面向点云分类和分割的形状自适应特征聚合网络  

Shape adaptive feature aggregation network for point cloud classification and segmentation

作　　者：蒋志豪 张美香 薛卫涛 付莉娜 文静[1] 李永强[2] 黄鸿[1] JIANG Zhihao;ZHANG Meixiang;XUE Weitao;FU Lina;WEN Jing;LI Yongqiang;HUANG Hong(Key Laboratory of Optoelectronic Technology and System,Ministry of Education,Chongqing University,Chongqing 400044,China;Product Testing Center,Beijing Institute of Space Machinery and Electronics,Beijing 100094,China)

机构地区：[1]重庆大学光电技术与系统教育部重点实验室,重庆400044 [2]北京空间机电研究所产品检测中心,北京100094

出　　处：《光学精密工程》2025年第5期777-788,共12页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.42071302);先进光学遥感技术北京市重点实验室开放基金资助项目(No.AORS202315);重庆市留学人员回国创业创新支持计划资助项目(No.cx2019144)。

摘　　要：点云分类与分割在机器人导航、虚拟现实以及自动驾驶领域应用广泛,大多面向点云处理的深度学习方法采用共享权重的多层感知机(MultiLayer Perceptron,MLP)以及单一的池化来聚合点云的局部特征,难以准确地描述排列复杂的点云结构信息。针对上述问题,提出一种点云形状自适应的局部特征编码方法,以有效表征形状多样的点云结构信息,提升点云分类和分割性能。该方法首先引入一种自适应特征增强模块,采用差分和可学习的调节因子对特征进行增强,弥补共享权重MLP描述能力不足的问题。在此基础上,设计了一种特征聚合模块,利用点云的绝对空间距离赋予不同点不同权重以适应形状多变的点云结构信息,突出有代表性的点集,更加准确地描述点云的局部结构信息。在3个大型公开点云数据集上进行实验,结果表明,在ModelNet40数据集上取得了93.9%的总体实例分类精度,在分割数据集ShapeNet和S3dis上分别取得了85.9%,59.7%的总体实例平均交并比(mean Intersection over Union,mIoU),本文提出的方法在点云分类和分割任务上表现优秀。The classification and segmentation of point clouds are widely applicable in robotic navigation,virtual reality,and autonomous driving.Most current deep learning approaches for point cloud processing employ multilayer perceptrons(MLPs)with shared weights and single pooling operations to aggregate local features.This methodology often hinders the accurate representation of structural information within point clouds exhibiting complex arrangements.To address these challenges,a novel point cloud shapeadaptive local feature encoding method was proposed,aimed at effectively capturing the structural informa⁃tion of point clouds with diverse geometric configurations while enhancing classification and segmentation performance.Initially,an adaptive feature enhancement module was introduced,this module utilized dif⁃ferentiation and learnable adjustment factors to strengthen the feature representation,compensating for the descriptive limitations inherent in shared weight MLPs.Building on this foundation,a feature aggregation module was designed to assign variable weights to distinct points based on their absolute spatial distances.This approach facilitates adaptation to the variable shapes of point cloud structures,accentuates representative point sets,and enables a more precise depiction of local structural information.Experimental evaluations conducted on three extensive public point cloud datasets reveal that the proposed method achieves exceptional performance in both classification and segmentation tasks,attaining an overall instance average classification accuracy of 93.9%on the ModelNet40 dataset,along with mean intersection over union(mIoU)scores of 85.9%and 59.7%on the ShapeNet and S3DIS datasets,respectively.

关 键 词：深度学习 点云分类 点云分割 局部特征聚合 

分 类 号：TP751.2[自动化与计算机技术—检测技术与自动化装置]