NeurINS: Neural Inertial Navigation System for Consistent and Drift-Free State Estimation

Abstract

The data-driven inertial navigation system (INS) utilizing a low-cost inertial measurement unit (IMU) is remarkably autonomous and impervious to external signals, demonstrating considerable potential in environments, such as indoors and underground. However, existing INS studies still face two challenges. First, they cannot guarantee global consistency, that is, the spatial correspondence between a personal IMU and the global world. Second, they are unable to support robust long-term state estimation since their open-loop integration may incur pose drift over time. To address the two challenges, we propose a neural network-aided INS (NeurINS) that leverages solely an IMU for consistent and drift-free state estimation. For the first challenge, we propose a consistency regression network (CR-Net) to extract the global consistency implicit in IMU measurements and predict consistency vectors. These vectors serve as global constraints for the subsequent joint state estimation, enabling the recovery of spatial correspondence between the IMU and world frames. For the second challenge, a novel neural-inertial fusion method is proposed to jointly estimate the six-degree-of-freedom (6-DoF) pose, velocity, and IMU biases, where a state-space description is modeled, and a resilient filter is utilized as the estimator. Considering the accuracy and smoothness of localization over short durations, we propose a velocity regression network (VR-Net) to provide relative motion constraints. Experimental evaluations on the IDOL dataset demonstrate that NeurINS achieves consistent and drift-free state estimation using only a consumer-grade IMU. Specifically, the root mean squared errors (RMSEs) of NeurINS are 0.95 m for absolute trajectory error (ATE) and 3.44° for absolute orientation error (AOE), representing reductions of 71% and 54%, respectively, compared with other methods. Trials on our self-collected dataset further prove the superior performance of NeurINS.