Multitarget direction of arrival trajectory estimation using BiLSTM and sequential allocation

Abstract The purpose of this study is multiple source direction of arrival (DoA) trajectory estimation using bidirectional long short‐term memory (BiLSTM) artificial neural network and sequential direction assignment. Conventional DoA tracking algorithms suffer from the necessity of accurate prior information and complex modeling. But in this article, the DoA trajectory estimation has been transformed into a mapping problem between a sequence of sample correlation matrices and a limited number of sampled directions from the corresponding trajectories. For this aim, several challenges must be resolved. One of these challenges is to define a hypothesis space of nonlinear and distinct trajectories, provided that they include most of the possible movement scenarios in the desired angular range. The second one is the lack of uniqueness between the sample correlation matrices and different permutations of DoAs. The last one is the assignment of the estimated DoAs to the corresponding trajectories. This article proposes polynomials for the path modeling of the moving sources. Then, directions and slopes of the paths at certain times are defined as the outputs of two separate BiLSTM networks. Finally, the directions and slopes of the paths are used for the sequential assignment of DoAs to the different trajectories. Simulation results indicate that the proposed method not only estimates the DoA trajectories of moving sources accurately, but also it significantly performs better than compressed sensing‐based and subspace‐based methods.