Quantifying VIO Uncertainty

We compute the uncertainty of XIVO, a monocular visual-inertial odometry system based on the Extended Kalman Filter, in the presence of Gaussian noise, drift, and attribution errors in the feature tracks in addition to Gaussian noise and drift in the IMU. Uncertainty is computed using Monte-Carlo simulations of a sufficiently exciting trajectory in the midst of a point cloud that bypass the typical image processing and feature tracking steps. We find that attribution errors have the largest detrimental effect on performance. Even with just small amounts of Gaussian noise and/or drift, however, the probability that XIVO's performance resembles the mean performance when noise and/or drift is artificially high is greater than 1 in 100.