特征自适应融合插值的点云语义分割算法  

Point cloud semantic segmentation based on adaptive⁃fusion feature interpolation

作　　者：朱芬芬 王蕾[1] 刘华 ZHU Fenfen;WANG Lei;LIU Hua(Department of Information Engineering,East China University of Technology,Nanchang 330013,China;School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013 [2]东华理工大学测绘工程学院,江西南昌330013

出　　处：《现代电子技术》2023年第12期175-181,共7页Modern Electronics Technique

基　　金：国家自然科学基金项目(61561003);国家自然科学基金项目(61761003);国家自然科学基金项目(62266001);国家自然科学基金项目(42001411);江西省核地学数据科学与系统工程技术研究中心基金项目(JETRCNGDSS201902;JELRGBDT202202);江西省自然科学基金项目(20202BABL212014);东华理工大学江西省数字国土重点实验室开放研究基金资助项目(DLLJ202004)。

摘　　要：点云语义分割是场景理解的基础问题。大多数针对大场景点云的语义分割方法精心设计局部特征聚合模块以减少降采样点云过程中的信息损失,然而未能高效地利用多尺度局部特征推理点云语义信息。为此,文中提出一种多尺度特征自适应融合的插值方法,以实现更准确的大场景点云语义分割算法。首先,通过注意力机制的局部特征聚合模块学习点云内部的语义关系,描述局部模式的各项异性;然后,以不同的采样率随机地采样点云产生多尺度的稀疏局部特征图;最后,使用特征自适应融合的插值法代替广泛使用的最近邻插值法,恢复全分辨率的特征图,为原始密度的点云提供更准确的语义信息。在SemanticKITTI和S3DIS两个大场景点云数据集上对所提算法进行评估,结果表明,所提算法的平均交并比(mIoU)分别为54.24%、75.5%,平均准确率(mACC)分别达到88.92%、86.5%,比大多数主流算法的分割效果更加准确。Semantic segmentation of point cloud is a fundamental problem in scene⁃level understanding.In the most of the semantic segmentation methods for large⁃scale scenic spot clouds,local feature aggregation modules are carefully designed to reduce the information loss in the process of downsampling point clouds,but multi⁃scale local features are not used efficiently infer point cloud semantic information.A multi⁃scale feature adaptive fusion interpolation method is proposed to achieve a more accurate semantic segmentation algorithm for large⁃scale scenic spot clouds.The local feature aggregation module of the attention mechanism is used to learn semantic relationships within the point cloud and describe the anisotropy of local patterns.Randomly sampling point clouds at different sampling rates can generate multi⁃scale sparse local feature maps.The interpolation method of feature adaptive fusion is used to instead of the widely used nearest neighbor interpolation method to restore the full resolution feature map and provide more accurate semantic information for the original density point cloud.The proposed algorithm is evaluated on two large⁃scale scenic spot cloud datasets,SemanticKITTI and S3DIS.The results show that the mean intersection of union(mIoU)of the proposed algorithm is 54.24%and 75.5%,respectively,and the mean accuracy(mACC)is 88.92%and 86.5%,respectively,which is more accurate than most mainstream algorithms in segmentation.

关 键 词：点云语义分割 注意力机制 自适应融合插值 局部特征聚合模块 深度学习 随机降采样 

分 类 号：TN919-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]