针对近几年三维视觉领域基于深度学习点云配准算法鲁棒性差和精度较差等问题，设计了一种基于深度学习的三维点云配准方法。首先抽取具有明显几何特征的点作为兴趣点，通过区域生长算法对兴趣点进行聚合，并基于多尺度分析方法进行特征提取以及特征融合。为进一步提取特征数据中包含的深层局部信息，使用多层感知机(multilayer perceptron，MLP)网络进行二次特征提取，并加入修改过的Transformer网络补充特征。设计了匹配矩阵生成及优化算法，并通过奇异值分解(singular value decomposition，SVD)计算得到变换矩阵。通过在ModelNet40 数据集上进行比较实验，证明本文的配准算法远优于传统配准算法，并在配准精度和鲁棒性方面优于近几年流行的深度配准网络DCPNet和RPMNet。本文分析结果可为提高点云配准鲁棒性以及精度提供参考。
Multi-scale deep learning method for 3D point cloud registration
In order to solve the problems of poor robustness and poor accuracy of the point cloud registration algorithm based on deep learning in the field of 3D vision in recent years,a 3D point cloud registration method based on deep learning is designed.Firstly,the points with obvious geometric features are extracted as points of interest,and the points of interest are aggregated by the regional growth algorithm,and feature extraction and feature fusion are carried out based on the multi-scale analysis method.In order to further extract the deep local information contained in the feature data,the multilayer perceptron(MLP)network is used for secondary feature extraction,and the modified Transformer network is added to supplement the features.The matching matrix generation and optimization algorithm is designed,and the transformation matrix is obtained by singular value decomposition(SVD)calculation.Through comparative ex-periments on the ModelNet40 dataset,it is proved that the registration algorithm in this paper is far superior to the tradi-tional registration algorithm,and is better than the popular deep registration networks DCPNet and RPMNet in recent years in terms of registration accuracy and robustness.The analysis results in this paper provide a reference for improv-ing the robustness and accuracy of point cloud registration.
point cloud registration3D visionfeature extractionfeature fusiondeep learningpoint of interest aggreg-ationmulti-scale featuresTransformer network
万方数据
维普
