Scholarlink

<p>In the forefront of LiDAR point cloud research, significant advancements have been seen in semantic segmentation, especially in single-frame point cloud segmentation. The application of multi-frame point clouds marks a new chapter the enhancement of dynamic and static object classification, effectively addressing occlusion problems. In this study, the spatial information of moving and static objects in a continuous multi-frame point cloud is different, so this study adopts a single spatial feature information for semantic segmentation modeling of dynamic and static objects, which breaks through the upper limit of the dynamic feature extraction capability of spatial information. Therefore, this study introduces the Adaptive Multi-Scale Kernel-Point Convolution Network (AMSKPconv), which bypasses temporal feature modeling and employs an encoder-decoder architecture that focuses solely on spatial information. In the coding phase of the network, an adaptive dynamic radius nearest neighbor method is introduced in this paper, which can automatically adjust the search radius according to the density distribution of the point cloud region. To better handle large multi-frame point clouds in outdoor environments, an adaptive rigid kernel is developed in conjunction with the Multi-Scale Kernel Point Convolution (MSKPconv). This combination allows the network to accurately capture both fine details and broader structural patterns. In this study, a dynamic and static perceptual contrastive loss function is proposed, which combined with the traditional cross-entropy loss, improves the network’s accuracy in distinguishing between dynamic and static objects within multi-frame point clouds. The AMSKPconv significantly improves the distinction between dynamic and static objects for accurate, large-scale, fine-grained semantic segmentation of outdoor multi-frame point clouds. Extensive testing on the SemanticKITTI and nuScenes datasets has confirmed the network’s superior performance, with a 3.73 per cent improvement compared to Mars3D and only a 1.7 per cent gap to Memoryseg, but outperforming Memoryseg in detection accuracy of dynamic objects. These results not only demonstrate the effectiveness of the AMSKPconv network, but also provide some references for future research in semantic segmentation of large-scale outdoor multi-frame point clouds.</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

Adaptive multi-scale KPconv: a semantic segmentation network for multi-frame point clouds of vehicle-mounted LiDAR

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Yi Wang,
  • Zihan Zhou,
  • Yufei Xu,
  • Hongmei Zhao,
  • Juan Li,
  • Junyao Gao,
  • Hai-kun Wang

摘要

In the forefront of LiDAR point cloud research, significant advancements have been seen in semantic segmentation, especially in single-frame point cloud segmentation. The application of multi-frame point clouds marks a new chapter the enhancement of dynamic and static object classification, effectively addressing occlusion problems. In this study, the spatial information of moving and static objects in a continuous multi-frame point cloud is different, so this study adopts a single spatial feature information for semantic segmentation modeling of dynamic and static objects, which breaks through the upper limit of the dynamic feature extraction capability of spatial information. Therefore, this study introduces the Adaptive Multi-Scale Kernel-Point Convolution Network (AMSKPconv), which bypasses temporal feature modeling and employs an encoder-decoder architecture that focuses solely on spatial information. In the coding phase of the network, an adaptive dynamic radius nearest neighbor method is introduced in this paper, which can automatically adjust the search radius according to the density distribution of the point cloud region. To better handle large multi-frame point clouds in outdoor environments, an adaptive rigid kernel is developed in conjunction with the Multi-Scale Kernel Point Convolution (MSKPconv). This combination allows the network to accurately capture both fine details and broader structural patterns. In this study, a dynamic and static perceptual contrastive loss function is proposed, which combined with the traditional cross-entropy loss, improves the network’s accuracy in distinguishing between dynamic and static objects within multi-frame point clouds. The AMSKPconv significantly improves the distinction between dynamic and static objects for accurate, large-scale, fine-grained semantic segmentation of outdoor multi-frame point clouds. Extensive testing on the SemanticKITTI and nuScenes datasets has confirmed the network’s superior performance, with a 3.73 per cent improvement compared to Mars3D and only a 1.7 per cent gap to Memoryseg, but outperforming Memoryseg in detection accuracy of dynamic objects. These results not only demonstrate the effectiveness of the AMSKPconv network, but also provide some references for future research in semantic segmentation of large-scale outdoor multi-frame point clouds.