面向低重叠率点云匹配的多层级过滤网络
Multi-level filter network for low-overlap point cloud registration
贺敏琦 1刘俐 1李尚 1吴浩 1朱大虎1
作者信息
- 1. 武汉理工大学 现代汽车零部件技术湖北省重点实验室,湖北 武汉 430070;武汉理工大学 汽车零部件技术湖北省协同创新中心,湖北 武汉 430070
- 折叠
摘要
针对点云测量过程中由于结构遮挡、视场约束、拼接误差等导致的匹配失真问题,提出一种多层级过滤网络(Multi-level Filter Network,MulFNet)用于实现单次测量点云低重叠率下的精确配准.通过特征金字塔编码网络提取点云的多层级特征,获得不同尺度的语义信息,同时嵌入注意力模块和信息编码模块以增强特征显著性.基于多尺度一致性决策机制对多层级特征进行过滤,筛选离群点并保留点云突出特征,获得初始对应关系.最后,将初始对应结点基于几何信息自适应分组,由局部至全局进行加权转换估计,获得基于多层级过滤筛选后的预测矩阵.实验结果表明,MulFNet网络在标准 3DMatch公共数据集上的匹配效果明显优于FCGF,PREDATOR等主流网络,在平均重叠率为10%的测量数据集上的匹配精度比ICP算法和GeoTransformer网络分别提高40.9%和85.4%,有效解决了低重叠率点云匹配失真的问题.
Abstract
Aiming at the problem of matching distortion caused by structural occlusion,field of view con-straints,and stitching errors during point cloud reconstructed,a multi-level filter network(MulFNet)is proposed to achieve single-shot scanning point clouds for low-overlap registration.Firstly,the multi-level features of the point clouds are extracted through the feature pyramid coding network to obtain semantic in-formation at different scales,and the attention module and the location module are embedded to enhance the feature significance;secondly,the multi-level features are filtered based on the multi-scale consistency voting mechanism,outliers are screened out and prominent features of the point clouds are retained to ob-tain the initial correspondence;and finally,the initial corresponding nodes are adaptively grouped based on the geometric relationships,and weighted estimation conversion is performed from local to global to obtain a prediction matrix based on the multi-level filtering.The experimental results show that the MulFNet is better than the popular networks such as FCGF and PREDATOR on the standard 3DMatch.The registra-tion accuracy of the MulFNet on the scanning dataset with an average overlap rate of 10%is 40.9%and 85.4%higher than the ICP and the GeoTransformer,respectively.It is verified that the proposed net-work can effectively solve the problem of low-overlap point cloud matching distortion.
关键词
点云匹配/匹配失真/低重叠率/多层级过滤/局部测量Key words
point cloud registration/matching distortion/low-overlap point cloud/multi-level filter/par-tial measurement引用本文复制引用
基金项目
国家自然科学基金资助项目(52375509)
湖北省重点研发计划资助项目(2022BAA067)
出版年
2024
国家科技期刊平台
NSTL
万方数据