TDoA positioning with data-driven LoS inference in mmWave MIMO communications

Location awareness is an essential feature to support various mobile services, and cooperative positioning with channel state information (CSI) in millimeter wave multiple-input multiple-output (MIMO) networks is promising. Meanwhile, strong non-line of sight (NLoS) effects in outdoor scenarios severely reduce the model-based localization accuracy, and existing fingerprint-based methods have a critical requirement for labeled data and lack scalability. To overcome these shortcomings, we propose a hybrid data-and-model driven cooperative localization scheme using uplink wideband MIMO CSI at multiple base stations (BS) as measurements. First, we design a data-driven line of sight (LoS) inference module, realized by a transformer with an attention mechanism, to estimate the statistics of LoS arrival times. Each BS is equipped with a module, and the counterpart LoS inferences are simultaneously and distributedly performed. Second, we locate the mobile user by a model-driven approximate maximum likelihood time difference of arrival algorithm with the estimated statistics. Experiment results in urban scenarios show that the proposed localization scheme is extensible and also robust to varying number of channel paths and combination of BS measurements. Considering a small training dataset, the proposed scheme significantly outperforms the data- and model-driven baselines in NLoS scenarios, in terms of positioning accuracy.