Tightly Coupled RTK/IMU Integration With a Multipath-Aware Stochastic Model-Based Ambiguity Resolution for Positioning in Urban Areas

Abstract

Tightly coupled integration of single-frequency global navigation satellite system (GNSS) real-time kinematic (RTK) observations and inertial measurement unit (IMU) data provides a practical solution for high-precision urban positioning. However, severe nonline-of-sight (NLOS) and multipath effects compromise urban positioning performance. In such environments, GNSS observation errors become non-Gaussian and heteroscedastic, degrading single-epoch ambiguity resolution (AR) robustness and navigation stability. To address this challenge, we propose a multipath-aware stochastic model-based AR method. The key idea is to incorporate offline-learned, location-aware mean-square multipath uncertainty priors into online stochastic modeling. Offline, a hexagonal-grid regressor learns satellite- and location-conditioned priors from receiver-accessible signal-quality, geometric, and contextual features, thereby retaining both persistent multipath/NLOS bias energy and local centered dispersion. Online, each long-term prior is adaptively combined with a short-term centered residual-dispersion estimate to form an effective pseudorange covariance-inflation term. The fused pseudorange uncertainty term and its scaled counterpart are combined with the code- and phase-tracking-noise variances, respectively, to yield a unified heteroscedastic covariance model. Using this model, float ambiguities and their covariance are estimated, after which ambiguity dilution of precision (ADOP)-guided partial AR (PAR) selects and fixes a reliable ambiguity subset. The resulting ambiguity solution is used in the RTK/IMU state update. Field tests demonstrate that the proposed method achieves an ambiguity-fixing rate of 93.17 %, outperforming full AR (FAR) (61.58 %), elevation-based PAR (73.42 %), and ADOP-based PAR (62.00 %). The 3-D positioning root-mean-square error (RMSE) is reduced to 0.24 m, with reductions of 77.57 %, 62.50 %, and 69.23 % relative to FAR, elevation-based PAR, and ADOP-based PAR, respectively.