Skip to content

Latest commit

 

History

History
72 lines (69 loc) · 3.97 KB

associating_IMU_earth.md

File metadata and controls

72 lines (69 loc) · 3.97 KB

Associating Uncertainty to Extended Poses for on Lie Group IMU Preintegration with Rotating Earth

Author : Brossard

Year : 2021

Notes:

  • $SE_2(3)$ represents orientation, velocity and position. Also called "extended pose" $$ S E_2(3):=\left{\mathbf{T}=\left[\begin{array}{c|c|c} \mathbf{R} & \mathbf{v} & \mathbf{p} \ \hline \mathbf{0}_{2 \times 3} & \mathbf{I}_2 \end{array}\right] \in \mathbb{R}^{5 \times 5} \mid \begin{array}{c} \mathbf{R} \in S O(3) \ \mathbf{v}, \mathbf{p} \in \mathbb{R}^3 \end{array}\right} $$
  • addresses uncertainty propagation, pre integration on manif
  • defines extended pose integration as $$ \mathbf{T}_{i+1}=\boldsymbol{\Gamma}_i \Phi\left(\mathbf{T}_i\right) \mathbf{\Upsilon}i $$ with $$ \begin{aligned} \Phi_t(\mathbf{T}) & :=\left[\begin{array}{c|c} \mathbf{R} & \mathbf{v} \mid \mathbf{p}+t \mathbf{v} \ \hline \mathbf{0}{2 \times 3} & \mathbf{I}_2 \end{array}\right], \ \boldsymbol{\Gamma}_t & :=\left[\begin{array}{c|c} \mathbf{I}3 & \operatorname{tg} \mid \mathbf{g} t^2 / 2 \ \hline \mathbf{0}{2 \times 3} & \mathbf{I}_2 \end{array}\right], \ \mathbf{\Upsilon}_t&:=\left[\begin{array}{c|cc} \Delta \mathbf{R}_t & \Delta \mathbf{v}_t & \Delta \mathbf{p}t \ \hline \mathbf{0}{2 \times 3} & \mathbf{I}_2 \end{array}\right] \end{aligned} $$
  • Derives a noise model as: $$ \boldsymbol{\Upsilon}_i:=\hat{\boldsymbol{\Upsilon}}_i \exp \left(\boldsymbol{\eta}_i\right), \quad \boldsymbol{\eta}_i \sim \mathcal{N}\left(\mathbf{0}_9, \mathbf{Q}\right) $$
  • Compute a closed form formula for uncertainty propagation $$ \boldsymbol{\Sigma}{i+1}:=\mathrm{E}\left[\boldsymbol{\xi}{i+1} \boldsymbol{\xi}{i+1}^T\right] \stackrel{(42)}{=} \underbrace{\mathrm{Ad}{\hat{\boldsymbol{\Upsilon}}i^{-1}} \mathbf{F}}{:=\mathbf{A}_i} \boldsymbol{\Sigma}i\left(\operatorname{Ad}{\hat{\boldsymbol{\Upsilon}}i^{-1}} \mathbf{F}\right)^T + Q + S{4th} $$
  • The 4th order term is due to derivation with the BCH formula and leads to more accuracy wrt to monte carlo covariance
  • On simulations $SE_2(3)$ representation leads to a banana shaped uncertainty while $SO(3) \times \mathbb{R}^6$ leads to a gaussian
  • Proposes a IMU pre integration in $SE_2(3)$ that leads to the following measurement: $$ \mathbf{\Upsilon}{i j} =\left(\boldsymbol{\Gamma}{i j} \Phi_{i j}\left(\mathbf{T}_i\right)\right)^{-1} \mathbf{T}_j $$
  • Same pre integration measurement as in Foster et al but uncertainty is encoded in $SE_2(3)$
  • Compare approaches by deriving IMU measurements on a KITTI trajectory by differentiating groundtruth
  • $SE_2(3)$ is the most consistent approach according to the NEES criterion, while Forster et al approach degrades for long integration time
  • Add random bias peturbation on measurements from KITTI, and it is more accurate for pose and velocity correction but similar for rotation
  • IMU equations taking earth rotation into account: $$ \begin{aligned} \dot{\mathbf{R}} & =-\boldsymbol{\Omega}{\times} \mathbf{R}+\mathbf{R}\left(\boldsymbol{\omega}^m-\mathbf{b}^{\boldsymbol{\omega}}-\eta^{\boldsymbol{\omega}}\right){\times} \ \dot{\mathbf{v}} & =\mathbf{R}\left(\mathbf{a}^m-\mathbf{b}^{\mathbf{a}}-\eta^{\mathbf{a}}\right)+\mathbf{g}-2 \boldsymbol{\Omega}{\times} \mathbf{v}-\boldsymbol{\Omega}{\times}^2 \mathbf{p}, \ \dot{\mathbf{p}} & =\mathbf{v}. \end{aligned} $$
  • Obtain pre integration factors takin earth rotation into account that do not change the covariance formulas (as is corresponds only to non noisy additive terms)
  • Implement it in GTSAM and performs a 4H (~150km) drive to test long term LiDAR-Inertial navigation using a fixed lag smoother based on ISAM2
  • The factor graph is fed with: pre integration factors, translationnal factors (from LiDAR) and zero upward velocity factor
  • LiDAR uses scan matching algorithm (whose orientation uncertainty is way higher than the one of IMU integration)
  • A Huber loss is used for the translationnal factor to mitigate outliers
  • Taking into account earth rotation seems beneficial for sequences of length more than 7 minutes