Time-Varying Direction-of-Arrival Estimation Exploiting Mamba Network

Direction-of-arrival (DOA) estimation for moving targets presents a significant challenge in array signal processing. Traditional DOA estimation and tracking methods often encounter limitations due to the infeasibility of acquiring large volumes of stationary data and performing subspace-based processing over many snapshots, and lead to high computational costs. Recently, deep learning techniques have been effectively applied in DOA estimation, owing to their reduced complexity during inference. In this paper, we propose the use of Mamba network as a state-space model-based approach to estimate and track DOAs that vary snapshot-by-snapshot. The proposed network is interpretable and hardware-efficient, making it advantageous for training and real-time inference.