Integration of navigation sensors based on advanced Bayesian estimation methods

Global Navigation Satellite Systems (GNSSs) represent the leading technology for radio-navigation and their use is widespread in the framework of outdoor localization. For some applications, the strict requirements in terms of accuracy cannot be complied with a standalone GNSS solution and it is worthwhile coupling the former technology with an Inertial Navigation System (INS). A INS/GNSS integrated navigation unit leverages the complementary characteristics of the two sensors in order to enhance the accuracy and robustness of the solution to the localization problem. Among the available hybridization approaches, a Tightly-Coupled (TC) architecture is implemented, where a centralized Bayesian estimator exploits the low-rate GNSS noisy measurements to correct the INS high-rate estimates. The state-of-art INS/GNSS fusion routine is represented by the Extended Kalman Filter (EKF), where the integrated system models are handled in a linearized fashion. While the model for the dynamic evolution of the system inertial states can be regarded as linear to a large extent, the measurement model is inherently non-linear. As such, the primary effort of this Thesis is oriented to study more advanced Bayesian strategies better fitting with non-linearities. A first investigated methodology concerns the Unscented Kalman Filter (UKF) and its reliance on the Unscented Transform (UT) function to overcome model linearization. However, the accuracy performance of Kalman-based methods might be strongly penalized by their underlying assumption of Gaussian-distributed input observables, especially when navigating in signal-degraded environments. The latter issue steers the research towards the Particle Filter (PF) which, besides preserving non-linearities, offers enough flexibility to accommodate multiple density models for the statistical description of measurement noises. The joint synthesis between the UT-concept, at the basis of the UKF, and the Sequential Importance Sampling (SIS) idea, at the basis of any PF, gives rise to the Unscented Particle Filter (UPF), which sets as the innovative algorithmic proposal aimed at boosting up the integrated system capability. A crucial requirement to achieve accurate state estimation builds upon the appropriate representation of noise statistics. As regards measurement noise, it frequently behaves as a non-stationary and non-ergodic stochastic process, and static Bayesian formulations may fail in tracking its evolution. Therefore, this research also targets the application of adaptive integration schemes meant to promptly reflect the changes of the physical world on the filter statistical information. Starting from low-complexity innovation-based (Innovation-based Adaptive Estimation (IAE)) and residual-based (Residual-based Adaptive Estimation (RAE)) approaches, a more sophisticated paradigm exploiting redundant measurements is developed. Furthermore multipath, which identifies the harshest error source affecting GNSS range measurements, is addressed, and its mitigation is pursued through the elaboration of a self-contained pre-processing mechanism acting on raw observables. With the primary intent of achieving a performance assessment, the proposed Bayesian algorithms are tested on real INS and GNSS data. The ultimate proposal of this thesis is to verify whether the expected accuracy gain, theoretically entailed by the investigated filters, truly justifies the increase of complexity w.r.t. the original EKF approach.