Abstract |
Location and mobility management are major functions and essential features for seamless
and ubiquitous environments. Self-organizing sensor networks, health care monitoring, personal
tracking and context dependent information services are some of the potential applications.
Received signal strength (RSS) fingerprinting is a highly accurate location technique
that has the major advantage of exploiting already existing infrastructure to avoid additional
deployment costs. Fingerprint based localization systems adopt a calibration phase in order
to create signature maps that represent the physical space by capturing the variations of the
dynamic nature of indoor propagation. These maps, or fingerprints, are compared to the
RSS at the runtime phase in order to perform localization.
This thesis explores the notion of sparsity and reformulates the problem of user localization
as a sparse approximation problem. The proposed fingerprint-based localization
techniques adopt the Compressed Sensing (CS) framework, which provides a new paradigm
for recovering signals being sparse in some basis by means of a limited amount of randomly
received measurements. Specifically, exploiting the observation that the base stations receive
correlated signals from the mobile devices, we propose two CS-based algorithms: a
centralized and a decentralized one. According to the centralized Jointly CS scheme all
local runtime measurements received from the mobile device are sent to a central unit to
perform location estimation. On the contrary, the decentralized scheme builds upon gossip
consensus based approaches to distribute decision estimations in the network.
Although fingerprint CS-based systems achieve high accuracy, issues that concern both
the calibration and the location estimation phases can potentially limit the accuracy and
scalability of these systems. Concerning the automation of the calibration phase required by
fingerprint based systems, we propose a wireless localization and laser-scanner assisted Fingerprinting
system that provides autonomous signature map generation. During the location
estimation phase, the system mitigates the existing problems adopting a Bayesian formalism that incorporates a sparsity prior and dynamically determines the sufficient number of
runtime measurements required for accurate positioning.
Further challenges related to typical fingerprint-based schemes arise since it is implicitly
assumed that communication occurs over the same frequency channel during the training
and the runtime phases. When this assumption is violated, the mismatches between training
and runtime fingerprints can significantly deteriorate the localization performance. Additionally,
the exhaustive calibration procedure required during training limits the scalability
of this class of methods, especially in the case where no additional hardware is utilized.
To address these limitations, we propose a novel fingerprint collection technique without
the need of additional hardware that significantly reduces the calibration time by pseudorandom
channel sampling. The sub-sampled signature map is reconstructed as an instance
of the Matrix Completion problem.
Finally, we propose a reduced effort recalibration technique for fingerprint-based indoor
positioning systems. The proposed method exploits the dynamic characteristics of an indoor
environment and considers that a sub-set of measurements may explicitly depend on
past measurements. Particularly, we minimize the number of RSS fingerprints by performing
pseudo-random sub-sampling in space. The proposed framework exploits the spatial
correlation structure of the RSS fingerprints while considering prior information provided
from previously observed measurements, to reconstruct the signature map.
|