Improving the Convergence of a Data-Driven Factor Graph Model for UWB-Based Positioning in Challenging Indoor Scenarios

Indoor scenarios can pose challenges for localization. Effects such as multipath and non-line-of-sight propagation between devices can cause errors in positioning systems that utilize radio technologies, even in ultra-wideband (UWB) technology-based systems. Additionally, the geometrical location of the anchor nodes can also impact the positioning error with lateration. In the case of iterative algorithms such as those based on Machine Learning, these effects can affect algorithm convergence. This work presents a solution to enhance the convergence stage of a Data-driven Factor Graph (FG) Model for UWB-based positioning in these scenarios. The FG is represented as a tree, a connected and acyclic graph that ensures convergence of the Gaussian Belief Propagation (BP) algorithm. The new solution utilizes Gaussian Mixture Model fitting to estimate the covariance of distances between the target and anchor nodes. The presented solution is compared with the classical solution that estimates the covariance with a Gaussian fitting. The comparison is conducted using real-life data collected from sensor network nodes compliant with IEEE 802.15.4, featuring a physical layer based on UWB technology. The results show that in challenging scenarios, the presented solution improves the rate of the convergence phase of BP, although at the expense of increased complexity in the FG-based algorithm.