Building Recognition from Aerial Images by Using Abduction
Loading...
Date
1997-08
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Ohio State University. Division of Geodetic Science
Abstract
Buildings are found in almost all kinds of cartographic mapping, and their robust detection and delineation require techniques that exploit knowledge about man-made features. This research proposes an approach using 3-D surface patches to solve the problem of identifying buildings from aerial images. It is based on both data-driven and hypotheses-driven processing to explain the surface patches. Based on the characteristics of buildings and Schenk's model of layered abductive building recognition, a structure-based generic building model, top-down abductive reasoning, and two phases of bottom-up processes were proposed. A structure-based generic building model focuses on part/whole and spatial/geometic relationships between buildings and nearby objects. The top-down abductive reasoning is based on the unique property of aerial images, namely, the top view of a scene, which implies that finding a building is to find roofs of a building. And the roofs are hypothesized with 3-D flat or curved surfaces. The two phases of bottom-up processes include (1) reconstructing surfaces from a pair of aerial images, (2) detecting humps from the reconstructed surface, and (3) computing properties of surface patches. Finally, four suggestions are made for further research: using multi-sensor data, introducing reliable weights associated with hypotheses, more clearly describing task-specific knowledge, and researching machine vision problems.