MUID: Multiperson Gait Identification With Commodity Wi-Fi

With the popularization of smart homes and the Internet of Things, human gait identification using Wi-Fi channel state information (CSI) has attracted much attention. However, it remains a challenge to accurately recognize multiperson identities when they walk simultaneously in the sensing area. Existing researches separate the walking signals of each person from the multipath superposed CSI and then constructs the Doppler spectrum of separated signals to extract human gait features for multiperson identification (MUID). However, these systems suffer from the unsatisfactory accuracy since the extracted gait features are inaccurate and inadequate from the original Doppler spectrum, which is position-dependent and has low temporal resolution. To this end, we propose MUID, a Wi-Fi-based system capable of accurately recognizing multiperson identities using commercial devices. In MUID, we first leverage an existing model proposed by our previous work to separate each person’s walking signals and propose a spectrum calibration method to remove the position dependency of the Doppler spectrum of separated signals. Next, we novelly extract gait temporal AutoCorrelation (TAC) information and use it to compensate for the low temporal resolution of the Doppler spectrum. Finally, we design an attention-based two-branch feature fusion network, which can effectively combine the gait features extracted from the calibrated Doppler spectrum and TAC for accurate multiperson identification. Extensive experiments are conducted in three different environments and MUID can, respectively, achieve 93.56% and 90.59% average accuracy for 2-person and 3-person identification, superior to the state-of-the-art methods.