Exactly sparse information filters for simultaneous localization and mapping

摘要

NO FULL TEXT AVAILABLE. Access is restricted indefinitely.  –  – This thesis is concerned with computationally efficient solutions to the simultaneous localization and mapping (SLAM) problem. The setting for the SLAM problem is that of a robot with a known kinematic model, equipped with on-board sensors, moving through an environment consisting of a population of features. The objective of the SLAM problem is to estimate the position and orientation of the robot together with the locations of all the features. Extended Kalman Filter (EKF) based SLAM solutions widely discussed in the literature require the maintenance of a large and dense covariance matrix. Recently, Extended Information Filter (ElF) based SLAM solutions have attracted significant attention due to the recognition that the associated information matrix can be made sparse. However, existing algorithms for ElF-based SLAM have a number of disadvantages, such as estimator inconsistency, a long state vector or information loss, as a consequence of the strategies used for achieving the sparseness of the information matrix. Furthermore, some important practical issues such as the efficient recovery of the state estimate and the associated covariance matrix need further work. The contributions of this thesis include three new exactly sparse information filters for SLAM: one is achieved by decoupling the localization and mapping processes in SLAM; the other two are aimed at SLAM in large environments through joining many small scale local maps. In the first algorithm, D-SLAM, it is shown that SLAM can be performed in a decoupled manner in which the localization and mapping are two separate yet concurrent processes. This formulation of SLAM results in a new and natural way to achieve the sparse information matrix without any approximation. In the second algorithm, the relative information present in each local map is first extracted, and then used to build a global map based on the D-SLAM framework. Both these algorithms, while computationally efficient, incur some information loss. The third algorithm that modifies the global map state vector by incorporating robot start and end poses of each local map, completely avoids the information loss while maintaining the sparseness of the information matrix and associated computational advantages. Two efficient methods for recovering the state estimate and the associated covariance matrix from the output of the ElF are also proposed. These methods exploit the gradual evolution of the SLAM information matrix, and allow the ElF-based SLAM algorithms proposed in this thesis to be implemented at a computational cost that is linearly proportional to the size of the map.