Efficient Deep Learning-based Device-free Indoor Localization Using Passive Infrared Sensors

Abstract

Internet of Things (IoT) technology has continuously advanced over the past decade. As a result, device-free indoor localization functions have become a crucial part of application areas such as healthcare, safety, and energy management. Passive infrared (PIR) sensors detecting changes in temperature in an environment are one of the suitable options for human localization due to their lower cost, low energy consumption, electromagnetic tolerance, and enhanced private awareness. Although existing localization methods, including machine/deep learning, have been proposed to detect multiple persons based on signal phase and amplitude, they still face challenges regarding signal quality, ambiguity, and interference caused by the complex, interleaving movements of multiple persons. This paper proposes a novel deep learning method for multi-person localization using channel separation and template-matching techniques. The approach is based on a deep CNN-LSTM architecture with ensemble models using a mean bagging technique for achieving higher localization accuracy. Our results show that the proposed method can estimate the locations of two participants simultaneously with a mean distance error of 0.55 m, and 80% of the distance errors are within 0.8 m.