BooM-Vio: Bootstrapped Monocular Visual-Inertial Odometry with Absolute Trajectory Estimation through Unsupervised Deep Learning

Machine learning has emerged as an extraordinary tool for solving many computer vision tasks by extracting and correlating meaningful features from high dimensional inputs in ways that often exceed the best human-derived modeling efforts. However, the area of vision-aided localization remains diverse with many traditional, model-based approaches (i.e. filtering- or nonlinear least- squares- based) often outperforming deep, model-free approaches. In this work, we present Bootstrapped Monocular VIO (BooM), a scaled monocular visual-inertial odometry (VIO) solution that leverages the complex data association ability of model-free approaches with the ability to exploit known geometric dynamics with model-based approaches. Our end-to-end, unsupervised deep neural network simultaneously learns to perform visual-inertial odometry and estimate scene depth while scale is enforced through a loss signal computed from position change magnitude estimates from traditional methods. We evaluate our network against a state-of-the-art (SoA) approach on the KITTI driving dataset as well as a micro aerial vehicle (MAV) dataset that we collected in the AirSim simulation environment. We further demonstrate the benefits of our combined approach through robustness tests on degraded trajectories.