Towards reliability and scalability in feature based simultaneous localization and mapping

Publication Type:
Issue Date:
Full metadata record
Simultaneous Localization and Mapping (SLAM) has always been an attractive topic in the vibrant field of robotics. Feature based representations of the problem can be seen as one of the most common definitions. In recent years, many SLAM researchers have realized some limitations of filtering based methods and started to focus more on optimization based SLAM techniques. However, this raises several questions surrounding convergence reliability and, similar to filtering, algorithm scalability. In SLAM, sensor noise and non-linearity often causes the problem to become difficult. Converging towards the global minimum in a non-linear least squares formulation is by no means easy. Typically, one would need to start from a good initial estimate, preferably already inside the basin of attraction of the global minimum. In this thesis, we introduce a technique called Iterative Re-Weighted Least Squares bootstrapping to achieve a good initial estimate even when the noise is exceptionally large. As a robot continues to traverse through its environment the complexity of SLAM tends to scale badly with the cumulative nature of graph nodes and edges. To solve large SLAM problems within a reasonable time scale one must also take into consideration elements of accuracy and consistency. In this thesis, we propose two alternative algorithms to handle complexity, Sparse Map Joining and Pose Graph Representation. Both of which contain unique advantages for handling the diverse scenarios within SLAM. A series of quantitative analyses are performed on a number of challenging datasets, both real and simulated. In addition to this we perform a comprehensive case study on a specific type of feature based SLAM problem, RGB-D SLAM. This demonstrates how our technique is capable of avoiding inaccuracies and failure scenarios that is otherwise common in other RGB-D SLAM algorithms.
Please use this identifier to cite or link to this item: