Indoor Human Tracking Using Electrostatic Sensors and Recurrent Neural Networks

Abstract

Device-free localization and tracking (DFLT) of human targets in indoor environments is of great importance for a variety of applications. This article presents a novel DFLT method that utilizes an array of electrostatic sensors for near-field remote sensing of human motions and an end-to-end neural network for joint estimation and tracking of the human location. In order to determine the instantaneous location during free walking, the electrostatic signals are partitioned into frames, which are zero-aligned and normalized according to the sensing characteristics of the electrostatic sensor. Several prevalent neural networks are used to encode the signal frame into a feature vector. To achieve accurate localization and a smooth trajectory, a recurrent neural network (RNN), which is essentially a weighted average filter but free of problem-specific design and tuning of the weights, is used to learn the temporal correlations between neighboring frames. Experiments were conducted to collect training and testing data generated by three subjects under different environmental conditions. A smooth trajectory and a median localization error of around 0.38 m were achieved, validating the superiority of the RNN-based method over framewise estimation.