基于二阶相似性测度的点云配准网络

Point Cloud Registration Network Based on Second-order Similarity Measure

为了提高点云配准在异常点移除以及位姿变换矩阵求取时的精度,设计了一种点云配准方法,在PointDSC网络的基础上,将二阶相似性测度引入到特征聚合模块,嵌入到匹配点对的特征之中,提高了特征提取时内点与异常点的初始特征差异;引入二阶兼容性矩阵到谱匹配模块中,改进了相容性矩阵的求取,通过嵌入的特征去估计特征点对是内点还是异常点的置信度,提高了异常点与内点的区分精度,更好地移除异常点;利用局部-全局的配准方法根据内点求解点云的位姿变换矩阵,提高了位姿变换矩阵的求取速度及精度。在3DMatch数据集上进行了配准实验,验证了提出的方法在多种不同的数据集场景下都能够很好地实现点云配准任务,提高了配准的速度与精度。

In order to improve the accuracy of point cloud registration when removing anomalous points and finding pose transformation matrix,a point cloud registration method is designed,and on the basis of PointDSC network,the second-order similarity measure is introduced into the feature aggregation module and embedded in the features of matching point pairs,which improves the initial feature difference between internal points and abnormal points during feature extraction.The second-order compatibility matrix is introduced into the spectral matching module,which improves the finding of the compatibility matrix,and estimates the confidence of whether the feature point pair is an internal point or an anomaly point through the embedded features,which improves the discrimination accuracy between the abnormal point and the inner point and better removes the abnormal point.The local-global registration method is used to solve the pose transformation matrix of the point cloud according to the internal point,which improves the detection speed and accuracy of the pose transformation matrix.Registration experiments are carried out on the 3DMathch dataset.