计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (24): 209-215.DOI: 10.3778/j.issn.1002-8331.2209-0004

• 图形图像处理 • 上一篇    下一篇

基于3D特征动态融合的点云特征提取网络

孙刘杰,翟仁杰,王文举,庞茂然   

  1. 上海理工大学 出版印刷与艺术设计学院,上海 200093
  • 出版日期:2023-12-15 发布日期:2023-12-15

Point Cloud Feature Extraction Network Based on 3D Feature Dynamic Fusion

SUN Liujie, ZHAI Renjie, WANG Wenju, PANG Maoran   

  1. College of Communication and Art Design, Shanghai University of Science and Technology, Shanghai 200093, China
  • Online:2023-12-15 Published:2023-12-15

摘要: 针对目前用于点云配准的点云特征提取方法并未充分提取点云中的有效信息等问题,提出了一种基于3D特征动态融合的点云特征提取网络(3D feature dynamic fusion and residual u-net,DFRUNet)。该网络通过3DFDF(3D feature dynamic fusion)模块将编码和解码模块的特征动态融合,以充分提取点云中的有效信息;同时采用SE-Res(squeeze and excitation residual)模块来提取点云特征,通过动态调整显著区域的权重,对该区域特征进行重点提取,以提高所提取特征的质量。将网络所提取特征映射到高维空间中,采用随机采样一致性(random sample consensus,RANSAC)算法完成点云配准。实验结果表明,在3DMatch数据集上,该算法特征匹配召回率(feature-match recall,FMR)达到了96.3%,相较于经典的FCGF算法提高了0.011。配准召回率(registration recall)达到了82.2%,提高了0.014。该方法充分提取了点云中的有效信息,达到了更高的召回率,对其他点云配准研究具有参考价值。

关键词: 特征提取, 点云配准, 特征动态融合, 深度学习

Abstract: To solve the problem that the feature extraction methods currently used for point cloud registration do not adequately extract effective information from point cloud, a point cloud feature extraction network DFRUNet (3D feature dynamic fusion and residual u-net) based on 3D feature dynamic fusion is proposed. The network dynamically fuses the features of encoding and decoding modules through 3DFDF (3D feature dynamic fusion) module to extract sufficient information from the point cloud. Meanwhile, the SE-Res (squeeze and excitation residual) module is used to extract point cloud features. By dynamically adjusting the weights of significant areas, the area features are extracted to improve the quality of the extracted features. Secondly, map the features extracted from the network into high-dimensional space, and complete point cloud registration using RANSAC (random sample consensus) algorithm. The experimental results show that on the 3DMatch dataset, the FMR (feature-match recall) of the algorithm is 96.3%, which is 0.011 higher than that of the classical FCGF algorithm. The registration recall rate is 82.2% and increased by 0.014. This method fully extracts the effective information from point clouds and achieves a higher recall rate, which has reference value for other point cloud registration studies.

Key words: feature extraction, point cloud registration, feature dynamic fusion, deep learning