Author: Nerurkar
Year: 2014
- marginalize non KF information and maintains sparsity
- discarded =! marginalized
- both visual and inertial measurements
- Gauss Newton iterative has a complexity of
$\mathcal{O}([K+N]^3)$ with$K$ the number of lmk and$N$ the number of robot poses - exteroceptive measurements between non keypose and keypose destroy sparsity
- to maintain sparsity it ignores the data association between landmarks in keypose with non keypose
- approximated marginalization will only fill information with pose constraint between the consecutive keyposes
- computationnal complexity analysis of marginalization ->
$\mathcal{O}(M_r^3)$ with$M_r$ the number of non key pose
Experiments:
- VIO with SIFT features matching
- CKLAM takes 4% of the full BA time with 5% more error