Statistical Data Fusion for Hybrid Localization of Mobile Terminals
In recent years, there is an increased interest in wireless location systems offering reliable mobile terminal (MT) location estimates. This is mainly due to upcoming and already available Location Based Services, such as intelligent transport systems, yellow page services, location sensitive billing and other promising services that rely on accurate MT location estimates. So far, a multitude of wireless location systems have been proposed that offer MT location estimates. The most promising solutions are based on the Global Navigation Satellite System (GNSS) and the cellular radio network, since both systems utilize an already existing infrastructure. Conventionally, these systems provide MT location estimates independently from each other. However, there exist scenarios where the signals that are exchanged between the satellites and the MT are blocked, e.g., in urban environments where tall buildings surround the MT or in indoor environments. In these scenarios, the number of measurements available from GNSS is often insufficient to determine the MT location. The signals that are exchanged between the base stations (BSs) of the cellular radio network and the MT are available in these scenarios, but generally they cannot offer the same accuracy as the signals from GNSS. In urban and indoor scenarios, these signals are often reflected at obstacles such as buildings or trees, so that a direct, line-of-sight (LOS) path between MT and the BS does not exist. In this case, the signal of the MT arrives via an indirect path at the BS, which is known as non-line-of-sight (NLOS) propagation. The errors due to NLOS propagation generally result in a decreased localization performance, and should be therefore taken into account in the MT localization algorithms.
This project deals with the problem of estimating the MT location using pseudorange (PR) measurements from the Global Positioning System (GPS) and round trip time (RTT) and received signal strength (RSS) measurements from the Global System for Mobile communications (GSM), which is termed hybrid localization. The measurements, which are available from off-the-shelf mobile phones and conventional GPS receivers, are efficiently combined by using statistical data fusion, so that it is possible to obtain MT location estimates even if the number of measurements available from GPS is insufficient to determine the MT location. The corresponding hybrid localization algorithms are designed such that good performance can be achieved in situations when the measurements are affected by either LOS propagation conditions or propagation conditions that switch between LOS and NLOS. It is investigated how the existing offset between the satellite clocks and the MT clock can be mitigated and the localization accuracy can be improved by using GNSS reference time (GRT) measurements.
In order to analyze the hybrid localization algorithms, a mathematical framework is introduced that describes the hybrid localization scenario. Statistical models for the MT movement and MT clock, as well as models for the measurements assuming LOS and NLOS propagation conditions are introduced. In this project, the following three types of hybrid localization algorithms are introduced:
• Non-recursive hybrid localization algorithms, that do not take into account existing temporal dependencies between time consecutive MT locations and measurements.
• Recursive hybrid localization algorithms, that take into account the information of MT estimates and measurements from previous time steps.
• Recursive hybrid localization algorithms with adaptive LOS/NLOS detection, that take into account the information of MT estimates and measurements from previous time steps, and that estimate the current propagation conditions.
The non-recursive hybrid localization algorithms are based on the maximum likelihood (ML) principle. The ML estimators for LOS propagation conditions and for propagation conditions that switch between LOS and NLOS are newly derived, and ML estimates are numerically obtained using suboptimal algorithms. In order to assess the theoretical best achievable performance of non-recursive estimators, the Cramer-Rao lower bound (CRLB) for hybrid localization is evaluated. Simulation and field trial results have shown that additionally taking into account PR measurements from GPS and GRT from GSM in the algorithms can significantly improve the localization accuracy compared to algorithms that only take into account RTT and RSS measurements from GSM.
The recursive hybrid localization algorithms developed in this project are Kalman filter (KF)-based estimators and particle filter (PF)-based estimators. Different estimators for LOS propagation conditions and for propagation conditions that switch between LOS and NLOS are newly proposed. The PF-based estimators additionally take into account road information to further improve the localization accuracy. The theoretical best achievable performance of recursive estimators is found by evaluating the posterior CRLB (PCRLB). It is shown that additionally taking into account road information into the estimators can significantly improve the localization accuracy. It is further demonstrated that recursive hybrid localization algorithms outperform non-recursive hybrid localization algorithms.
The recursive hybrid localization algorithms with adaptive LOS/NLOS detection that are proposed in this work are based on the interacting multiple model (IMM) estimator that is combined with extended KFs (EKFs) and two multiple model PF-based estimators. The multiple model PF-based estimators additionally take into account road information to further improve the localization accuracy. A novel method is proposed to determine the PCRLB for recursive estimators with adaptive LOS/NLOS detection.
It is shown that multiple model PF-based estimators with road constraints generally outperform the IMM-EKF. It is further demonstrated that the IMM-EKF achieves the best trade-off between performance and computational complexity, as long as road constraints are not considered in the multiple model PF-based estimators.