Scholarlink

Direction-of-Arrival (DoA) estimation is crucial in spatial hearing, robotics, and human-computer interaction. Traditional methods like GCC-PHAT and MUSIC perform adequately in static settings but degrade under motion or reverberation. This study presents a binaural sound localization and distance estimation system using deep learning models—Convolutional Neural Network (CNN), Residual Network (ResNet), and Convolutional Recurrent Neural Network (CRNN) integrated with traditional signal processing and enhanced by a rotation-aware algorithm for full 360 \(^\circ \) coverage. A custom dataset simulates static, head rotation, and walking scenarios. The CRNN-based model achieves 5.79 \(^\circ \) DoA error, 0.97 m distance error, and 91.54% accuracy in the static case. Under dynamic conditions, it maintains robust performance with 92.33% accuracy for walking and 89.87% during rotation. Results confirm the model’s robustness across realistic environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of Direction-of-Arrival and Distance for Binaural Sound Sources Using a CRNN-Based Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Abhinay Chanda,
  • Gautam Mohanty,
  • Ishaan Sharma,
  • K. Thriveni,
  • G Ajay Kumar Naik,
  • Ch. V. Rama Rao

摘要

Direction-of-Arrival (DoA) estimation is crucial in spatial hearing, robotics, and human-computer interaction. Traditional methods like GCC-PHAT and MUSIC perform adequately in static settings but degrade under motion or reverberation. This study presents a binaural sound localization and distance estimation system using deep learning models—Convolutional Neural Network (CNN), Residual Network (ResNet), and Convolutional Recurrent Neural Network (CRNN) integrated with traditional signal processing and enhanced by a rotation-aware algorithm for full 360 \(^\circ \) coverage. A custom dataset simulates static, head rotation, and walking scenarios. The CRNN-based model achieves 5.79 \(^\circ \) DoA error, 0.97 m distance error, and 91.54% accuracy in the static case. Under dynamic conditions, it maintains robust performance with 92.33% accuracy for walking and 89.87% during rotation. Results confirm the model’s robustness across realistic environments.