Author: Kaess
Year: 2008
- smoothing information matrix is naturally sparse, but in filtering approaches it becomes dense because of marginalization
- periodic variable reordering for loops (?)
- fast algorithm for data association
- solution based on QR factorization
- iterative optimization avoids the problem arising from wrong linearization point
- QR factorization with givens rotation -> set to zeros all entries below diagonal, constructs the Q matrix
- number of givens rotation is independant of the length of the trajectory
- With loops, information matrix remains sparse but R becomes dense -> variable reordering with COLAMD
- Uses ML data association (better than NN as it takes uncertainties into account, but slower)
- The full covariance matrix is never calculated, marginals are obtained performing backsubstitution on R matrix
Commentaires:
- Equations trés claires