《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (10): 3202-3208.DOI: 10.11772/j.issn.1001-9081.2023020119

• 多媒体计算与计算机仿真 • 上一篇    

基于多特征融合的点云场景语义分割

郝雯(), 汪洋, 魏海南   

  1. 西安理工大学 计算机科学与工程学院,西安 710048
  • 收稿日期:2023-02-15 修回日期:2023-04-03 接受日期:2023-04-07 发布日期:2023-08-14 出版日期:2023-10-10
  • 通讯作者: 郝雯
  • 作者简介:汪洋(1998—),男,安徽合肥人,硕士研究生,主要研究方向:点云场景分割
    魏海南(1998—),女,河北承德人,硕士研究生,主要研究方向:点云场景分割。
  • 基金资助:
    国家自然科学基金资助项目(61602373);陕西省自然科学基金资助项目(2021JM-342);西安市碑林区研发项目(GX2242)

Semantic segmentation of point cloud scenes based on multi-feature fusion

Wen HAO(), Yang WANG, Hainan WEI   

  1. Faculty of Computer Science and Engineering,Xi’an University of Technology,Xi’an Shaanxi 710048,China
  • Received:2023-02-15 Revised:2023-04-03 Accepted:2023-04-07 Online:2023-08-14 Published:2023-10-10
  • Contact: Wen HAO
  • About author:WANG Yang,born in 1998, M. S. candidate. His research interests include point cloud scene segmentation.
    WEI Hainan, born in 1998, M. S. candidate. Her research interests include point cloud scene segmentation.
  • Supported by:
    National Natural Science Foundation of China(61602373);Natural Science Foundation of Shaanxi Province(2021JM-342);Xi’an BeiLin Science Research Plan(GX2242)

摘要:

为挖掘特征间的语义关系以及空间分布信息,并通过多特征增强进一步改善点云语义分割的效果,提出一种基于多特征融合的点云场景语义分割网络(MFF-Net)。所提网络以点的三维坐标和改进后的边特征作为输入,首先,利用K-近邻(KNN)算法搜寻点的近邻点,并在三维坐标和近邻点间坐标差值的基础上计算几何偏移量,从而增强点的局部几何特征表示;其次,将中心点与近邻点间的距离作为权重信息更新边特征,并引入空间注意力机制,获取特征间的语义信息;再次,通过计算近邻特征间的差值,利用均值池化操作进一步提取特征间的空间分布信息;最后,利用注意力池化操作融合三边特征。实验结果表明,所提网络在S3DIS(Stanford 3D large-scale Indoor Spaces)数据集上的平均交并比(mIoU)达到了67.5%,总体准确率(OA)达到了87.2%,相较于PointNet++分别提高10.2和3.4个百分点,可见MFF-Net在大型室内/室外场景均能获得良好的分割效果。

关键词: 点云, 语义分割, 空间注意力, 注意力池化, 特征融合

Abstract:

In order to mine the semantic relationships and spatial distribution among features, and further improve the semantic segmentation results of point cloud through multi-feature enhancement, a Multi-Feature Fusion based point cloud scene semantic segmentation Network (MFF-Net) was proposed. In the proposed network, the 3D coordinates and improved edge features were used as input, firstly, the neighbor points of the point were searched by using K-Nearest Neighbor (KNN) algorithm, and the geometric offsets were calculated based on 3D coordinates and coordinate differences among neighbor points, which enhanced the local geometric feature representation of points. Secondly, the distance between the central point and its neighbor points were used to as weighting information to update the edge features, and the spatial attention mechanism was introduced to obtain the semantic information among features. Thirdly, the spatial distribution information among features was further extracted by calculating the differences among neighbor features and using mean pooling operation. Finally, the trilateral features were fused based on attention pooling. Experimental results demonstrate that on S3DIS (Stanford 3D large-scale Indoor Spaces) dataset, the mean Intersection over Union (mIoU) of the proposed network is 67.5%, and the Overall Accuracy (OA) of the proposed network is 87.2%. These two values are 10.2 and 3.4 percentage points higher than those of PointNet++ respectively. It can be seen that MFF-Net can achieve good segmentation results in both large indoor and outdoor scenes.

Key words: point cloud, semantic segmentation, spatial attention, attention pooling, feature fusion

中图分类号: