Vehicle Inertial Localization With Adaptive Noise Estimation and Pseudo-Measurement Constraints in GNSS Denied Environment

Accurate and robust localization is essential for autonomous driving in complex, dynamic environments and across various motion scenarios. Due to signal blockage, Global Navigation Satellite Systems (GNSS) positioning is not reliable in urban complex environments, and inertial navigation systems (INS) is widely integrated using the Kalman filtering for continuous and robust vehicle localization. On the one hand, INS suffers from accumulated errors and it poses a challenge for precise vehicle localization in complex environments, where GNSS has degraded performance. On the other hand, the Kalman filtering takes certain empirical assumptions regarding the probability models of process noise and measurements, which may be not matched with the real-world conditions. To address these issues, we propose a data and model jointly driven neural-Kalman solution for vehicle inertial localization to enhance positioning accuracy and robustness in various motion states under GNSS-denied environments. The proposed solution exploits the data-driven neural networks to real-time predict the covariance of process noise and three-dimensional (3D) vehicle velocity. Within the neural networks, the proposed loss function dynamically adjusts the error penalty weight and establish a connection between the uncertainty and estimation to improve the localization robustness. Finally, the predicted covariance of process noise and 3D velocity are incorporated with the adaptive Kalman filtering to improve the vehicle localization accuracy. We evaluate the proposed method using two datasets, and the results indicate that the average RMSE of 3D velocity regression is reduced by 56.33% and 50.08%, respectively, compared to two deep learning methods. In the positioning accuracy evaluation experiment, the proposed neural Kalman filter method reduces the average ATE and RTE by 21.10% and 23.96% compared with the state-of-the-art baseline methods in two datasets.