核相关神经网络点云自动配准算法

Automatic Point Cloud Registration Algorithm Based on Kernel Correlation Neural Network

摘要:

点云配准是点云数据智能处理的重要问题，也是将点云应用于智慧城市、自动驾驶和智能三维重建等方面的关键。针对现有点云配准方法效率低、鲁棒性差的问题，提出了一种基于核相关神经网络的点云自动配准算法。首先构建点云核用于计算点云中每个点的核相关度，然后通过多层感知机对点云进行特征编码，基于编码特征向量估计点间对应关系并求解变换参数，最后以迭代方式来使待配准点云不断逼近目标点云，完成点云配准。使用斯坦福大学3D扫描模型库中的Bunny、Dragon、Happy、Elephant、Horse点云数据，对该算法以及迭代最近邻点算法（ICP）等多个算法进行对比实验。实验结果表明，所提算法能够对不同物体点云实现精确配准，精度和效率均优于所对比算法，且在点云数据存在噪声和密度不一致的情况下仍具有良好的稳定性和精度。

Abstract:

Point cloud registration is an important issue in the intelligent processing of point cloud data. It is also the key to applying point clouds in smart cities， autonomous driving， and 3D reconstruction. To solve the problem of low efficiency and poor robustness of existing point cloud registration methods， an automatic point cloud registration algorithm based on kernel correlation neural network is proposed. First kernel correlation of each point in point cloud is calculated by using several point kernels. Then the point cloud information is coded through multi-layer perceptron， and point correspondence is generated based on the coded feature vector to calculate the transformation parameters. Finally， the registration result is optimized through iteration. Using the Bunny， Dragon， Happy， Elephant， Horse in the 3D scanning model library of Stanford University， the algorithm proposed in this paper and other algorithms such as iterative closest point （ICP） are compared. The experiment shows that the algorithm proposed can achieve an accurate registration of different object point clouds， with a better accuracy and efficiency. It has a good stability even in the presence of noise and different density in the point cloud data.