A modular hybrid SLAM for the 3D mapping of large scale environments

Abstract

Underground mining environments pose many unique challenges to the task of creating extensive, survey quality 3D maps. The extreme characteristics of such environments require a modular mapping solution which has no dependency on Global Positioning Systems (GPS), physical odometry, a priori information or motion model simplification. These restrictions rule out many existing 3D mapping approaches. This work examines a hybrid approach to mapping, fusing omnidirectional vision and 3D range data to produce an automatically registered, accurate and dense 3D map. A series of discrete 3D laser scans are registered through a combination of vision based bearing-only localization and scan matching with the Iterative Closest Point (ICP) algorithm. Depth information provided by the laser scans is used to correctly scale the bearing-only feature map, which in turn supplies an initial pose estimate for a registration algorithm to build the 3D map and correct localization drift. The resulting extensive maps require no external instrumentation or a priori information. Preliminary testing demonstrated the ability of the hybrid system to produce a highly accurate 3D map of an extensive indoor space.