Human Trajectory Tracking with Wi-Fi Integrating Machine Learning and Radio Feature Extraction

Abstract

This paper focuses on indoor human trajectory tracking with Wi-Fi sensing. There have been two conventional approaches to realizing passive tracking by exploiting the channel state information (CSI) obtained for a single Wi-Fi link: the machine learning (ML) and signal processing (SP) approaches. The ML approach using CSI as its input enables accurate estimation for a specific training environment; however, it suffers from large generalization error, i.e., it gives large estimation error when applied to a new environment. On the other hand, in the SP approach, radio features, such as angle of arrival (AoA), time of flight (ToF), and Doppler frequency shift (DFS), are first obtained by processing CSIs of multipath signals propagating through a target human. Then, the human trajectory is estimated by using a geometric model, which makes the estimation less-dependent on a specific environment. However, it is hard to stably extract the radio features from the weak signals reflecting on the human body. Considering these advantages and disadvantages, this paper proposes an integrated approach that uses the CSI and radio features obtained by the SP approach as input to the Transformer based ML model. Our experimental results show that the proposed approach employing the Transformer model maintains a high estimation accuracy even when applied to different environments from the training.