一种室内确定目标的点云图像分割方法
A Point Cloud Image Segmentation Method for Indoor Determining Objects
颜玉祥 1徐晓龙 1张辰旭 1张卓 1邵晓琦1
作者信息
- 1. 河海大学物联网工程学院 常州 213022
- 折叠
摘要
针对室内确定目标点云图像分割应用场景,在语义分割层面,改进了PointNet++网络,为提高下采样的精度和效率,用重要性采样代替了最远点采样;为提升点云稀疏区域的特征提取效果,加入了包围球特征提取的阈值.在实例分割层面,为有效分割同类的多个实例重叠交错的点云图像,在DBSCAN粗聚类的基础上,使用最小割算法对每个实例进行逐一分离.实验结果表明,在语义分割层面,论文算法在分割准确率、训练耗时、点云稀疏区域的分割表现上均优于PointNet++;在实例分割层面,论文算法的总精确度比Jsnet网络提高了 6.04%.
Abstract
Aiming at the application scenario of determining the target point cloud image segmentation in the room,at the se-mantic segmentation level,the PointNet++network is improved.In order to improve the accuracy and efficiency of down sampling,importance sampling is used to replace the farthest point sampling.To improve the feature extraction of the sparse area of the point cloud,threshold for feature extraction of the bounding ball is added.At the instance segmentation level,in order to effectively seg-ment the overlapping and interlaced point cloud images of multiple instances of the same type,on the basis of DBSCAN coarse clus-tering,each instance is separated one by one using the minimum cut algorithm.Experimental results show that at the level of seman-tic segmentation,the algorithm in this paper is superior to PointNet++in segmentation accuracy,training time-consuming,and seg-mentation performance in sparse point cloud regions.At the instance segmentation level,the total accuracy of the algorithm in this paper is 6.04%higher than that of the Jsnet network.
关键词
室内确定目标/点云分割/PointNet++/最小割Key words
indoor determining objects/point cloud segmentation/PointNet++/minimum cut引用本文复制引用
基金项目
国家重点研发计划(2018YFC0407101)
国家自然科学基金项目(61671202)
出版年
2024
国家科技期刊平台
NSTL
万方数据