Random finite sets for robot mapping and SLAM. New concepts in autonomous robotic map representations. (Q2275354)

Mathematical interpretations of the vehicle's environment which consider all the relevant uncertainty are of fundamental importance to an autonomous vehicle and its ability to function reliably within that environment. While a mathematical model which considers the vast complexities of the physical world remains an extremely challenging task, stochastic mathematical representations of a robots operating environment are widely adopted by the autonomous community. Probability densities on the chosen map representation are often derived and then recursively propagated in time via the Bayesian framework, using appropriate measurement likelihoods. Of crucial importance in autonomous navigation is the computational representation of a robot's surroundings and its uncertainty, referred to as the map. The presented book directly addresses this issue, initially in the area of feature based robotic mapping (FBRM) which assumes known robot location, and the area of feature based simultaneous localization and map building (SLAM). The book is divided into three parts. In Part I, the question ``Why use random finite sets?'' is addressed. Chapter 1, as an introduction, presents a detailed description of the contents of the book. Chapter 2 summarizes fundamental differences between random finite set (RFS) and vector based representations of features. The fundamental mathematical relationship between map states, observations, and vehicle pose are examined under both vector and FRS based frameworks. The issues of map representation, data association, map management, map error quantification, and the concise application of the Bayes theorem are summarized in this chapter. Chapter 3 introduces mathematical representation which can be used for RFSs. As in the case of vector based approaches, full Bayesian estimation in the space of features and robot trajectory is intractable for all realistic scenarios. This chapter poses the fundamental question, ``Given the posterior distribution of the map (and trajectory in the case of SLAM), what is the Bayes optimal estimate?''. Principal estimators are therefore presented here, which are capable of representing RFSs in a Bayes optimal manner. The probability hypothesis density (PHD) filter is introduced as one of the simplest approximations to Bayesian estimation with RFSs. This chapter therefore provides the foundations for most of the filtering algorithms for both FBRM and SLAM used throughout the book. Chapter 4, which coincides with Part II, presents new insights to motivate an RFS approach to mapping. By focussing on the mapping only problem, an estimation framework which yields Bayes optimal map estimates in the general case of unknown feature number, spurious sensor measurements, feature detection and data associations uncertainty are developed. Chapter 4 examines also in more detail the concept of FB map estimation error for the general case of an unknown number of features. Part III addresses the full SLAM problem. Chapter 5 offers a ``brute force'' solution to the SLAM problem using RFSs as it models the joint vehicle trajectory and map as a singular RFS and recursively propagates its first order moment. A first order approximation of the RFS state recursion is implemented which utilizes the PHD filter. Under Gaussian noise assumption, an extended Kalman Gaussian mixture implementation is used to implement the PHD-SLAM filter. The filter jointly estimates the vehicle pose, feature number in the map and their corresponding locations. Assuming a mildly non-linear Gaussian system, an extended Kalman Gaussian mixture implementation of the recursion is then tested for SLAM. In Chapter 6, a RAO-Blackwellised (RB) implementation of the PHD-SLAM filter is proposed based on the Gaussian mixture PHD filter for the map and a particle filter for the vehicle trajectory. Analysis are carried out, both in simulated environment through Monte Carlo trials and an outdoor SLAM experimental dataset based on a millimeter wave radar sensor. Results demonstrate the robustness of the proposed filter, particularly in the presence of large data association uncertainty and clutter, illustrating the merits of adopting an RFS approach to SLAM. Chapter 7 demonstrates that the RFS-FBRM and SLAM frameworks allow other approximations and implementations, besides those of the basic PHD filter, to be made. PHD-SLAM estimates the PHD of the map, encompassing the expected number of features, and the vehicle trajectory. A Multi-Bernoulli representation of an RFS is introduced, which allows each map feature to have its own probability of existence, yielding a valid density function which jointly captures its existence as well as spatial uncertainty. At the end three appendices dealing with some special mathematical questions are added.