基于自注意力特征融合组卷积神经网络的三维点云语义分割  

Semantic segmentation of 3D point cloud based on self-attention feature fusion group convolutional neural network

作　　者：杨军 李博赞 YANG Jun;LI Bozan(Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学测绘与地理信息学院,甘肃兰州730070 [2]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《光学精密工程》2022年第7期840-853,共14页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.61862039);甘肃省科技计划资助项目(No.20JR5RA429);2021年度中央引导地方科技发展资金资助项目(No.2021-51);兰州市人才创新创业项目(No.2020-RC-22);兰州交通大学天佑创新团队(No.TY202002)。

摘　　要：针对现有算法忽略点云数据全局单点特征和局部几何特征的深层关系,导致捕获的局部几何信息缺乏鉴别性且难以有效识别复杂形状的问题,提出基于自注意力特征融合组卷积神经网络的三维点云语义分割算法。首先,设计轻量化网络框架的代理点图卷积提取点云局部几何特征,并加入组卷积操作减少计算量和复杂度,以较少的冗余信息增强特征的丰富性;其次,通过Transformer模块进行不同分支间特征信息的交流,使全局特征和局部几何特征相互补偿,增强特征的完备性;然后,将点云底层语义特征与原始点云融合以扩大局部邻域感受野,获得高级上下文语义信息;最后,将特征输入到分割模块完成细粒度语义分割。实验结果表明,该算法在S3DIS数据集和SemanticKITTI数据集上的分割精度分别达到79.3%和56.6%,能够提取三维点云的关键特征信息,网络参数量较少且具有较高的语义分割鲁棒性。The existing algorithms ignore the profound relationship between global single point features and local geometric features.This results in the lack of discriminative captured local geometric information and increases the difficulty of effectively identifying complex shape categories.This paper proposes a semantic segmentation algorithm for three-dimensional point clouds based on a self-attention feature fusion group convolutional neural network.First,the proxy point graph convolution of lightweight network is designed to extract the local geometric features of the point cloud.Then,the group convolution operation is added to reduce the amount of calculation and complexity and enhance the richness of features with less redundant information.Second,the feature information exchange between different branches is carried out through the Transformer module to ensure mutual compensation between the global and local geometric features and to enhance the completeness of features.Then,the underlying semantic features of the point cloud are fused with the original point cloud to expand the local neighborhood perception field and obtain high-level context semantic information.Finally,the features are input into the segmentation module to complete fine-grained semantic segmentation.The experimental results show that the segmentation accuracy reaches 79.3%and 56.6%in the S3DIS and SemanticKITTI datasets,respectively.This algorithm can extract the key feature information from a 3D point cloud using fewer network parameters and exhibits high robustness of semantic segmentation.

关 键 词：三维点云 语义分割 图卷积 组卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]