Deep Learning TSPI from Wave Patterns for Indoor Localization

Abstract

Congested environments resulting in numerous reflections from one or more radio frequency (RF) sources exacerbate the accuracy of Time Space Positioning Information (TSPI). The St. Mary’s Unmanned Aerial Systems (UAS) Lab is a highly reflective building (almost entirely metal), so using GPS signals for indoor localization would be impractical. This brings up the possibility of using RF reflections to determine localization. This paper presents a new approach for positioning using deep learning algorithms to decipher the interaction of the RF reflections with physical obstacles. We utilized a calibrated infrared camera system with real-time TSP, vali- dating the synthetic signal strengths (SS) data with the experimental SS data, and implemented a Convolutional Neural Network (CNN)-based TSPI localization capability in an indoor, enclosed environment. We achieved an accuracy of 94% within a 1-2 meter radius, showing the possibil- ity of utilizing reflections of RF waves for indoor localization. Future work consists of refining and optimizing the algorithm, operating the model in different environments, and introducing the algorithm in real-world spaces.