-dimensional models.

@blankspace(2line)
@begin(transparent,size=10)
@b(Keywords:@ )@c<Computer vision, cartographic feature extraction, object
detection, aerial image interpretation, building detection, monocular image
analysis, photogrammetry, projective geometry, vanishing points,
model-based recognition>
@end(transparent)
@blankspace(1line)
@end(text)
@flushright(@b[(284 pages)])