ارائه‌ي روشي براي همجوشي حسگرهاي GNSS/IMU در وسايل نقليه‌ي خودران مبتني بر ESKF/RTS

Accurate, continuous, an‎d reliable localization is a fundamental requirement for autonomous vehicle navigation. An Inertial Measurement Unit (IMU) provides high-rate motion information, but integration of its measurement errors over time leads to accumulated position drift. In contrast, a Global Navigation Satellite System (GNSS) provides an absolute position reference, although its measurements are affected by noise an‎d a lower updat‎e rate. This thesis presents an‎d eva‎luates a two-dimensional GNSS/IMU localization framework based on an Error-State Kalman Filter (ESKF) combined with Rauch–Tung–Striebel (RTS) smoothing. During the forward stage, inertial measurements are used to continuously propagate vehicle position an‎d velocity, while periodic GNSS observations correct the accumulated estimation error. After forward filtering is completed, RTS smoothing uses information from the entire measurement sequence in a backward pass to refine the estimated trajectory. The final eva‎luation is conducted on KITTI sequences 00, 08, an‎d 09. Real OXTS acceleration an‎d angular-rate measurements are used as inertial inputs, while controlled experimental GNSS observations are generated by approximately 1 Hz subsampling of the reference trajectory an‎d addition of Gaussian position noise. The proposed ESKF+RTS framework is compared under a common protocol with IMU-only, GNSS-only, forward ESKF, UKF-M, an‎d an adapted KalmanNet implementation. Across five independent realizations of GNSS noise, the equal-sequence mean position RMSE is 0.774 m for ESKF+RTS an‎d 1.635 m for forward ESKF, corresponding to an approximately 52.6% relative reduction under the eva‎luation protocol used in this study. The proposed method also achieves an equal-sequence mean MAE of 0.688 m an‎d a CEP95 of 1.340 m. Trajectory an‎d error-distribution analyses are consistent with these quantitative results an‎d show that the backward smoothing stage improves the accuracy an‎d consistency of the filtered trajectory in the eva‎luated conditions. Because the study uses a simplified two-dimensional model, controlled GNSS observations, an‎d a limited set of KITTI sequences, the conclusions are restricted to the experimental protocol of this thesis an‎d are not intended as a general claim of superiority

محمدرضا احمدزاده , محمد داورپناه جزي

الهام محمودزاده , مينا اميري