基于语义信息的三维点云全景分割方法研究  

Research on Panoptic Segmentation of 3D Point Clouds based on Semantic Information

作　　者：任不凡 黄小燕[1] 吴思东 蒋涛[1] 袁建英[1] REN Bufan;HUANG Xiaoyan;WU Sidong;JIANG Tao;YUAN Jianying(College of Automation,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学自动化学院,四川成都610225

出　　处：《成都信息工程大学学报》2023年第5期535-542,共8页Journal of Chengdu University of Information Technology

基　　金：国家自然科学基金资助项目(62103064);四川省科技厅资助项目(2021YFG0295、2021YFG0133、2021YFN0104、2021YFH0069、2022YFN 0020、2022YFS0565);四川省无人系统智能感知控制技术工程实验室开放课题资助项目(WRXT2020-001、WRXT2020-002、WRXT2020-005)。

摘　　要：针对端到端点云全景分割网络精度不足的问题,设计一种基于点云语义信息的全景分割算法。首先利用语义分割模型获取点云数据语义信息,然后结合点云语义和空间信息,对前景目标(车、人等)进行聚类。具体地,为避免同类别相邻目标被聚类为一个目标,提出融合法向量夹角特征、空间位置、语义信息的聚类算法进行准确的前景实例分割。最后,提出一种新的类别划分方法,在不影响后续决策处理情况下,显著增加分割质量。SemanticKITTI数据集上的实验结果表明,提出的方法在全景质量、分割质量、识别质量、平均交并比4个指标上取得了较好的效果,分别达到56.6%、82.3%、68.2%、68.1%,并保持较快的速度(175 ms),充分证明其有效性和实用性。The accuracy of end-to-end point cloud panoptic segmentation network is insufficient,this paper designs a panoptic segmentation algorithm based on point cloud semantic information to address this problem.This paper uses a semantic model to obtain the semantic information of point cloud,and then combined with spatial information to cluster the foreground objects(cars,people,etc.).Specifically,to avoid the adjacent objects of the same category from being clustered into one object,a clustering algorithm that integrates normal angle,spatial localization,and semantic information is proposed to accurately segment foreground instance.Finally,through further analysis and research on data categories,a new category classification method is proposed,which can significantly increase the segmentation quality without affecting the subsequent decision processing.The experimental results on the SemanticKITTI dataset show that the proposed method has achieved good results in four indicators:panoptic quality,segmentation quality,recognition quality,and average intersection ratio,respectively reaching 56.6%,82.3%,68.2%,68.1%,and maintain a relatively fast speed(175 ms),which further indicates the effectiveness and practicability of this method.

关 键 词：全景分割 深度学习 三维点云 聚类算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]