Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (13): 217-226.DOI: 10.3778/j.issn.1002-8331.2011-0388

• Graphics and Image Processing • Previous Articles     Next Articles

Multi-Scale Hierarchical Point Clouds Recognition Network Based on Weight Index of Point Density

QIN Xinyu, HAN Shuai, SHEN Xueli, YANG Ying   

  1. 1.School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China 
    2.Quanzhou Equipment Manufacturing Research Institute, Haixi Research Institute, Chinese Academy of Sciences, Quanzhou, Fujian 362216, China
  • Online:2022-07-01 Published:2022-07-01

基于点密度加权的多尺度分层点云识别网络

秦鑫宇,韩帅,沈学利,杨莹   

  1. 1.辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
    2.中国科学院 海西研究院 泉州装备制造研究所,福建 泉州 362216

Abstract: Unlike dense and regularly distributed 2D raster images, 3D point clouds are irregular and disordered, thus applying convolution on them can be difficult. A method for direct convolution of original 3D point clouds is proposed. This method uses Gaussian kernel density estimation and multi-layer perceptron(MLP) network to learn the density function. Besides, the density scale of the learning point and the relative position of the point are fitted by the weight function of the approximated MLP network to obtain the weight value of each point in the local area. The whole convolution kernel can be seen as a nonlinear function of 3D local coordinates composed of the weight function and the density function, which can be used for translation invariant and permutation invariant convolution of any set of points in 3D space. In addition, multi-scale sampling grouping and normal features are employed to achieve the best effect of this network. In the classification experiment of point cloud datasets of ModelNet40 and ModelNet10, the accuracy of the network is 92.8% and 94.7% respectively, which acquires better performance than baseline. The image datasets of CIFAR-10 and MNIST are converted into point clouds to test, experiments on aforementioned datasets show that the performance of the network in 2D images is basically equivalent to that of the traditional 2D convolutional network.

Key words: image processing, 3D point cloud, kernel density estimation, convolution computation, multi-scale features, normal features

摘要: 与密集且规则分布的2D栅格状图像不同,3D点云是不规则且无序的,对其进行卷积可能会存在一定的困难,因此,提出了一种针对原始3D点云的卷积运算。该方法使用高斯核密度估计和多层感知器(multi-layer perceptron,MLP)网络来学习密度函数,将学习到的点的密度尺度结合点的相对位置,通过由MLP网络近似的权重函数之中,得到局域中每个点的权重值。整个卷积核可视为由权重函数和密度函数组成的3D点局域坐标的非线性函数,可用于对3D空间中任意点集进行平移不变和置换不变的卷积,并融合多尺度采样分组和法向特征使网络达到最佳效果。在ModelNet40和ModelNet10数据集的分类实验中,该网络分别取得了92.8%和94.7%的准确率,均高于所对比的同类方法的性能水平。将CIFAR-10和MNIST图像数据集转为点云并进行测试,结果表明网络在2D图像中的性能基本等效于传统2D卷积网络。

关键词: 图像处理, 3D点云, 核密度估计, 卷积运算, 多尺度特征, 法向特征