A Multi-Stage Stereo Matching Algorithm Based on Geometry Constraints

A geometry constraint applicable to natural terrain is derived based on the analysis of lunar surface. With this constraint a multi-stage stereo matching algorithm is presented to provide the fairly accurate and dense disparity map needed for lunar rover autonomous navigation. Points with larger intensity gradient are matched with a multi-feature matching strategy in earlier stages. Then the information of matched points is used to limit search region in the matching of other points in later stages within the help of several geometry constraints. Finally the unmatched points are dealt with to resolve the occlusion caused by large raised objects. Experimental results with real images of various types of natural terrain indicate that the algorithm improves the accuracy of dense disparity maps effectively, which lays the foundation for the reconstruction of 3D terrain map