Feature Extraction Papers

 
Research on Cartographic Feature Extraction

A necessary step towards the automatic compilation of cartographic databases in urban and suburban areas is the ability to detect and delineate manmade structures. A system which can robustly and accurately solve this problem must be able to handle a wide variety of viewing angles, object shape complexity, object density, object occlusions, and shadow effects.

A common theme of our work in feature extraction is that no one technique can solve all of the detection and delineation problems that arise in aerial images. For instance, a shadow-based building detection is useless when the sun is directly above the imaged scene; line-corner analysis is ineffective when building roofs have the same intensity as their immediate surroundings. To address this problem, we employ the cooperative methods paradigm in many of our research systems. This means that multiple methods are employed and their results are combined in a principled manner. Examples of the use of this information fusion paradigm include monocular fusion of building boundary cues across multiple images, refinement of stereo disparity estimates using intensity/surface material information, refinement of multispectral imagery using high resolution panchromatic imagery, and road tracking using independent surface and boundary information.

Another focus of our work is the use of rigorous modeling of the image acquisition process. The utility of image acquisition knowledge has been largely unexplored in the computer vision community; our use of photogrammetric techniques provides additional leverage for cartographic feature extraction problems. In particular, these techniques provide a source of valuable geometric constraints for detecting manmade structure, and they permit modeling to take place in object space, providing a common ground for fusing information from stereo and multispectral analysis.

We continue to work on a variety of projects in cartographic feature extraction, including road network detection and analysis; line-corner based building detection; shadow-based building detection, verification and grouping; monocular fusion of building hypotheses; photogrammetric vanishing point analysis for building detection; multi-image feature matching for building delineation; and semi-automated model matching techniques.