Fully Convolutional Network-Based DOA Estimation with Acoustic Vector Sensor

In this paper, a learning-based direction of arrival (DOA) estimation pipeline for acoustic vector sensor (AVS) is proposed. In the proposed pipeline, a fully convolutional network (FCN) is introduced for uncontaminated time-frequency (TF) point extraction, which is a crucial step for AVS-based DOA estimation. Unlike conventional direct path dominant (DPD) or single source points (SSP) detection, the uncontaminated TF point extraction problem is modeled as an image segmentation problem, where the direct DOA cues from the spatial response of AVS is utilized for ground truth labeling to generate the training data of the network. With the extracted uncontaminated TF points, the final DOA can be generated using the proposed fuzzy geometric median (FGM) clustering. Simulation results show that the proposed pipeline is capable of improving the accuracy in the cases of small angular difference between acoustic sources and improving robustness in strong reverberation and noise situations.