Dynamic Real-Time Anchor Selection for Accurate UWB Indoor Positioning-Based Deep Neural Networks

In wireless localization systems, enhancing location estimation performance is critical, particularly in challenging environments, such as military urban operations and emergency response scenarios. Ultra-wideband (UWB) positioning systems using two-way-ranging (TWR) schemes avoid synchronization issues, but face challenges related to scalability, anchor selection, and poor channel characteristics. Existing methods often rely on exhaustive geometric calculations, leading to inefficiencies in dynamic and time-critical scenarios. This paper proposes a lightweight sequential branching deep network for dynamic indoor positioning (LSB-DIP Net) is proposed to address these challenges. By integrating multi-scale feature extraction, sequential learning, and advanced activation functions with the conventional linearized least squares method, the LSB-DIP Net enables robust, accurate, and dynamic UWB positioning in real-time. The model effectively mitigates non-line-of-sight (NLOS) ranging errors, evaluates anchor channel quality online, and selects optimal anchor combinations, ensuring scalability and adaptability for diverse deployment scenarios. The proposed approach demonstrates exceptional performance in dynamic setups, achieving low mean squared error (MSE) of 0.0051m2, high accuracy in identifying anchor channels of 99.44%, with a maximum positional error of less than 0.17 m in harsh environments. Validated across public datasets, the system ensures generalizability and outperforms state-of-the-art counterparts in the market, making it a reliable tool for real-time applications in communication and navigation systems.