Scholarlink

<p>The primary objective of point cloud registration is to determine the optimal spatial correspondence between multiple point cloud datasets, which typically involves computing a transformation that aligns the point cloud data accurately. Despite the existence of various registration schemes, they are still affected by factors such as noise, complex surfaces, partial overlap, occlusion, and scale differences, which lead to inaccuracies and incompleteness in feature extraction, thus limiting the improvement in registration performance. This paper aims to effectively exploit information at different resolution levels within point cloud data and propose a multi-scale feature fusion point cloud registration algorithm (MSM, MultiScaleMatcher) combined with a cross-attention mechanism. This method extracts point cloud features at different resolution levels in each edge convolution layer of DGCNN, integrates them through a cross-attention mechanism, and generates the final multi-scale fused features through weighted merging. Experimental results show that on the ModelNet40 dataset, the RMSE(R) and RMSE(t) of the MSM algorithm are 1.79760 and 0.0027, respectively, while the MAE(R) and MAE(t) are 0.1557 and 0.0047, respectively. On the ScanObjectNN dataset, the RMSE(R) and RMSE(t) of the MSM algorithm are 2.1144 and 0.0073, respectively, while the MAE(R) and MAE(t) are 0.2833 and 0.0091, respectively, effectively improving the performance of point cloud registration.</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

MSM: point cloud alignment algorithm through multi-scale feature fusion and cross-attention mechanism

  • 期刊论文

  • 发表日期：

    2025年5月9日

查看全文
  • Keyuan Qiu,
  • Yingjie Zhang,
  • Feng Chen

摘要

The primary objective of point cloud registration is to determine the optimal spatial correspondence between multiple point cloud datasets, which typically involves computing a transformation that aligns the point cloud data accurately. Despite the existence of various registration schemes, they are still affected by factors such as noise, complex surfaces, partial overlap, occlusion, and scale differences, which lead to inaccuracies and incompleteness in feature extraction, thus limiting the improvement in registration performance. This paper aims to effectively exploit information at different resolution levels within point cloud data and propose a multi-scale feature fusion point cloud registration algorithm (MSM, MultiScaleMatcher) combined with a cross-attention mechanism. This method extracts point cloud features at different resolution levels in each edge convolution layer of DGCNN, integrates them through a cross-attention mechanism, and generates the final multi-scale fused features through weighted merging. Experimental results show that on the ModelNet40 dataset, the RMSE(R) and RMSE(t) of the MSM algorithm are 1.79760 and 0.0027, respectively, while the MAE(R) and MAE(t) are 0.1557 and 0.0047, respectively. On the ScanObjectNN dataset, the RMSE(R) and RMSE(t) of the MSM algorithm are 2.1144 and 0.0073, respectively, while the MAE(R) and MAE(t) are 0.2833 and 0.0091, respectively, effectively improving the performance of point cloud registration.