基于2D-3D语义传递的室内三维点云模型语义分割  

Semantic Segmentation of Indoor 3D Point Cloud Model Based on 2D-3D Semantic Transfer

作　　者：熊汉江[1] 郑先伟[1] 丁友丽 张艺 吴秀杰 周妍 XIONG Hanjiang;ZHENG Xianwei;DING Youli;ZHANG Yi;WU Xiujie;ZHOU Yan(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China;School of Mathematics and Statistics,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079 [2]武汉大学数学与统计学院,湖北武汉430079

出　　处：《武汉大学学报（信息科学版）》2018年第12期2303-2309,共7页Geomatics and Information Science of Wuhan University

基　　金：国家重点研发计划(2018YFB0505401);国家自然科学基金(41871361,41701445);测绘遥感信息工程国家重点实验室自主科研项目~~

摘　　要：针对现有三维点云模型重建对象化和结构化信息缺失的问题,提出一种基于图模型的二维图像语义到三维点云语义传递的算法。该算法利用扩展全卷积神经网络提取2D图像的室内空间布局和对象语义,基于以2D图像超像素和3D点云为结点构建融合图像间一致性和图像内一致性的图模型,实现2D语义到3D语义的传递。基于点云分类实验的结果表明,该方法能够得到精度较高的室内三维点云语义分类结果,点云分类的精度可达到73.875 2%,且分类效果较好。In this paper, we propose an effective algorithm based on graph model for semantic transfer from 2 D images to 3 D point clouds, which can effectively solve the problem of objectification and lack of structured information of 3 D point cloud model. Our proposed method uses the extended full convolutional neural network to extract the indoor space layout and object semantics of 2 D images, and then implements the transfer of 2 D semantics to 3 D semantics based on the 2 D image superpixels and 3 D point clouds as nodes to construct a graph model of consistency between images and intra-image consistency. The experiment from 3 D point cloud shows that the proposed method can obtain accurate indoor 3 D point cloud semantic classification results. The accuracy of point cloud classification can reach 73.875 2%, and the classification effect is better.

关 键 词：语义三维点云模型 语义传递 语义标记 点云分类 

分 类 号：P208[天文地球—地图制图学与地理信息工程]