Liu Dengzhi, Zhang Yu, Luo Lin, Li Jinlong, Gao Xiaorong
Appl Opt. 2021 Apr 10;60(11):2990-2997. doi: 10.1364/AO.418304.
It is important to improve the registration precision and speed in the process of registration. In order to solve this problem, we proposed a robust point cloud registration method based on deep learning, called PDC-Net, using a principal component analysis based adjustment network that quickly adjusts the initial position between two slices of the point cloud, then using an iterative neural network based on the inverse compositional algorithm to complete the final registration transformation. We compare it on the ModelNet40 dataset with iterative closest point, which is the traditional point cloud registration method, and the learning-based methods including PointNet-LK and deep closest point. The experimental results show that the registration error is not worse with the increase of the initial phase between point clouds, avoiding the algorithm falling into the local optimal solution and enhancing the robustness of registration.
在配准过程中提高配准精度和速度很重要。为了解决这个问题，我们提出了一种基于深度学习的鲁棒点云配准方法，称为PDC-Net，它使用基于主成分分析的调整网络快速调整两片点云之间的初始位置，然后使用基于逆合成算法的迭代神经网络完成最终的配准变换。我们在ModelNet40数据集上，将其与传统点云配准方法迭代最近点以及包括PointNet-LK和深度最近点在内的基于学习的方法进行比较。实验结果表明，随着点云之间初始相位的增加，配准误差不会变差，避免了算法陷入局部最优解，增强了配准的鲁棒性。
