Reliable underwater multi-target direction of arrival estimation with optimal transport using deep models

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Multi-target direction of arrival (DoA) estimation is crucial but challenging in sonar signal processing.
• Existing methods often struggle with accuracy and robustness in complex acoustic environments.

Purpose of the Study:

• To develop a novel deep learning method for accurate multi-target DoA estimation using a single model.
• To introduce an optimal transport (OT) loss to better handle the continuous nature of angular data in DoA estimation.

Main Methods:

• Proposed the learning direction of arrival with optimal transport (LOT) approach, modeling DoA estimation as a multi-label classification task.
• Introduced an OT loss with a custom cost matrix to capture angular grid properties, enhancing prediction accuracy.
• Developed a lightweight channel mask data augmentation module for covariance matrix-based deep models.

Main Results:

• The LOT approach demonstrated improved accuracy in DoA estimation compared to baseline methods.
• The proposed methods showed effectiveness and robustness across various experimental scenarios and measurements.
• Experiments on SwellEx-96 data confirmed the practicality and real-world applicability of the LOT approach.

Conclusions:

• The LOT method offers a robust and accurate solution for multi-target DoA estimation in sonar.
• The proposed optimal transport loss and data augmentation module enhance the performance of deep learning models for DoA estimation.
• The developed techniques are portable across different network architectures and show promise for practical sonar applications.