Object modeling and matching from multi-view ground images for automated Mars rover localization

This paper presents an innovative method for object modeling and matching from multi-view ground images for automated Mars rover localization. In this method, rocks are first extracted from a selection of 3D ground points that are generated from dense matching. The extracted rocks are then modeled using analytical surfaces such as ellipsoids, hemispheres, cones, and tetrahedrons. The extracted rocks of two rover stations are matched through a robust algorithm that matches the geometric configuration patterns of the rocks from the two stations using an improved Hough transform technique, followed by a heuristic refinement. Finally, peaks of the matched rocks serve as cross-site tie points in bundle adjustment of the rover image network. Experiments conducted using Navcam images acquired from the 2003 Mars Exploration Rover mission have demonstrated that the proposed method is capable of selecting cross-site tie points for two rover stations that are 26 m apart. The issues of integration of visual odometry with bundle-adjustment are also briefly discussed