Multipath assisted positioning using machine learning

Abstract

The multipath propagation of the radio signal was considered a problem for positioning systems that had to be eliminated. However, a groundbreaking new approach called multipath assisted positioning caused a paradigm shift, where multipath propagation improves the positioning performance. Moreover, the multipath assisted positioning algorithm called Channel-SLAM shows the possibility of using a single physical transmitter in a multipath environment for positioning. In this thesis, I open a discussion on some problems that have vital importance for multipath assisted positioning algorithms with a focus on pedestrian positioning. Using the idea of multipath assisted positioning, I present a single frequency network positioning algorithm. I evaluated the single frequency network-based positioning algorithm for positioning in a real scenario using a terrestrial digital video broadcasting transmission. I propose a novel pedestrian transition model utilizing the inertial measurements from a handheld inertial measurement unit. The proposed pedestrian transition model improves the precision and reliability of the Channel-SLAM. Comparing the proposed transition model with the Rician transition model previously used in Channel-SLAM quantifies the performance improvement. This thesis proposes a joint data association technique that overcomes the strong dependence on the radio channel estimation algorithm used in Channel-SLAM. The joint data association allows reusing the previously observed virtual transmitters after an outage of multipath component tracking. The evaluation based on the walking pedestrian scenario shows that the joint data association algorithm provides superior positioning precision. The virtual transmitter position estimation yields a significant computational load in Channel-SLAM. I propose a method that represents the virtual transmitter by a Gaussian mixture model and learns its parameters. The evaluation shows that the proposed method outperforms the previous approach while decreasing the computational load. Also, the current methods for radio channel estimation yield a considerable computational load that prohibits a real-time deployment. The thesis investigates the possibility of using artificial neural networks trained to estimate the number of multipath components and corresponding delays in a noisy measurement of a channel impulse response. The artificial neural network-based delay estimator provides a superresolution performance and faster runtime than the classical approaches. The precision of the trained artificial neural network architecture is evaluated and compared to the Cramer-Rao lower bound theoretical limit and classical channel estimation algorithms.