Read the paper and write down the requirements

[prashanthr05]

Read and understand the papers

• [x] Probabilistic Terrain Mapping for Mobile Robots With Uncertain Localization
• [x] An invariant-EKF VINS algorithm for improving consistency [2]
• [x] A Multi-State Constraint Kalman Filter for Vision-aided Inertial Navigation [1]

[prashanthr05]

The multi-state constraint Kalman filter [1] is basically a EKF based VINS algorithm that tries to avoid complexity blow up due to maintaining feature vectors within the state by the standard EKFs. It does so by proposing a measurement model that tracks same features over multiple camera frames and imposes a geometric constraint between these features. This way, the state is reduced to maintaining the IMU state of dimensions 15 in the tangent space and N camera poses of 6 dimensions each.

This algorithm is a very good deal for a real time navigation system that can be implemented with simple formulations, offering robust outlier rejection and optimal estimates under specific conditions.

However, this algorithm still suffers from consistency issues. Which is why a Right Invariant EKF based VINS is proposed [2] to avoid consistency issues by propagating uncertainties in a Lie group space. However, the complexity of RIEKF-VINS is very due to the aforementioned issue of maintaining too many features in the state vector. However, the RIEKF-VINS can be modified into a RI-MSCKF having similar theoretical guarantees.

[prashanthr05]

I will close this issue. Since I realized within a EKF framework, it quickly becomes cumbersome to add multiple factors and the fusion is performed in a loosely coupled manner. We will investigate the optimization based approaches for estimation, and maybe come back to this MSCKF approach for a baseline comparison.