Pi-Loc: a Pareto indoor localization using deep learning in 5GB

Abstract

Abstract Ultra-accurate indoor localization is essential for critical applications such as emergency drone tracking, healthcare operations, and autonomous systems that rely on 5G New Radio (NR) and beyond networks. Sub-centimeter positioning accuracy is increasingly necessary for safe and effective operation in complex indoor environments. However, existing localization methods, including global positioning systems (GPS), face fundamental limitations when dealing with complex blockage conditions, signal heterogeneity, and the coexistence of line-of-sight (LoS) and non-line-of-sight (NLoS) propagation conditions, particularly in single-cell positioning scenarios. In response to these challenges, this work introduces Pi-Loc, a novel Pareto indoor localization framework that explicitly addresses the trade-off between LoS and NLoS signal conditions through multi-objective optimization. Pi-Loc operates in two phases: first, Pareto optimization is applied to min-max normalized 5G NR channel state information (CSI) features, minimizing the competing LoS and NLoS localization errors simultaneously; second, the Pi-Loc convolutional network is trained on the Pareto-selected optimal CSI feature weightings. Extensive experiments on multiple simulated 5G NR benchmark datasets including three DeepMIMO indoor ray-tracing scenarios compliant with the 3GPP 5G NR cluster delay line (CDL) channel model, a 3GPP Indoor Office scenario, and an NYUSIM complex blockage dataset demonstrate that Pi-Loc outperforms established deep learning baselines (NN, DNN, LSTM, BiLSTM) in both accuracy and convergence efficiency, achieving near-mm-level positioning accuracy on simulated DeepMIMO datasets. Ablation studies confirm the framework’s stability and the individual contribution of each CSI feature category.