This paper covers the global aspects of a new SLAM framework introduced in . The map is a graph of overlapping range views, linked by multiple uncertain hypotheses of motion and correspondence. They form the interface between local pose estimation and globally consistent mapping. In order to prune the hypotheses and reduce the brittleness of the local algorithms, we propose a novel cooperation between image-based and odometric motion estimates, and a geometric-probabilistic visibility model for oriented surface features which can also discern moving objects. Global loop closing works by exchanging hypotheses, priorized on a node ambiguity measure to bound the update complexity. The cycle error in frame space is used as a consistency criterion. The new concepts were tested and evaluated during several indoor exploration tours.
This paper introduces a new framework for incremental localization and mapping (3D-SLAM) that integrates submaps of surface features extracted from dense range images. Its local layer estimates unknown motion and feature associations between adjacent views. The global layer tries to ensure map consistency and to close loops, using multiple hypotheses offered by the local layer. The local algorithm crucially affects the map quality. An interpretation tree (IPT) is compared to Orthogonal Surface Assignment (OSA), a new algorithm tracing a building coordinate system inside man-made work spaces. During indoor experiments, performed with the rotating laser scanner RoSi, OSA proved much more reliable than IPT. Another experiment, running a simple on-line surface classification, indicates that the maps are useful for mission planning.