基于深度学习的大规模三维点云处理综述①王振燕,孙红岩,孙晓鹏(辽宁师范大学计算机与信息技术学院,大连116029)通信作者:孙红岩,E-mail:3256153627@qq.com摘要:随着三维视觉的快速发展,基于深度学习的大规模三维点云实时处理成为研究热点.以三维空间分布无序的大规模三维点云为背景,综合分析介绍并对比深度学习实时处理三维视觉问题的最新进展,对点云分割、形状分类、目标检测等方面算法优势与不足进行详细分析,给出详细的性能分析与优劣对比,并对点云常用数据集进行简要介绍,并给出不同数据集的算法性能对比.最后,指出未来在基于深度学习方法处理三维点云问题上的研究方向.关键词:深度学习;目标检测;目标追踪;形状分类;点云分割引用格式:王振燕,孙红岩,孙晓鹏.基于深度学习的大规模三维点云处理综述.计算机系统应用,2023,32(2):1–12.http://www.c-s-a.org.cn/1003-3254/8743.htmlSurveyonLargeScale3DPointCloudProcessingUsingDeepLearningWANGZhen-Yan,SUNHong-Yan,SUNXiao-Peng(DepartmentofComputerandInformationTechnology,LiaoningNormalUniversity,Dalian116029,China)Abstract:Withtherapiddevelopmentof3Dvision,large-scale3Dpointcloudprocessinginrealtimebasedondeeplearninghasbecomearesearchhotspot.Takingalarge-scale3Dpointcloudwithdisorderedspatialdistributionasthebackground,thisstudycomprehensivelyanalyzes,introducesandcomparesthelatestprogressofdeeplearninginreal-timeprocessingof3Dvisionproblems.Then,itanalyzesindetailandcomparestheadvantagesanddisadvantagesofalgorithmsintermsofpointcloudsegmentation,shapeclassificationandtargetdetection.Further,itbrieflyintroducesthecommondatasetsofpointcloudsandcomparesthealgorithmperformanceofdifferentdatasets.Finally,thestudypointsoutthefutureresearchdirectionof3Dpointcloudprocessingbasedondeeplearning.Keywords:deeplearning(DL);targetdetection;targettracking;shapeclassification;pointcloudsegmentation深度学习(deeplearning,DL)指基于数据预测、并改进其预测结果或行为的方法[1],训练阶段以最小化损失函数为引导,通过梯度下降调整计算模型的权重和偏置;测试阶段以输入数据和训练好的模型参数计算预测值[1],广泛应用于二维目标检测、分割和分类[2,3]等领域.随着三维扫描仪、激光雷达、RGB-D相机(如Kinect、RealSense和Apple深度...
