Information Sparsification in Visual-Inertial Odometry

Author: Michael Kaess

Year: 2018

Notes:

"To bound computational complexity, fixed-lag smoothers typically marginalize out variables, but consequently introduce a densely connected linear prior which significantly deteriorates accuracy and efficiency"
sparsification enables to maintain information sparsity, structural similarity and non linearity
fixed-lag smoothing framework: sliding window
drawbacks of sliding window:
- Marginalizes out variable=> fixing linearization points
- Will not converge to optimal solution as marginalized variables are no longer optimizable
- prior densely connected => high computationnal cost
perform VIO with marginalization & sparsification in RT
VINS MONO stratefy: discard measurements (ie landmarks) for sparsity & marginalize additionnal variables => loose the capabilities of re-estimating the position of the landmarks

Problem:

window of states $\mathcal{X}_w = { \mathcal{K}_w, \mathcal{F}_w, \mathcal{L}_w }$
relative IMU meas with "pre integration factors", each landmark is a 3D point in world frame: "stereo projection factors"

System:

marginalization is done with Schur Complement on linearized information matrix of the markov blanket
mid frame marginalization: if a KF is not voted: all projection factors are discarded, but only include inertial constraint
Figure 3 is excellent!
KF marginalization: When a KF is voted, the last one is marginalized with all the landmarks only associated to it, but the other one are kept (unlike VINS MONO and co.)
topology: independent unary prior factors between the frames and relative pose factors (interesting) with all landmarks
Covariances recovery is computed in closed form as measurements models always provide full rank invertibles jacobian

Results:

odometry drift measurement of 0.3m/170m on hardware demonstration
most of the time for marginalization is spend on closed form information matrix computation (it seems that it is highly parallelizable?)
next step: explore new factor graph topologies