A precise urban vehicle navigation method based on 3D velocity Mamba constraint
TL;DR
The proposed GNSS/INS/Mamba-ODO framework learns a direct and efficient mapping from raw IMU outputs to 3D velocity through asymmetric sequence segmentation, multi-scale temporal modeling, and a state-space sequence learning backbone.
Abstract
In complex urban environments, vehicle-integrated navigation systems based on Global Navigation Satellite Systems (GNSS) aided Inertial Navigation Systems (INS) typically rely on nonholonomic constraints (NHC) and odometer velocity to improve navigation accuracy. However, the performance of such systems is often degraded due to inaccurate estimation of the installation angles between the Inertial Measurement Unit (IMU) and the vehicle frame. To address this challenge, this paper proposes a precise urban vehicle navigation method based on a 3D velocity constraint using the Mamba model. The proposed GNSS/INS/Mamba-ODO framework learns a direct and efficient mapping from raw IMU outputs to 3D velocity through asymmetric sequence segmentation, multi-scale temporal modeling, and a state-space sequence learning backbone. By replacing traditional velocity constraints, the method significantly enhances positioning robustness in GNSS-denied urban areas. Ground vehicle experiments demonstrate that the proposed approach achieves a mean Root Mean Square Error (RMSE) of 0.22 m/s in 3D velocity estimation. Furthermore, during a 120 s GNSS outage, the 3D positioning RMSE remains within 0.20 m, confirming the accuracy and robustness of the proposed method in real-world urban scenarios.