| Ananth Ranganathan College of Computing |
|||||||
|
Real-time Smoothing and Mapping (SAM)
However, the originally proposed SAM algorithm had the disadvantage of being a batch algorithm that had to recompute the solution for all the poses and features at each step. This project has the aim of producing incremental SAM algorithms that run in constant time except in rare cases. To achieve this, we have been following two parallel paths. One technique leverages the linear algebra literature relating to the update of matrix factorizations, while the other view proceeds from a graphical model view of the SLAM problem where we try to perform incremental inference. The result has been two algorithms - incremental SAM, based on Givens rotations to update the QR factorization of a matrix, and Loopy SAM, based on loopy belief propagation performed in an incremental manner so that the message updates can be done in constant time per step. ![]() The SAM problem as an MRF
![]() Loopy Belief Propagation on the MRF solves the SAM problem. Only the nodes near the robot (in red) are updated here. ![]() When the factorization of matrix (here the triangular matrix R of a QR factorization) is updated incrementally upon the addition of a few rows to the original matrix, only a small portion of the entries get changed (shown in red). Publications
|