A Mutual Information–Based Infrastructureless Radio Frequency Positioning System via Deep Learning

Abstract

We proposed a novel, scalable, spatial positioning solution based on existing radio frequency (RF) signals and available infrastructure by harvesting information from the signals, distilling the readings to get the mutual information gain of each perceived RF characteristic at the corresponding location of interest, and processing such information with a deep neural network for localization. The solution leveraged all available characteristics of perceived RF signal sources instead of filtering them out. We built and tested a scalable positioning prototype on the Google Kubernetes containerization framework accordingly. The preliminary results achieved at least 1.8 m (6 ft) of accuracy within 90% cumulative distribution and 1.08 m (3.3 ft) mean accuracy by solely using 802.11ac received signal strength fingerprints. The prototype performed approximately three times as good as the referenced baseline of weighted k-nearest neighbors.