例如,Yue Wang 和 Solomon 提出了一种基于深度学习的配准算法,基本流程是首先用动态图卷积神经网络(DGCNN)将未对齐的点云数据嵌入(embed)到一个共同的空间之中,然后用一个基于注意力的模块结合指针网络(pointer network)预测两片点云的近似匹配,最后用一个奇异值分解模块提取刚体变换,得到最终的结果,参见《Deep Closest Point: Learning Representations for Point Cloud Registration》。
同时配准和重建(simultaneous registration and reconstruction):
假设输入的数据基本对准了,算法的基本思想是把空间划分成预定义分辨率的网格,在每个网格内部根据扫描数据拟合出潜在的曲面(latent surface),然后在各个网格中配准扫描数据与潜在曲面,再又优化潜在曲面,如此交替地优化,最终可配准所有的扫描数据,并且也得到了重建的曲面,参见《High Quality Pose Estimation by Aligning Multiple Scans to a Latent Map》;
参见《Registration of 3D point clouds and meshes: A survey from rigid to nonrigid》。 阅读材料
《A Method for Registration of 3-D Shapes》Paul J. Besl and Neil D. McKay. IEEE Transactions on Pattern Recognition and Machine Intelligence. 1992.
《Efficient Variants of the ICP Algorithm》Szymon Rusinkiewicz and Marc Levoy. 2001.
《Geometry and Convergence Analysis of Algorithms for Registration of 3D Shapes》Helmut Pottmann, Qixing Huang, Yongliang Yang, and Shimin Hu. International Journal of Computer Vision. 2006.
《High Quality Pose Estimation by Aligning Multiple Scans to a Latent Map》Qixing Huang and Dragomir Anguelov. IEEE International Conference on Robotics and Automation 2010.
《DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-Time》Richard Newcombe, Dieter Fox, and Steve Seitz. CVPR 2015.