Virtual Covariance Assisted Dual-Task Vision Transformer for Distributed Array DOA Estimation

Abstract

Data driven methods based on deep learning have shown strong potential for direction of arrival (DOA) estimation using distributed arrays. However, most existing methods still rely on angular grids and a fixed source-number assumption, resulting in grid mismatch and reduced flexibility under varying numbers of incident sources. To address these limitations, this paper proposes a dual-task vision Transformer (DT-ViT) framework for DOA estimation. At the input stage, a virtual array is constructed by vectorizing the covariance matrix, with a hole mask introduced to mitigate the effects of zero-filled holes. A multi-scale patch embedding strategy is then employed, where the virtual covariance matrix is partitioned into patches of different sizes and fed into independent ViT encoder branches. Dedicated multi-layer perceptron (MLP) heads are attached to each branch to jointly predict the number of sources and the DOAs through source-number classification and angle regression, respectively. A weighted combination of smooth ${l}_{1}$ loss and cross-entropy loss is adopted to achieve coordinated optimization of the two tasks. Experimental results demonstrate that the proposed method achieves robust and high-accuracy source-number prediction and DOA estimation.