基于密度自适应深度网络的点云场景语义分割算法

Point Cloud Scene Semantic Segmentation Algorithm Based on Density Adaptive Deep Network

摘要:

由于传感器噪声干扰，点云密度不均匀，场景复杂多样以及物体之间存在遮挡现象等问题，使得三维点云场景语义分割问题的研究工作极具挑战性。针对三维点云数据采样密度不均匀以及图卷积网络深度有限的问题，提出一种密度自适应的方法。该方法通过多层感知器学习一个权重函数，利用核密度估计学习一个密度函数，对非均匀采样的点云数据进行卷积操作。同时，受深度学习在图像领域的启发，引入残差连接、空洞卷积等结构，训练更深层的点云分割网络。该算法在多个点云分割的标准数据集上取得了优秀的性能。

Abstract:

Sensor noise interference， uneven point cloud density， complex and diverse scenes， and the occlusion between objects pose great challenges to the research of the semantic segmentation of 3D point cloud scenes. In the view of the uneven sampling density of 3D point cloud data and limited depth of graph convolutional networks， this paper proposes a density adaptive method， which uses a multi-layer perceptron to learn a weight function， and uses kernel density estimation to learn a density function. The convolution operation is performed on the evenly sampled point cloud data. At the same time， inspired by deep learning in the image field， residual connection， hole convolution， and other structures were introduced to train deeper point cloud segmentation networks. The algorithm achieves an excellent performance on standard data sets of multiple point cloud segmentation.