基于点云卷积的点云分类方法研究

A Point Cloud Classification Method Based on Point Cloud Convolution

基于深度学习方法,借鉴二维图像卷积的思想,设计了一种适合三维点云的卷积操作。点云卷积的作用域是局部球形邻域,输入为三维坐标和空间几何关系。通过点云卷积提取局部特征,使用最远点采样算法采集邻域中心点,根据半径构建球形局部邻域,使用多层感知器(multi-layer perceptron,MLP)网络学习空间关系权重,将学习到的关系权重和输入特征相乘,实现卷积操作。基于三维点云卷积,构建了一个多层分类网络模型实现点云分类。使用道路场景的黄石路数据集进行分类实验,结果证明了所提方法的有效性。

Based on deep learning methods,we design a con⁃volution operation suitable for 3D point clouds by referring to the idea of 2D image convolution.The scope of point cloud convolution is a local spherical neighborhood,and its inputs are 3D coordinates and spatial geometric relations.Local fea⁃tures are extracted by point cloud convolution,and the farthest point sampling algorithm is used to collect the neighborhood center points.The spherical local neighborhood is constructed according to the radius,and the multi-layer perceptron(MLP)is used to learn the spatial relation weights.We multiply the learned relation weights and the input features to achieve the convolution operation.Based on 3D point cloud convolution,we construct a multi-layer classification network model to realize point cloud classification task.A classification experi⁃ment on Huangshi Road dataset of the road scene verifies the effectiveness of the proposed method.