Label-Free Range-Based Indoor Tracking With Physics-Guided Deep State Space Model

Abstract

Accurate indoor tracking is essential to modern location-based services, fundamentally transforming the way we interact with indoor environments. Traditional state space model (SSM)-based tracking approaches often exhibit limitations in complex environments due to their reliance on fixed and overly simplified transition and observation functions, which restricts their capability to adequately capture intricate target dynamics and measurement uncertainties. To address these challenges, we propose a novel deep state space model (DSSM) that augments these fixed physics-based model functions with trainable neural networks (NNs). This innovative integration enables the DSSM to effectively learn previously unknown or inadequately modeled dynamics and uncertainties inherent in indoor tracking systems, while preserving critical physical constraints. Our proposed DSSM retains the structured representation and Bayesian inference of SSMs while significantly improving the capacity to characterize complex dynamics in both target motion and measurement errors. By leveraging this hybrid structure, the proposed DSSM facilitates maximum likelihood parameter learning directly from range measurements, eliminating the need for ground truth data. We further develop inference schemes of both online filtering and offline smoothing for the proposed DSSM. Extensive evaluations using real-world time of flight (ToF) measurements from two datasets across five diverse indoor scenarios demonstrate competitive or superior tracking performance compared to other state-of-the-art methods.