Dense 3D reconstruction for video stabilization and georegistration

Aerial motion imagery georegistration and stabilization is recognized as a challenging problem over urban or mountainous regions and oblique viewpoints. Several current solutions solve this problem by using high-resolution geographic references such as a LiDAR DEM to account for motion parallax, however such data sets are not readily available over a vast majority of the world. This paper presents an automatic, hierarchical, area based 3D reconstruction technique capable of extracting a dense 3D surface from a motion imagery collection. The algorithm then uses the surface reconstruction to form predicted images, which are registered to the video frames, and solves for the corrected sensor model, including camera pose and interior parameters such as radial distortion and focal length. We present results for wide area motion imagery over rugged terrain as well as suburban terrain.