基于自适应分层采样局部-非局部编码点云分类网络  

Point AHLN:Point Clouds Classification Using Local-non-local Neural Networks with Adaptive Hierarchical Sampling

作　　者：邱阳 王旭智[1,2] 万旺根[1,2] Qiu Yang

机构地区：[1]上海大学通信与信息工程学院,上海200072 [2]上海大学智慧城市研究院,上海200072

出　　处：《工业控制计算机》2022年第9期80-83,共4页Industrial Control Computer

基　　金：安徽省自然科学基金(1908085MF178);安徽省重点研究和开发计划项目(202104b11020031);中国博士后基金项目(2020M681264)。

摘　　要：点云分类是点云数据处理的一个重要研究方向,也是一个具有挑战性的课题。原始点云中不可避免地包含大量对全局影响很小的点以及会对模型产生恶劣影响的离群点,提出了一种用于点云分类的新型网络,它可以有效地处理模型中带有的异常值和带有噪音的点云,并且能够尽量避免采样到对全局没有贡献的点,该方法的关键就是自适应分层点云下采样和局部-非局部模块。自适应分层采样首先得到每个点对全局的重要性程度,当选取下一个采样点时会综合考虑点的重要程度以及点空间之间的距离,这不仅能够得到比较均匀的采样模型还能够选取出对模型较重要的点。为了进一步捕获采样点邻域的依赖关系,提出了一个局部-非局部(L-NL)模块,这种L-NL模块使得网络在学习过程中对噪声不太敏感。大量实验验证了该方法在点云处理任务中的鲁棒性和优越性,该网络在公开数据集上取得了较好的分类性能。Point cloud classification is an important research direction of point cloud data processing, and it is also a challenging subject.The original point cloud inevitably contains a large number of points that have little impact on the global and outliers that will have a bad impact on the model.A new type of end-to-end point cloud classification network is proposed in this paper.It can effectively deal with outliers and noisy point clouds,and can avoid sampling to insignificant points.The key to our method is adaptive hierarchical point cloud down sampling and local-non-local modules.Adaptive hierarchical sampling first obtains the global importance of each point.When the next sampling point is selected,the importance of the point and the distance between the point spaces are comprehensively considered.This not only can obtain a more uniform sampling model,but also pick points that are more important to the model.In order to further capture the dependency of the sampling point neighborhood,this paper proposes a local-non-local(L-NL) module.This L-NL module makes the network less sensitive to noise during the learning process.A large number of experiments have verified the robustness and superiority of this method in point cloud processing tasks.This network has achieved good classification performance on public data sets.

关 键 词：点云分类 重要性程度 自适应分层采样 局部-非局部 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]