A Tensor Decomposition Technique for Source Localization from Multimodal Data

Abstract

This paper studies the problem of localizing a source based on different types of signals measured at different sensing locations, where propagation models of the signals are not known. A tensor observation model is proposed to arrange such multimodal data into different layers to form a 3D data array. It is proven that the vectors extracted from the least squares rank-1 approximation of the tensor under the Tucker's model are location signature vectors of the source, where the vectors are unimodal and their peak locations correspond to the source location. Numerical experiments demonstrate that the proposed localization method based on tensor decomposition outperforms the baseline that heuristically averages the estimates individually from different types of data.