融合图注意力的摄影测量点云语义分割研究  

Research on Semantic Segmentation of Photogrammetric Point Clouds with Image Attention

作　　者：徐俊[1] 杜宣萱 宋俊锋[2] 陆佳炜[1] 程振波[1] 肖刚[1] XU Jun;DU Xuan-xuan;SONG Jun-feng;LU Jia-wei;CHENG Zhen-bo;XIAO Gang(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;College of Engineering,Lishui University,Lishui 323000,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]丽水学院工学院,浙江丽水323000

出　　处：《小型微型计算机系统》2022年第7期1464-1470,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61976193)资助;浙江省重点研发计划项目(2021C03136)资助;浙江省自然科学基金项目(LY19F020034)资助.

摘　　要：随着无人机倾斜摄影测量技术的发展,通过密集影像匹配可以快速获得类比激光扫描数据精度的大规模室外点云,但是这些点云存在着不规则、遮挡严重、数据量庞大的特点,同时因为缺乏对象信息无法深入进行语义分析.针对上述问题,本文提出一种融合图注意力的摄影测量点云语义分割方法.首先构建了一种新的图卷积模块,在网络的每一层动态的更新点云局部邻域图,将跨层点描述与上下文特征结合起来并逐层汇聚点云空间潜在语义信息;然后在每个网络层引入通道注意力机制使网络能够自适应学习通道间的权重,并由此建立基于一种新的图注意模块的点云语义分割网络,实现复杂点云的细粒度语义分割.通过在两个公开的室外点云基准数据集上的实验结果表明,该方法能够显著提升网络对局部拓扑特征信息的学习能力,且对复杂场景点云语义分割具有良好的泛化能力.With the development of UAV tilt photogrammetry technology,dense image matching can quickly obtain large-scale outdoor point clouds with analogue laser scanning data accuracy,but these point cloud have the characteristics of irregularity,serious occlusion,and huge data volume.Because of the lack of object information,it is impossible to conduct in-depth semantic analysis.To solve the above problems,this paper proposes a method for semantic segmentation of photogrammetric point clouds based on graph attention.First,a new graph convolution module is constructed,which dynamically updates the local neighborhood graph of the point cloud at each layer of the network,combines the cross-layer point description with the context feature,and gathers the potential semantic information of the point cloud layer by layer.Then the channel attention mechanism is introduced to each network layer,so that the network can adaptively learn the weights between channels.Finally,a point cloud semantic segmentation network based on a new graph attention module is established to achieve fine-grained complex point clouds Semantic segmentation.Experimental results on two public outdoor point cloud benchmark data sets show that this method can significantly improve the network′s learning ability of local topological feature information,and has good generalization ability for point cloud semantic segmentation in complex scenes.

关 键 词：摄影测量点云 局部邻域图 通道注意池 语义分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]