基于边缘感知的点云配准算法

Edge-Aware Point Cloud Registration

在使用深度学习方法进行点云配准时,直接利用特征相似性作为采样依据,往往会导致采样过于集中且大量分布在平面等非显著区域内,不利于变换矩阵的推导.针对此问题,提出一种边缘感知的点云配准算法.首先通过分析点云中每点与其邻域点的空间分布对边缘区域进行检测;然后针对现有的特征描述子和联合学习框架,将对应关系和关键点的采样区域限定在边缘区域,提高特征的匹配能力;最后将特征相似性和显著性作为采样概率,得到一组分布良好的对应关系或关键点并用于配准.在真实数据集和合成数据集上的大量实验结果表明,所提算法可以使现有的特征描述符达到与现有联合学习框架相当的性能,对于现有联合学习框架,在低重叠点云场景(3DLoMatch)中,边缘区域采样关键点可以平均提高约5%的配准召回率.

When using deep learning methods for point cloud registration,directly using feature similarity as the sampling basis,the samples will often be too concentrated and most of them distributed in the non-salient region such as plane,which is not conducive to the estimation of the transformation matrix.To solve this problem,an edge-aware point cloud registration algorithm is proposed.Firstly,the edge region is detected by analyzing the spatial distribution of each point in the point cloud and its neighbors.Secondly,according to the existing feature descriptors and joint learning frameworks,the sampling areas of correspondences and keypoints are limited to the edge region,so as to improve the ability of feature matching.Finally,a probabilistic sampling method based on feature similarity and saliency is used to obtain a set of well-distributed correspondences or keypoints for registration.A large number of experiments on real and synthetic datasets show that our method can make feature descriptors achieve comparable performance to existing joint learning frameworks,for the latter,sampling keypoints in the edge region can increase the registration recall by an average of 5%in the low overlap scenarios(3DLoMatch).