自适应生成卷积核的动态图注意力三维点云识别及分割  

Recognition and Segmentation of 3D Point Cloud by Dynamic Graph Attention with Adaptive Generated Convolutional Kernel

作　　者：杨军[1,2] 郭佳晨 YANG Jun;GUO Jiachen(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,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：《湖南大学学报（自然科学版）》2024年第12期139-152,共14页Journal of Hunan University:Natural Sciences

基　　金：国家自然科学基金资助项目(42261067)。

摘　　要：针对现有算法在对点云数据进行平移、缩放以及旋转等几何变换时网络不能充分提取局部特征,导致网络精度显著下降的问题,提出基于自适应生成卷积核的动态图注意力三维点云识别及分割算法.首先,利用感受野中心点位置信息增强邻点感知上下文信息能力,通过改进的自注意力机制重构感受野,使感受野内特征信息充分交互,增强感受野的上下文信息.其次,构造自适应生成卷积核,通过捕获变化的点云拓扑信息,自适应生成卷积核权重,进而提升网络性能.最后,构建动态图注意力卷积算子,并设计点云识别的动态网络与分割的U形网络.实验结果表明,本文算法在ModelNet40点云识别数据集的识别精度达到了94.0%,在ShapeNet Part点云部件语义分割数据集的平均交并比达到了86.2%.本文算法能够提取三维点云的关键特征信息,具有较好的三维点云识别与分割能力.As the current algorithms fail to fully extract local features and result in significant degradation of network accuracy when performing geometric transformations such as translation,scaling,and rotation on point cloud data,this paper proposes a dynamic graph attention-based 3D point cloud recognition and segmentation algorithm based on adaptive generated convolutional kernels.Firstly,the positional information of the center point in the receptive field is used to enhance the contextual information perception of neighboring points.The receptive field is reconstructed to enable sufficient interaction of feature information within the receptive field and enhance the contextual information by improving the self-attention mechanism.Then,an adaptive generated convolutional kernel is constructed to capture changing point cloud topology information and adaptively generate convolutional kernel weights to enhance network performance.Finally,a dynamic graph attention convolutional operator is built,and a dynamic network for point cloud recognition and a U-shaped network for segmentation are designed.The experimental results show that the proposed algorithm achieves a recognition accuracy of 94.0%in the ModelNet40 point cloud recognition dataset,and the instance mean intersection over union reaches 86.2%in the ShapeNet Part point cloud semantic segmentation dataset.The algorithm proposed can extract critical feature information from 3D point clouds and is capable of 3D point cloud recognition and segmentation.

关 键 词：三维点云 动态图注意力卷积 自适应算法 模型识别 语义分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]