A knowledge based approach to automatic image registration

The presented work addresses the problem of automatic control point matching for the registration of remotely sensed images. The inaccuracy of flight parameters and the sensor specific appearance of objects are the difficulties automatic registration suffers from. To overcome these problems the presented system uses prior knowledge to select appropriate structures for matching, i.e. control points, from a GIS and to extract their corresponding features from the sensor data. The knowledge is represented explicitly using semantic nets and rules. The best correspondence between the GIS data and the image is found by an A*-algorithm. The automatic control point matching is demonstrated for crossroads in aerial and SAR imagery