计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (10): 217-226.DOI: 10.3778/j.issn.1002-8331.2301-0230

• 图形图像处理 • 上一篇    下一篇

可学习动态分组卷积神经网络的大规模点云分割

康玥,杨军   

  1. 1.兰州交通大学 自动化与电气工程学院,兰州 730070
    2.兰州交通大学 电子与信息工程学院,兰州 730070
  • 出版日期:2024-05-15 发布日期:2024-05-15

Large-Scale Point Cloud Segmentation by Learnable Dynamic Grouping Convolutional Neural Network

KANG Yue, YANG Jun   

  1. 1. School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
    2. School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2024-05-15 Published:2024-05-15

摘要: 针对现有大规模点云语义分割算法提取特征时冗余干扰信息过多,导致神经网络分割性能较差的问题,提出可学习动态分组卷积神经网络架构,高效准确地实现大规模点云分割。对输入点云以分组的方式进行局部几何特征提取,并通过动态筛选和修剪冗余特征通道来减少无用特征信息对神经网络特征识别的干扰,进一步提高网络模型语义分割精度。构建位置编码模块,将点云位置特征映射到高维频域空间,使神经网络充分挖掘点云频域特征信息,增强特征的丰富性。对提取到的局部几何特征和全局单点位置特征进行融合,并构建可学习动态分组卷积神经网络,完成解码得到最终分割结果。实验结果表明,该算法在大规模点云分割数据集S3DIS和SemanticKITTI上的mIoU分别为69.6%和58.3%。与现有点云语义分割方法相比,所提出的网络模型具有更高的分割准确率和较低的参数量。

关键词: 大规模点云, 语义分割, 可学习动态分组卷积, 位置编码

Abstract: There exists too much redundant interference information when large-scale point cloud semantic segmentation algorithms extract features, which results in the poor segmentation performance of neural networks. To solve this problem, a learnable dynamic grouping convolutional neural network architecture is proposed to efficiently realize large-scale point cloud segmentation. Firstly, the algorithm extracts local geometric features from the input point cloud in a grouped manner and reduces the interference of useless feature information on neural network feature recognition by dynamically filtering and pruning redundant feature channels, while improving the accuracy of semantic segmentation. Secondly, a positional encoding module is built to map the position feature of the point cloud to the high-dimensional frequency domain space, so that the neural network can fully mine the feature information of the point cloud and enhance the richness of features. Finally, the extracted local geometric feature and position feature are fused, while building a learnable dynamic grouping convolutional neural network to get the final segmentation result. Experimental results show that the mIoU of this algorithm on large-scale point cloud segmentation datasets S3DIS and SemanticKITTI is 69.6% and 58.3%, respectively. Compared with existing point cloud semantic segmentation methods, the proposed network model has higher segmentation accuracy and fewer network parameters.

Key words: large-scale point cloud, semantic segmentation, learnable dynamic grouping convolution, positional encoding