A multisensor SLAM for dense maps of large scale environments under poor lighting conditions

Abstract

This thesis describes the development and implementation of a multisensor large scale autonomous mapping system for surveying tasks in underground mines. The hazardous nature of the underground mining industry has resulted in a push towards autonomous solutions to the most dangerous operations, including surveying tasks. Many existing autonomous mapping techniques rely on approaches to the Simultaneous Localization and Mapping (SLAM) problem which are not suited to the extreme characteristics of active underground mining environments. Our proposed multisensor system has been designed from the outset to address the unique challenges associated with underground SLAM. The robustness, self-containment and portability of the system maximize the potential applications.The multisensor mapping solution proposed as a result of this work is based on a fusion of omnidirectional bearing-only vision-based localization and 3D laser point cloud registration. By combining these two SLAM techniques it is possible to achieve some of the advantages of both approaches – the real-time attributes of vision-based SLAM and the dense, high precision maps obtained through 3D lasers. The result is a viable autonomous mapping solution suitable for application in challenging underground mining environments.A further improvement to the robustness of the proposed multisensor SLAM system is a consequence of incorporating colour information into vision-based localization. Underground mining environments are often dominated by dynamic sources of illumination which can cause inconsistent feature motion during localization. Colour information is utilized to identify and remove features resulting from illumination artefacts and to improve the monochrome based feature matching between frames.Finally, the proposed multisensor mapping system is implemented and evaluated in both above ground and underground scenarios. The resulting large scale maps contained a maximum offset error of ±30mm for mapping tasks with lengths over 100m.