摘要
针对目前的点云分类是直接将原始点云作为输入并提前预设点云分类数存在的缺陷,本文提出一种改进的方法,在输入前对原始点云进行预处理,对密集的点云降低密度以减少计算量,对稀疏的点云进行三角形内部线性插值以便提取完整的特征,以此提高点云分类的精度。将预处理后的点云数据输入SOM-K(K-Means优化的自组织映射神经网络)模型进行聚类,再将聚类后的点云数据并行通过PointNet网络进行点云数据特征的提取,这种先进行聚类后、进行特征提取的方法可以充分保留点云在点云空间中的分布特性,并且不额外增加数据特侦提取的计算时间。
For the current point cloud classification, the original point cloud is directly used as input and the number of point cloud classifications is preset in advance. This paper proposes an improved method to preprocess the original point cloud before input, and lower the density of the dense point cloud and reduce the amount of calculation, and the sparse point cloud is linearly interpolated inside the triangle in order to extract the complete features, thereby improving the accuracy of point cloud classification. The preprocessed point cloud data is input into SOM-K(K-Means optimized self-organizing map neural network) model for clustering, then the point cloud data after clustering is extracted through the PointNet network in parallel to extract the features of the point cloud data. This method of first clustering and then feature extraction can fully retain the distribution characteristics of the point cloud in the point cloud space, and could not increase the calculation time for special detection and extraction of data.
作者
邬春学
胡真豪
WU Chunxue;HU Zhenhao(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《智能计算机与应用》
2022年第11期172-179,共8页
Intelligent Computer and Applications
作者简介
邬春学(1964-),男,博士,教授,博士生导师,主要研究方向:图像分类、智能家居、物联网技术及应用;胡真豪(1996-),男,硕士研究生,主要研究方向:图像分类。
