摘要
城市点云语义分割通过解析城市元素,为城市规划提供精确的场景理解能力.城市环境的数据分布不一致性对点云语义分割方法的实际应用提出了极高的要求.现有点云分割方法尚不足以高效地完成任意点云场景上的语义分割流程.本文提出一种混合点云语义分割方法,基本思想是结合深度学习预训练模型对不同数据分布之间分布共性的良好提取和在线学习方法对分布差异的良好表达,共同高效完成任意点云场景上的语义分割流程.预训练模型对点云场景的初始分割结果可以替代在线学习所需的大量标注数据,在线学习机制则能够纠正预训练模型因泛化能力不足导致的分割错误.实验证明,本文算法在新点云场景上的分割准确率和所需时间均优于主流点云分割方法.
Urban point cloud semantic segmentation provides precise scene understanding capabilities for urban planning by parsing urban elements.The inconsistency in data distribution of urban environments poses high demands on the practical application of point cloud semantic segmentation methods.Existing point cloud segmentation methods are not efficient enough to complete the semantic segmentation process on arbitrary point cloud scenes.This paper presents a hybrid point cloud semantic segmentation method.The basic idea is to efficiently complete the semantic segmentation process on arbitrary point cloud scenes by combining the good extraction of distribution commonalities between different data distributions using pre-trained deep learning models and the good representation of distribution differences using online learning methods.The initial segmentation results of point cloud scenes by pre-trained models can replace a large amount of annotated data required for online learning,while the online learning mechanism can correct segmentation errors caused by the insufficient generalization ability of the pre-trained models.Experimental results show that the algorithm in this paper outperforms mainstream point cloud segmentation methods in terms of segmentation accuracy and required time in new point cloud scenes.
作者
郝鹏举
秦健翔
张严辞
HAO Pengju;QIN Jianxiang;ZHANG Yanci(College of Computer Science,Sichuan University,Chengdu 610065,China;National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065,China)
出处
《小型微型计算机系统》
北大核心
2025年第6期1373-1382,共10页
Journal of Chinese Computer Systems
基金
国家重大专项项目(GJXM92579)资助
四川省重点研发项目(2023YFG0122)资助.
关键词
点云处理
语义分割
在线学习
随机森林
交互分割
point cloud processing
semantic segmentation
online learning
random forest
interactive segmentation
