Scholarlink

Accurate Direction-of-Arrival (DOA) estimation is crucial for the efficacy of Unattended Ground Sensor (UGS) systems, enhancing sound localization, situational awareness, resource optimization, and integration with other sensor data for comprehensive monitoring. The need for sensors to be lightweight and miniaturized for versatility in diverse environments is growing. Despite deep learning advancements, DOA estimation of low-frequency reverberant signals with dense small-aperture microphone arrays poses significant challenges for conventional and neural network methods. This paper presents the Spatial-Temporal Attention Residual Network (STAR-Net), aimed at precise DOA estimation of low-frequency sounds in noisy environments, utilizing densely packed arrays with apertures smaller than the half-wavelength of the sound source. STAR-Net’s performance was evaluated against Brownian and Gaussian noise simulations, reflecting wind and general background disturbances. Remarkably, STAR-Net achieves a DOA estimation accuracy of 75.61% and a Mean Angular Error (MAE) of just 11.00 \(^\circ \) in demanding low-frequency signal scenarios characterized by a low SNR of 0 dB and an RT60 of 0.6 s. Experimental results validate that enhancing DOA precision by increasing the number of microphones, without changing the array’s aperture, remains effective in deep learning contexts.

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

Advancing DOA Estimation with Ultra-dense Small-Aperture Acoustic Arrays: A Spatial-Temporal Transformer Approach for Low-Frequency Reverberation Signals

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kunkun Zhang,
  • Kehao Wu,
  • Qianwei Zhou,
  • Jun Tao,
  • Haigen Hu,
  • Xiaoxin Li

摘要

Accurate Direction-of-Arrival (DOA) estimation is crucial for the efficacy of Unattended Ground Sensor (UGS) systems, enhancing sound localization, situational awareness, resource optimization, and integration with other sensor data for comprehensive monitoring. The need for sensors to be lightweight and miniaturized for versatility in diverse environments is growing. Despite deep learning advancements, DOA estimation of low-frequency reverberant signals with dense small-aperture microphone arrays poses significant challenges for conventional and neural network methods. This paper presents the Spatial-Temporal Attention Residual Network (STAR-Net), aimed at precise DOA estimation of low-frequency sounds in noisy environments, utilizing densely packed arrays with apertures smaller than the half-wavelength of the sound source. STAR-Net’s performance was evaluated against Brownian and Gaussian noise simulations, reflecting wind and general background disturbances. Remarkably, STAR-Net achieves a DOA estimation accuracy of 75.61% and a Mean Angular Error (MAE) of just 11.00 \(^\circ \) in demanding low-frequency signal scenarios characterized by a low SNR of 0 dB and an RT60 of 0.6 s. Experimental results validate that enhancing DOA precision by increasing the number of microphones, without changing the array’s aperture, remains effective in deep learning contexts.