Two Key-Frame State Marginalization for Computationally Efficient Visual Inertial Navigation

In this paper we perform a detailed evaluation of two key-frame state marginalization for visual inertial navigation filters to show that the method is significantly more computationally efficient than generic visual inertial odometry (VIO) methods while being sufficiently accurate for micro aerial vehicle (MAV) navigation. For this purpose, we use the EuRoC MAV dataset [1] for comparing the drift of MSCKF-Generic [2], MSCKF-Mono [3], MSCKF-Two way [4], and Two key-frame [5] VIO filters. The error state formulation of the two key-frame based and multi key frame based VIO is presented, then the drift, accuracy, and execution time of each filter is compared. The results indicate close to 90% faster execution of two key-frame based VIO algorithm on all datasets compared while having less than 3% drift in position for the total distance traversed.