Describing, Navigating and Recognising Urban Spaces - Building an End-to-End SLAM System

摘要

This paper describes a body of work being undertaken by our research group aimed at extending the utility and reach of mobile navigation and mapping. Rather than dwell on SLAM estimation (which has received ample attention over past years), we examine sibling problems which remain central to the mobile autonomy agenda. We consider the problem detecting loop-closure from an extensible, appearance-based probabilistic view point and the use of visual geometry to impose topological constraints. We also consider issues concerning the intrinsic quality of 3D range data/maps and finally describe our progress towards substantially enhancing the semantic value of built maps through scene de-construction and labeling.