A deep learning method for reflective boundary estimation.

Area of Science:

• Acoustics
• Machine Learning
• Robotics

Background:

• Environment estimation in reverberant settings like underwater and indoor acoustics is challenging.
• Existing boundary estimation methods require high signal-to-noise ratios and accurate echo assignment.
• Limitations hinder accurate localization and mapping in complex acoustic environments.

Purpose of the Study:

• To develop a robust two-dimensional boundary estimation method using convolutional neural networks (NNs).
• To overcome limitations of existing methods, including low signal-to-noise ratios and spurious echo detection.
• To enable accurate environment estimation in reverberant acoustic domains.

Main Methods:

• A convolutional neural network (NN) approach is employed for boundary estimation.
• A Hough transform-inspired algorithm converts echo times of arrival into images.
• Multi-resolution regression by NNs refines boundary estimation sequentially.
• The method does not require correct echo labeling and is robust to reverberation.

Main Results:

• The proposed NN method demonstrates robust two-dimensional boundary estimation.
• Performance was validated through simulations and real-world water tank data.
• The method significantly outperforms state-of-the-art alternatives in challenging conditions.

Conclusions:

• The developed data-driven method offers a robust solution for environment estimation in reverberant settings.
• This approach holds significant potential for advancing underwater acoustic detection and tracking.
• Future work can extend this to three-dimensional environment estimation.