Geometric constraints for object detection and delineation - Jefferey Shufelt

The ability to extract generic 3D objects from images is a
crucial step towards automation of a variety of problems in
cartographic database compilation, industrial inspection and assembly,
and autonomous navigation. Many of these problem domains do not have
strong constraints on object shape or scene content, presenting
serious obstacles for the development of robust object detection and
delineation techniques. Geometric Constraints for Object
Detection and Delineation addresses these problems with a
suite of novel methods and techniques for detecting and delineating
generic objects in images of complex scenes, and applies them to the
specific task of building detection and delineation from monocular
aerial imagery.

PIVOT, the fully automated system implementing these techniques, is
quantitatively evaluated on 83 images covering 18 test scenes, and
compared to three existing systems for building extraction. The
results highlight the performance improvements possible with rigorous
photogrammetric camera modeling, primitive-based object
representations, and geometric constraints derived from their
combination. PIVOT's performance illustrates the implications of a
clearly articulated set of philosophical principles, taking a
significant step towards automatic detection and delineation of 3D
objects in real-world environments.

Geometric Constraints for Object Detection and Delineation is
suitable as a textbook or as a secondary text for a graduate-level
course, and as a reference for researchers and practitioners in
industry.