Enhanced Localization Systems with Multipath Fingerprints and Machine Learning

Abstract

We propose a new method to enhance the performance of radio frequency localization in strong multipath and non-line-of-sight (NLOS) situations. The knowledge about the geometrical structure of multipath propagation environment is exploited by using a ray-tracing tool. We further apply the Random Forest (RF) algorithm embedded in a machine learning framework to extract a reference data-set of Time Differences of Arrival (TDOA) fingerprints in multipath outdoor scenarios. Site-specific fingerprints are processed with a multidimensional cross-correlation, called Volume Cross-Correlation function (VCC), to extract the multipath features from measurements. The performance and feasibility of our method was evaluated by simulations and measurements.