马天恩, 刘涛, 杜萍, 陈朴一, 令振飞. 一种聚合全局上下文信息的三维点云语义分割方法[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20230143
引用本文: 马天恩, 刘涛, 杜萍, 陈朴一, 令振飞. 一种聚合全局上下文信息的三维点云语义分割方法[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20230143
MA Tian-en, LIU Tao, DU Ping, CHEN Po-yi, LING Zhen-fei. A 3D Point Cloud Semantic Segmentation Method for Aggregating Global Context Information[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20230143
Citation: MA Tian-en, LIU Tao, DU Ping, CHEN Po-yi, LING Zhen-fei. A 3D Point Cloud Semantic Segmentation Method for Aggregating Global Context Information[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20230143

一种聚合全局上下文信息的三维点云语义分割方法

A 3D Point Cloud Semantic Segmentation Method for Aggregating Global Context Information

  • 摘要: 现有基于深度学习的三维点云语义分割方法在很大程度上忽略了全局上下文信息,未能充分利用点云的局部几何结构、颜色信息和高层语义特征之间的互补性。针对该问题,本文提出一种融合局部特征编码和密集连接的点云语义分割模型。首先,设计一个局部特征提取模块,使得模型能够同时捕获空间几何结构、颜色信息和语义特征;其次,结合局部特征聚合模块保留原始点云数据中丰富的几何信息,减少特征提取过程中几何信息的损失;最后,利用密集连接模块来聚合全局上下文信息,实现低层特征和高层语义信息的互补。采用S3DIS和Semantic3D两个大型基准数据集进行实验验证,发现本文提出的模型在两个数据集上的平均交并比分别达到71.8%和77.8%。实验结果表明,本文方法在三维点云语义分割方面具有较好的性能。

     

    Abstract: Objectives: Existing deep learning-based 3D point cloud semantic segmentation methods often overlook global contextual information and do not fully leverage the synergy between the local geometric structure of the point cloud, color information, and high-level semantic features. To effectively capture the geometric structure, color variations, and semantic features of point clouds while retaining global context information. Methods: We propose a point cloud semantic segmentation model that integrates local feature encoding and dense connectivity. First, a local feature extraction module is employed to enable the model to concurrently capture spatial geometric structure, color information, and semantic features. Second, a local feature aggregation module is incorporated to preserve the rich geometric data within the original point cloud, minimizing information loss during feature extraction. Finally, we utilize a dense connectivity module to aggregate contextual semantic information, promoting synergy between low-level features and high-level semantic data. Results: Our model is benchmarked against two large datasets, S3DIS and Semantic3D. The results show that our proposed network model achieves an Overall Accuracy (OA) and mean Intersection over Union (mIoU) of 88.3% and 71.8% on the S3DIS dataset, improving the baseline set by RandLA-Net by 0.3% and 1.8% respectively. On the Semantic3D dataset, we registered an OA of 94.9% and an mIoU of 77.8%, marking respective improvements of 0.1% and 0.4% over RandLA-Net. Conclusions:Our model effectively preserves local geometric and color information through local feature encoding. The Local Feature Aggregation Module refines point proximity along boundaries to align with similar feature domains. Dense connections successfully integrate global context and key geometric features. Overall, our approach delivers more accurate semantic labels and a superior geometric feature representation, enhancing the precision of local segmentations.

     

/

返回文章
返回