Scholarlink

<p>In autonomous driving, 3D object detection gives the vehicle the perception that enables it to drive safely and accurately in complex traffic environments. Camera-LiDAR 3D object detection is currently widely studied. However, there is still a great challenge to deal with the inherent data differences between the two modalities and achieve accurate feature fusion. Therefore, this paper introduces RoI-Guided Pseudo-LiDAR Point cloud Feature Enhancement Network (RPFE-Net), which contains a pseudo-point cloud discard module and point voxel ensemble network based on anti-noise submanifold convolution. It generates pseudo-point clouds with high density and semantic information from depth completion and merges them with LiDAR data. RPFE-Net effectively improves 3D object detection accuracy by solving the problem of computational cost due to high density of pseudo-point clouds and the problem of information loss during the voxelization of pseudo-point clouds. Experiments on the KITTI dataset demonstrate that our RPFE-Net can achieve decent detection accuracy and fast inference speed.</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

RPFE-Net: RoI-guided pseudo-LiDAR point cloud feature enhancement network for multi-modal 3D object detection

  • 期刊论文

  • 发表日期：

    2025年10月10日

查看全文
  • Ruifan Lin,
  • Xinxin Feng,
  • Yuren Chen,
  • Haifeng Zheng

摘要

In autonomous driving, 3D object detection gives the vehicle the perception that enables it to drive safely and accurately in complex traffic environments. Camera-LiDAR 3D object detection is currently widely studied. However, there is still a great challenge to deal with the inherent data differences between the two modalities and achieve accurate feature fusion. Therefore, this paper introduces RoI-Guided Pseudo-LiDAR Point cloud Feature Enhancement Network (RPFE-Net), which contains a pseudo-point cloud discard module and point voxel ensemble network based on anti-noise submanifold convolution. It generates pseudo-point clouds with high density and semantic information from depth completion and merges them with LiDAR data. RPFE-Net effectively improves 3D object detection accuracy by solving the problem of computational cost due to high density of pseudo-point clouds and the problem of information loss during the voxelization of pseudo-point clouds. Experiments on the KITTI dataset demonstrate that our RPFE-Net can achieve decent detection accuracy and fast inference speed.