Researchers have reported some progress in automatically building maps from rich sensor data, but they have paid little attention to taking advantage of existing knowledge structures. Teller [13] reports on Argus, a system that also infers knowledge from unlabelled data, but that does not have the advantage of preexisting background knowledge. Argus constructs a 3D model of a scene from a series of digital images taken by a mobile camera platform. As in our effort, he acknowledges the need for an absolute reference system to build the database, and, like our system, Argus uses GPS traces. The system employs a GPS receiver on the camera platform to establish the absolute coordinate system. In this case, however, the positions themselves only provide a reference point for processing, and the principal algorithms operate on the images.
Automated mapping approaches have also focused on special-purpose, labor-intensive efforts to exhaustively map a target area. For example, the GPSVan [4] combines many sensors, including multiple GPS receivers, laser cameras, and stereo vision, to capture detailed information about the roadways it travels. However, the system is prohibitively expensive and requires dedicated personnel to encode features as the vehicle drives. The fields of machine learning and data mining have examined techniques to extract useful knowledge from large data sets not specifically designed to support modeling a particular phenomenon, making up in volume what is lacking in focus and detail. This approach has the potential to reduce mapping costs, covering a target area roughly at first, then with higher precision as more data become available.