L220, 9:00 3.2.2012

This seminar will summarize some theoretical contributions in the field of simultaneous localization and mapping (SLAM), which enable long-term robotic navigation. Those are methods that I've been developing during the last three years.
First methodology is a filtering approach to pose SLAM, a variant of SLAM that estimates robot position rather than landmarks in the environment. It uses information theory to select only informative links and relevant robot poses and the result is a compact representation of the SLAM problem.
Looking at SLAM in terms of graphs has led to several novel and exciting developments in the past few years. Subgraph preconditioned conjugate gradient (SPCG) is obtained by re-interpreting the method of conjugate gradients in terms of the graphical model representation of the SLAM problem. It combines the advantages of direct and iterative methods for solving linear systems. This led to important computational speed-up in large scale SLAM and structure from motion (SFM) applications.
The Bayes tree is another concept we developed at Georgia tech. It provides an algorithmic foundation which enable a better understanding of existing graphical model inference algorithms and their connection to sparse matrix factorization methods. Based on that, we derived a fully incremental smoothing and mapping algorithm for efficient online robot localization and mapping of the environment called iSAM2.

Speakers