Forster 2017 IMU Preintegration
Authors: Forster et al
Abstract:
- First contribution: preintegration theory (building up on Lupton’s work)
- what’s different from Lupton’s:
- addresses manifold structure of the rotation group, analytic derivation of all Jacobians
- Lupton’s work uses Euler angles
- Using Euler angles and techniques of Euclidian spaces for state propagation/covariance estimation is not properly invariant under rigid transformations
- uncertainty propagation, a-posteriori bias correction
- same as Lupton: integration performed in local frame, eliminating need for reintegrating when linearisation point changes
- what’s different from Lupton’s:
- Second contribution: integration of the preintegrated IMU model into a visual-inertial pipeline
- The system presented uses incremental smoothing for fast computation of the optimal MAP estimate
- Uses structureless model (3D landmarks are not part of the variables to be estimated)
- for visual measurements –> allows eliminating large numbers of variables
Motivation:
- optimisation-based approaches are more accurate than filtering-based ones, but come at the cost of high computational complexity
- preintegration theory allows reduction of the computational complexity by accurately summarising multiple inertial measurements into a single relative motion constraint
IMU preintegration over frames
Introduction Why use the visual-inertial sensor combination? [Handling the computational complexity of optimisation-based SLAM](handling the-computational-complexity-of-optimisation-based-slam.md)
Preliminaries [SO(3) 3D rotation group](so(3) 3d rotation group.md), lie group,-lie-algebra Exponential map
SE(3) Special Euclidian Group Gauss-Newton Method on Manifold
MAP VI state estimation [System in a VIN problem with IMU preintegration](system in a-vin-problem-with-imu-preintegration.md) MAP estimation
IMU model and motion integration IMU kinematic model using Euler integration IMU preintegration on manifold
Preintegration math
- To do…