Learning-Aided Inertial Odometry With Nonlinear State Estimator on Manifold

Abstract

Relying only on inertial measurement units (IMUs) for robust state estimation is critical to vehicle safety when imaging sensors abruptly fail. In this paper, we propose to consider learning-based method as a complement to the kinematic model, and obtain ego-motion based on the nonlinear filter pipeline. To be specific, we first model the state of the IMU on the manifold such that the beliefs of prior model are propagated correctly. Then, we construct an uncertainty-aware network to simultaneously learn the integral terms in the kinematic equations, and recursively compute the rigid body position and velocity as pseudo-measurements. We additionally use a nonlinear estimator to properly fuse the model with the learned observations, whose prior information is endowed by model on the manifold, while the updating correction signals are provided by the network on pattern learning, and finally, the split covariance intersection (SCI) is utilized to reasonably handle the unknown correlated information in both. The performance of method is evaluated in terms of accuracy, robustness, extensibility and server aspects using both simulated and real-world dataset. Experimental results demonstrate a promising performance of the proposed method to traditional or learning-based ones.