FreeTrack: Device-Free Human Tracking With Deep Neural Networks and Particle Filtering

Abstract

Human tracking is of key importance to a variety of applications. Different from the previous approaches requiring the targets to carry electronic devices, this paper proposes a device-free tracking method based on WiFi channel state information (CSI) through deep neural networks (DNN) and particle filtering (PF). In the area covered with WiFi, human movements may cause observable variations of WiFi signals. By analyzing the CSI fingerprint patterns and modeling the dependency between CSI fingerprints and locations through DNN, the proposed method is able to localize the targets according to the measured CSI fingerprints through DNN regression. Localization with DNN is accurate for static targets, yet unable to form reasonable trajectories of moving targets. This paper proposes to apply PF on DNN localization results and form the walking trajectories of the targets. To avoid the trajectories to cross obstacles, map matching is integrated to constrain the unreasonable transfer of particles, which further improves the precision of trajectories. To combat with environmental variants, fine tuning on DNN is proposed to enhance the adaptivity of the method. Extensive evaluations in two representative scenarios with only one pair of WiFi transmitter and receiver achieve the high accuracy of around half a meter and precise trajectories as well as adaptation to environmental variants.