A New Paradigm for Dealing With Manifold Structures in Visual Inertial Odometry by Using Stable Embedding

Abstract

Visual inertial odometry (VIO) is a crucial area of research in robotics that enables navigation in GPS-denied environments by using an inertial measurement unit (IMU) and a camera to estimate sensor platform poses. Recent developments in VIO have adopted manifold or Lie group representations of state space, leading to increased accuracy and robustness. These VIO algorithms represent poses as elements of a manifold or Lie group and use extended Kalman filter (EKF) or unscented Kalman filter (UKF) on the tangent bundle or Lie algebra of the set. In this study, we propose a novel approach for dealing with the manifold structures of state space in VIO using the mathematical theory of stable embedding, which eliminates the need for consideration of tangent bundles or Lie algebras. We apply stable embedding to VIO and evaluate our algorithm on two real-world micro aerial vehicle (MAV) flight datasets. Our VIO outperforms the compared VIOs in terms of position trajectory error and computational complexity. An experiment video is available at: https://www.youtube.com/watch?v=F8zSzX_eeCo.