摘要
点云是目前自动驾驶、机器人、遥感、增强现实(AR)、虚拟现实(VR)、电力、建筑等领域最常用的三维数据处理形式,深度学习方法能够处理大型数据,且可自主提取特征,因此点云深度学习方法已逐渐成为研究热点。本文综述了近十年来基于深度学习的三维点云分析方法的研究进展。首先给出了三维点云深度学习的相关概念;然后针对点云目标检测与跟踪、分类分割、配准和匹配以及拼接这4种任务,分别阐述了相应的深度学习方法的原理,分析并比较了各自的优缺点;随后整理了18种点云数据集和4种点云分析任务的性能评价指标,并给出了性能对比结果;最后总结了点云分析方法目前存在的问题,对进一步的研究工作进行了展望。
Point cloud is the most commonly used form of 3D data processing in the fields of autonomous driving,robotics,remote sensing,augmented reality(AR),virtual reality(VR),electric power,architecture,etc.Deep learning methods can not only handle large-scale data,but also extract features independently.Therefore,point cloud deep learning methods have gradually become a research hotspot.This article reviews the research progress of 3D point cloud analysis methods based on deep learning in the past decade.Firstly,the relevant concepts of deep learning for 3D point cloud are presented.Then,for the four tasks of point cloud object detection and tracking,classification and segmentation,registration and matching,and stitching,the principles of the corresponding deep learning methods are elaborated.Their advantages and disadvantages are analyzed and compared.Next,eighteen kinds of point cloud datasets and performance evaluation indexes for four types of point cloud analysis tasks are introduced.The performance comparison results are given.Finally,the existing problems of point cloud analysis methods are pointed out,and the further research work is prospected.
作者
陈慧娴
吴一全
张耀
Chen Huixian;Wu Yiquan;Zhang Yao(College of Electronic and Information Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2023年第11期130-158,共29页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(61573183)项目资助
关键词
三维点云分析
深度学习
目标检测
分类分割
点云配准
点云拼接
3D point cloud analysis
deep learning
object detection
classification and segmentation
point cloud registration
point cloud stitching
作者简介
陈慧娴,2021年于辽宁大学获得学士学位,现为南京航空航天大学硕士研究生,主要研究方向为图像处理与机器视觉。E-mail:chenhuixian2017@163.com;通信作者:吴一全,1998年于南京航空航天大学获得博士学位,现为南京航空航天大学教授、博士生导师,主要研究方向为视觉检测与图像测量、遥感图像处理与理解、红外目标检测与识别、视频处理与智能分析等。E-mail:nuaaimage@163.com
