Scholarlink

<p>3D object detection is a crucial technology for enabling automated driving. Methods with LiDAR employing Bird’s-Eye View (BEV) representations of point cloud in a Cartesian coordinate system have become mainstream. This technology requires recognizing objects at long distances, exceeding 100&#xa0;m. In autonomous driving scenarios, such as when monitoring oncoming lanes for a lane change across opposing traffic, long-range perception is essential. However, methods using BEV representations of point cloud face a significant challenge: the computational resources required increase enormously when attempting to recognize objects at long distances. To address this, we propose a novel approach that combines Cartesian and polar coordinate systems. This method efficiently enables perception of objects up to approximately 140&#xa0;m away, all while maintaining high detection accuracy for small objects at short-range. Furthermore, with real-time operation in mind, we focused on accelerating our method and achieved an inference time of 83.30 ms per frame.</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

Real-Time 3D Object Detection with Distance-Aware Hybrid Point Cloud Representation toward Long Range Detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月26日

查看全文
  • Keigo Hariya,
  • Hiroki Inoshita,
  • Yukiya Fukuda,
  • Keisuke Yoneda,
  • Naoki Suganuma

摘要

3D object detection is a crucial technology for enabling automated driving. Methods with LiDAR employing Bird’s-Eye View (BEV) representations of point cloud in a Cartesian coordinate system have become mainstream. This technology requires recognizing objects at long distances, exceeding 100 m. In autonomous driving scenarios, such as when monitoring oncoming lanes for a lane change across opposing traffic, long-range perception is essential. However, methods using BEV representations of point cloud face a significant challenge: the computational resources required increase enormously when attempting to recognize objects at long distances. To address this, we propose a novel approach that combines Cartesian and polar coordinate systems. This method efficiently enables perception of objects up to approximately 140 m away, all while maintaining high detection accuracy for small objects at short-range. Furthermore, with real-time operation in mind, we focused on accelerating our method and achieved an inference time of 83.30 ms per frame.