Performance Improvement of Iterative Closest Point-Based Outdoor SLAM by Rotation Invariant Descriptors of Salient Regions

Abstract

To navigate in an unknown environment, a robot should build a model for the environment. For outdoor environments, a three-dimensional (3-D) map is usually used as a main model. This study considers outdoor simultaneous localization and mapping (SLAM) to build a global 3-D map by matching local 3-D maps. An iterative closest point (ICP) algorithm is used to match local 3-D maps and estimate a robot pose, but an alignment error is generated by the ICP algorithm due to the false selection of corresponding points. We propose a new method to extract 3-D points that are valid for ICP matching. Rotation-invariant descriptors are introduced for robust correspondence. 3-D environmental data acquired by tilting a 2-D laser scanner are used to build local 3-D maps. Experimental results in real environments show the increased accuracy of the ICP-based matching and a reduction in matching time.