Autoregressive Attention Neural Networks for Non-Line-of-Sight User Tracking with Dynamic Metasurface Antennas

User localization and tracking in the upcoming generation of wireless networks have the potential to be revolutionized by technologies such as Dynamic Metasurface Antennas (DMAs). Commonly proposed algorithmic approaches rely on assumptions about relatively dominant Line-of-Sight (LoS) paths or may require pilot transmission sequences whose length is comparable to the number of DMA elements, thus leading to limited effectiveness and considerable measurement overheads in blocked LoS and dynamic multipath environments. Therefore, this paper proposes a two-stage machine-learning-based approach for user tracking, specifically designed for non-LoS multipath settings. A newly proposed Attention-based neural network is first trained to map noisy channel responses to potential user positions regardless of user-mobility patterns. This architecture constitutes a modification of the prominent Vision Transformer, specifically modified for extracting information from high-dimensional frequency response signals. As a second stage, its predictions for the past user positions are passed through a learnable autoregressive model to exploit the timecorrelated information and obtain the final position predictions; thus the problems of localization and tracking are decomposed. The channel estimation procedure leverages a DMA architecture with partially-connected Radio Frequency Chains (RFCs), which results to reduced numbers of pilot signals. The numerical evaluation over an outdoor ray-tracing scenario illustrates that despite LoS blockage, this methodology is capable of achieving high position accuracy across various multipath settings.