Elastic view graphs: A new framework for sequential 3D-SLAM

Abstract

This paper presents a new approach for incremental 3D SLAM from segmented range images with unknown feature association. For feature and motion tracking, an any-time interpretation tree is compared to a new algorithm called Orthogonal Surface Assignment. Both algorithms may utilize uncertain pose estimates from vehicle sensors, and yield several hypotheses. An elastic graph of surface submaps handles global pose correction. By stochastic error propagation and spatial indexing, loops are detected within bounded time. Loop closing with hypothesis exchange is priorized on a measure of node ambiguity. Indoor experiments performed with a rotating laser scanner demonstrate the effectiveness of our approach.