Graph Network-Based UWB Localization via Learning Spatial-Temporal and Geometric Features

In this letter, we propose a Graph-Attention-Recurrent Neural Network (Graph-ARNN) to improve UWB localization in complex environment by incorporating spatial, temporal and geometric information. We first build the ranging measurements from UWB sensors as a large spatial-temporal graph structure, and then the Graph-ARNN including the graph convolutional model, graph-attention model and deep RNN model are designed to extract the high-level spatial-temporal and geometric features which beneficial to tag location estimation. Thus, the localization performance can be improved. We also conduct three real-world experiments with both LOS and NLOS environments to suggest the advantages of our proposed method.