曲线和多头移动通道自注意力机制融合的点云语义分割  

Curve and multi-head shifted channel self-attention mechanism fusion for point cloud semantic segmentation

作　　者：卢健[1] 郑雨飞 梁有成 罗立果 苏盛斌 LU Jian;ZHENG Yufei;LIANG Youcheng;LUO Liguo;SU Shengbin(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2025年第2期28-38,共11页Journal of Xi’an Polytechnic University

基　　金：陕西省自然科学基础研究计划重点项目(2018JZ6002);西安市碑林区应用技术研发项目(GX2305)。

摘　　要：针对点云语义分割中存在局部空间结构与深层次点云特征提取不充分问题,提出一种基于曲线和多头移动通道自注意力机制融合的三维点云语义分割网络。首先,曲线模块通过动态行走策略对点云进行分组和行走操作,获取远程点之间的关联性与几何相关性。其次,引入多头移动通道自注意力机制模块,通过滑动窗口对通道进行划分,并构建多头自注意力聚合通道特征,以捕获点云深层次的语义信息。最后,提出了反向瓶颈模块,通过将低维度MLP嵌入到插值结构中加深网络的层次,增强特征的表达能力,同时有效改善了梯度消失和过拟合问题。实验结果表明:该模型在S3DIS第五区域数据集上的准确率为90.1%,平均交并比为68.6%;在ScanNet数据集上用于测试的平均交并比为70.9%。In order to solve the problem of inadequate extraction of local spatial structure and deep-level point cloud features in point cloud semantic segmentation.We proposed a 3D point cloud semantic segmentation network based on the fusion of curve and multi-head shifted channel self-attention mechanism.First,the curve module performed grouping and walking operations on the point cloud through a dynamic walking strategy to obtain the correlation and geometric correlation between remote points.Secondly,the multi-head shifted channel self-attention mechanism module was introduced to segment the channels by sliding windows and construct multi-head self-attention aggregated channel features to capture the deep semantic information of the point cloud.Finally,the reverse bottleneck module was proposed to deepen the hierarchy of the network by embedding low-dimensional MLP into the interpolation structure to enhance the expression of the features,and at the same time to effectively improve to the gradient vanishing and overfitting problems.The experimental results show that the accuracy of this paper′s model is 90.1%and the mean intersection over union is 68.6%on the S3DIS Area 5 dataset;the mean intersection over union used for testing in ScanNet is 70.9%.

关 键 词：曲线模块 多头移动通道自注意力机制 点云 语义分割 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]