Localization in Urban Canyon: Machine Learning based Localization Using LTE or LoRa Signal for 'GNSS-denied' Areas

Abstract

Primary goal of this paper is to improve accuracies of fingerprint-based localization systems by analyzing RSS data using machine learning techniques. As we know, generating and managing fingerprint database is quite costly and time-consuming. However, to the best of our knowledge, there are no means to estimate qualities of fingerprint databases. Considering that the localization accuracies highly depends on the quality of fingerprint database, providing such means has an huge impact on improving localization accuracies. In this paper, to estimate qualities of fingerprint database, we employ unsupervised learning to observe how reference positions with similar RSS vectors deploys in the interested area and show how the deployment varies as learning more and more RSS data. The result of unsupervised learning allows partitioning the fingerprint database and with the partitions, we are able to achieve more accurate localization. In addition, we employ Naive-Bayes classification for selecting partitions which is employed for filtering out reference positions that increase location ambiguities. Finally, we propose 'pattern matching'-based localization that is originated from regression analysis on RSS data in each partition.