Scholarlink

<p>This study investigates the feasibility of repurposing the Blue Robotics Ping360—a low-cost, mechanically scanned single-beam sonar—for underwater object detection beyond its navigation role. It addresses three key questions: (i) what challenges arise when interpreting navigation sonar data in cluttered or reflective environments; (ii) how effective is manual annotation when combined with segmentation models such as U-Net; and (iii) whether affordable sonar can support complex object-level perception. Controlled pool experiments were conducted to examine acoustic artifacts including reflections, shadows, and range-dependent distortions. A manually annotated dataset was used to evaluate classical and deep-learning-based segmentation methods. Results show that preprocessing—near-field exclusion, denoising, and polar resampling—significantly improves detection clarity. Leveraging GPU-accelerated and data-parallel processing, the framework achieves scalable, near-real-time performance aligned with high-performance computing principles. The dataset and code are publicly available to encourage further research in dynamic and multi-sensor underwater perception.</p>

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

Interpreting navigation sonar data for object detection: a feasibility study using the Ping360

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月1日

查看全文
  • Md Junayed Hasan,
  • Somasundar Kannan,
  • Ali Rohan,
  • Amira Samy Talaat

摘要

This study investigates the feasibility of repurposing the Blue Robotics Ping360—a low-cost, mechanically scanned single-beam sonar—for underwater object detection beyond its navigation role. It addresses three key questions: (i) what challenges arise when interpreting navigation sonar data in cluttered or reflective environments; (ii) how effective is manual annotation when combined with segmentation models such as U-Net; and (iii) whether affordable sonar can support complex object-level perception. Controlled pool experiments were conducted to examine acoustic artifacts including reflections, shadows, and range-dependent distortions. A manually annotated dataset was used to evaluate classical and deep-learning-based segmentation methods. Results show that preprocessing—near-field exclusion, denoising, and polar resampling—significantly improves detection clarity. Leveraging GPU-accelerated and data-parallel processing, the framework achieves scalable, near-real-time performance aligned with high-performance computing principles. The dataset and code are publicly available to encourage further research in dynamic and multi-sensor underwater perception.