Abstract:

A point cloud data representation method based on the octree tree and K-D tree is proposed to realize end to end representation point cloud model in deep learning. This method transforms the unorganized point cloud into a voxel space, and uses the octree model to divide and improves octree coding method. The adjacency relationship between nodes is constructed, and the octree is constructed in parallel on the GPU. A 3D K-D tree is used to index a single 3D spatial point in order to overcome the inefficiency of octree coding retrieval. Experimental results show that the method can truly reflect the details of the model itself and improve the construction time and retrieval efficiency of point cloud model. The new data structure transforms the point cloud into a form of data that can be received by convolutional neural network.

Key words: point cloud data, representation method, deep learning, octreetree, K-D tree

摘要：

为了实现在深度学习中能够端到端表示点云模型，提出基于八叉树和K-D树（OctKD）的点云数据表示方法。该方法将无组织的点云转换为体素空间，在体素空间对三维模型进行八叉树剖分，改进了八叉树编码方式；构建节点间的邻接关系，在GPU端并行构建八叉树；为了克服八叉树编码检索效率低的问题，采用三维K-D树索引单个三维空间点。实验结果表明该方法能够真实反映模型本身的细节特征，提高了点云模型的构造时间和检索效率。这种新的数据结构实现将点云转换为卷积神经网络可以接收的数据形式。

关键词: 点云数据, 表示方法, 深度学习, 八插树, K-D树