Taking SLAM Outdoors: Mapping and Viewing the World

Most visual monocular SLAM has been conducted in indoor office-like environments, and although it can work quite well under these conditions, significant differences in the type of scene encountered outdoors lead to a lot of challenges. Issues we intend to address include incorrect depth estimates for large distances, scaling of maps to very wide areas, dealing with moving objects and changing brightness, and capturing scenes at multiple levels of detail.

We are also working to give the SLAM maps a more meaningful representation for humans. Traditionally the maps consist of a sparse set of points, appropriate for robot localisation but impossible for a human to tell what they are looking at. For purposes of remote viewing, real-time scene modeling, virtual walkthroughs and so on, it would be useful to add visual information to the map - including a background image sphere, textured surfaces, and rendered graphics representing scene objects. Adding higher level structure will help make scenes recognisable, as well as simplifying the stored representation. Ideally this will make it possible to synthesise novel views and generate semi-realistic virtual and augmented environments in real-time.