Scholarlink

<p>Body-in-white (BIW) is the skeletal framework of a vehicle body, composed of welded stamped metal parts and serving as the foundational structure in automotive manufacturing. In automotive production, point cloud data of BIW components plays a pivotal role in defect detection and rectification before pre-assembly. The presence of holes in scanned point clouds of BIW components poses a significant challenge to accurate quality inspection. This paper proposes a hybrid geometric-deep learning solution for BIW parts hole repair. Our approach begins with a robust detection of hole boundaries on double-sided thin-walled structures. Subsequently, a Boundary-Constrained Centroidal Voronoi Tessellation (BC-CVT) method generates an irregular yet uniform set of 2D points within the hole region, providing a strong geometric prior. A dedicated point cloud repair network, empowered by a Self-Attention EdgeConv (SAE) module, then learns local shape contexts from the boundary to accurately predict the corresponding 3D coordinates of the 2D points. The experiments demonstrate that our algorithm achieves superior repair accuracy and produces better overall repair effect.</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

Accurate hole repair in Body-in-White parts point clouds via a deep neural network

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Hui Wang,
  • Ming Li,
  • QingYue Wei

摘要

Body-in-white (BIW) is the skeletal framework of a vehicle body, composed of welded stamped metal parts and serving as the foundational structure in automotive manufacturing. In automotive production, point cloud data of BIW components plays a pivotal role in defect detection and rectification before pre-assembly. The presence of holes in scanned point clouds of BIW components poses a significant challenge to accurate quality inspection. This paper proposes a hybrid geometric-deep learning solution for BIW parts hole repair. Our approach begins with a robust detection of hole boundaries on double-sided thin-walled structures. Subsequently, a Boundary-Constrained Centroidal Voronoi Tessellation (BC-CVT) method generates an irregular yet uniform set of 2D points within the hole region, providing a strong geometric prior. A dedicated point cloud repair network, empowered by a Self-Attention EdgeConv (SAE) module, then learns local shape contexts from the boundary to accurately predict the corresponding 3D coordinates of the 2D points. The experiments demonstrate that our algorithm achieves superior repair accuracy and produces better overall repair effect.