Simultaneous Localization and Mapping ( SLAM ) : Part II BY TIM BAILEY AND

Simultaneous localization and mapping (SLAM) is the process by which a mobile robot can build a map of the environment and, at the same time, use this map to compute its location. The past decade has seen rapid and exciting progress in solving the SLAM problem together with many compelling implementations of SLAM methods. The great majority of work has focused on improving computational efficiency while ensuring consistent and accurate estimates for the map and vehicle pose. However, there has also been much research on issues such as nonlinearity, data association, and landmark characterization, all of which are vital in achieving a practical and robust SLAM implementation. This tutorial focuses on the recursive Bayesian formulation of the SLAM problem in which probability distributions or estimates of absolute or relative locations of landmarks and vehicle pose are obtained. Part I of this tutorial (IEEE Robotics & Auomation Magazine, vol. 13, no. 2) surveyed the development of the essential SLAM algorithm in state-space and particle-filter form, described a number of key implementations, and cited locations of source code and real-world data for evaluation of SLAM algorithms. Part II of this tutorial (this article), surveys the current state of the art in SLAM research with a focus on three key areas: computational complexity, data association, and environment representation. Much of the mathematical notation and essential concepts used in this article are defined in Part I of this tutorial and, therefore, are not repeated here. SLAM, in its naive form, scales quadratically with the number of landmarks in a map. For real-time implementation, this scaling is potentially a substantial limitation in the use of SLAM methods. The complexity section surveys the many approaches that have been developed to reduce this complexity. These include linear-time state augmentation, sparsification in information form, partitioned updates, and submapping methods. A second major hurdle to overcome in the implementation of SLAM methods is to correctly associate observations of landmarks with landmarks held in the map. Incorrect association can lead to catastrophic failure of the SLAM algorithm. Data association is particularly important when a vehicle returns to a previously mapped region after a long excursion, the so-called loop-closure problem. The data association section surveys current data association methods used in SLAM. These include batch-validation methods that exploit constraints inherent in the SLAM formulation, appearance-based methods, and multihypothesis techniques. The third development discussed in this tutorial is the trend towards richer appearance-based models of landmarks and maps. While initially motivated by problems in data association and loop closure, these methods have resulted in qualitatively different methods of describing the SLAM problem, focusing on trajectory estimation rather than landmark estimation. The environment representation section surveys current developments in this area along a number of lines, including delayed mapping, the use of nongeometric landmarks, and trajectory estimation methods. SLAM methods have now reached a state of considerable maturity. Future challenges will center on methods enabling large-scale implementations in increasingly unstructured environments and especially in situations where GPS-like solutions are unavailable or unreliable: in urban canyons, under foliage, under water, or on remote planets.