REPOSITORY > RESULTS

Abstract

Indoor localization is increasing in importance in the modern world, be it as a standalone
service or as a supporting service for other applications. Global navigation satellite
systems can provide a reliable outdoor localization service where a line-of-sight link between
a receiver and the minimum required number of navigation satellites is fulfilled. In indoor
environments, satellite signals are mostly too weak to be used for localization purposes.
Thus, indoor localization requires additional custom equipment which can be part of an
existing communication infrastructure or purposely deployed localization infrastructure
that supplies reference localization points.
As part of the doctoral dissertation we conducted measurements of radio signal propagation
characteristics in four representative indoor environments using ultra-wideband
radios. To this end, we used off-the-shelf ultra-wideband radios that enabled thorough
measurements suitable for the performance analysis of indoor localization algorithms. We
analysed the impact of non-line-of-sight signal propagation conditions on ranging error
based on time-of-flight measurements, which directly influences the accuracy of rangebased
localization algorithms.
The main focus was on the algorithms for detection of non-line-of-sight propagation
conditions and estimation of ranging errors based on raw channel impulse response. We
proposed a novel non-line-of-sight detection algorithm and a ranging error estimation algorithm
based on convolutional neural networks which do not require any preprocessing
of channel impulse response data. We built models for four representative indoor environments
and we showed with an analysis of a thorough cross-validation process that
algorithms based on convolutional neural networks outperform standard classification algorithms
based on the complex feature extraction process.
We evaluated the computational complexity of proposed non-line-of-sight classification
algorithms on different computational platforms. Despite high computational complexity
of proposed convolution neural networks, their efficient implementation enables using the
proposed models on small single-board-computers found in sensor nodes as well as on the
best data-center-grade hardware.
We also analysed the performance of different least-squares-based localization algorithms
and proposed a tracking algorithm based on an extended Kalman filter. By using
ranging error estimation for each measurement contribution and modifying their weights,
this algorithm substantially improved the tracking and localization performance in the selected
indoor environment compared to the algorithms that did not use any range quality
information.