Square-Root Extended Information Filter for Visual-Inertial Odometry for Planetary Landing

A novel sequential information filter formulation for computationally efficient visual-inertial odometry and mapping is developed in this work and applied to a realistic moon landing scenario. Careful construction of the square-root information matrix, in contrast to the full information or covariance matrix, provides easy and exact mean and covariance recovery throughout operation. Compared to an equivalent extended Kalman filter implementation, which provides identical results, the proposed filter does not require explicit marginalization of past landmark states to maintain constant-time complexity. Whereas measurements to opportunistic visual features only provide relative state information, resulting in drift over time unless a priori mapped landmarks are identified and tracked, the tight coupling of the inertial measurement unit provides some inertial state information. The results are presented in a terrain-relative navigation simulation for both a purely orbital case (with no active propulsion) and a landing case with a constant thrust.

Automated summary

This work develops a novel sequential information filter formulation for computationally efficient visual-inertial odometry and mapping. By carefully constructing the square-root information matrix, the mean and covariance can be easily and accurately recovered throughout operation. Results show that this filter does not require explicit marginalization of past landmark states to maintain constant-time complexity.