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

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。

Aiming at the problem of matching distortion caused by structural occlusion,field of view constraints,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 information 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 obtain 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 registration 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 network can effectively solve the problem of low-overlap point cloud matching distortion.