基于空间序列化的高效室内点云语义分割方法

打开文本图片集

关键词：点云语义分割；空间序列化；几何感知；三维场景理解 中图分类号：TP391 文献标识码：A doi：10.37188/CJLCD.2025-0197 CSTR：32172.14.CJLCD.2025-0197

Efficient indoor point cloud semantic segmentation method basedonspatial serialization

CHENMingtao 1 ，WANGHaoting，SHANGYanfei²，ZHANGPengbo 3 ，CHENHuil （1.College ofAutomation Engineering， Shanghai University ofElectric Power， Shanghai 2Oo09o，China; 2.Lithium Yue Neru Energy Group，Shanghai 2Ol6OO，China; 3.Pylon Technologies Co.Ltd.， Shanghai 2OOl2O，China）

Abstract： In order to address the insufficient global structure modeling and poor geometric detail preservation in large sparse point clouds indoors，which cannot be applied to various scene issues directly，this paper proposes a novel network architecture that integrates Spatial Point Filing Serialization（SPFS） with Geometry-aware Channel Propagation（GCP）. The SPFS module employs an adaptive space-filling curve to order neighborhood points，explicitly preserving directional and spatial proximity information while reducing reliance on explicit coordinate encoding.The GCP module leverages geometric relationships between points to guide weighted channel interactions and residual fusion，thereby enhancing discrimination of complex structures and boundary regions. Experimental results show that the proposed method achieves superior performance across mainstream evaluation metrics，with particularly effective on challnging categories and small-scale targets. On the Stanford Large-Scale 3D Indoor Spaces （S3DIS） dataset， compared with RandLA-Net，our method improves mIoU from 70.0% to 76.2% and mAcc from 82.4% to （204号 83.4% . This study provides a scalable solution for efficient and accurate semantic segmentation of largescale point clouds，achieving higher overall accuracy while maintaining comparable inference eficiency and memory footprint.

Key words： point cloud semantic segmentation; spatial serialization; geometry-aware; 3D scene understanding

1引言

场景的高效建模[6-7]。（剩余14411字）