SLAM techniques in the context of 3D Mapping [lecture 2]

Abstract

6‐DoF registration methods can be used to integrate a sequence of range scans into a 3D map. But when doing this, individual errors in the registrations accumulate and the global accuracy in the map can become unboundedly error prone. Simultaneous Localization and Mapping (SLAM) is the remedy for this problem, which requires extending the registration methods with several additional techniques. First of all, there is the need for uncertainty measures that so to say indicate the quality of each registration. Second, there is the need for loop closing techniques, i.e., strategies to determine good candidates for registering non‐sequential scan pairs or so to say current data with previously seen parts of the environment. Finally, it is necessary to employ a global error minimization by suited optimization techniques, also known as the SLAM back‐end. The lecture discusses these aspects from the viewpoint of 3D mapping. In doing so, a special emphasis is laid on unstructured environments where applications benefit most from 3D mapping but where it is also the most challenging. Especially, the problem of ambiguous sensor data is discussed in this context. A novel form of graph‐based SLAM is presented, which can handle multi‐modal distributions in the registration results. It hence extends standard graphbased SLAM algorithms, which all cannot handle ambiguities by design. Furthermore, it is shown that the novel method outperforms particle‐filter SLAM, which in theory can handle ambiguities but in reality severely suffers in performance if they occur