Mapping object movement to aerial mosaics

This paper addresses the task of making a mosaic map with moving object information from images gathered by a downward-looking camera on an airborne platform so that the picture of the ground movement can be exploited by higher level analysis. We have previously successfully shown in simulations that the bundle adjustment techniques results in consistent, undistorted maps. In this paper techniques for feature matching between overlapping, but non-sequential, images, where track information is not available, have been developed in order to apply bundle adjustment on loop closure with real data. These techniques are based on geometric and appearance based feature matching. Tracks of moving objects are also placed on the mosaic image using the homographies of in-between frames and performing local bundle adjustment. The algorithms have also been extended to use a graph decomposition technique in order to provide a scalable solution to real scenarios.