基于全局上下文注意力的点云语义分割  

Point cloud semantic segmentation based on global contextual attention

作　　者：侯伟鹏 王蕾[1] HOU Weipeng;WANG Lei(Institute of Information 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

出　　处：《现代电子技术》2023年第9期120-125,共6页Modern Electronics Technique

基　　金：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(JELRGBDT202202)。

摘　　要：点云语义分割是三维环境感知的基础,直接基于点的语义分割方法避免了因点云结构化处理所造成的信息损失,但大多数深度学习模型的研究主要集中在提取局部几何特征,没有考虑点云不同局部结构之间的上下文关系,并且忽略了低级与高级特征之间的语义差距,限制了特征表示的能力,影响了点云语义分割的精度。因此,文中提出一种基于全局上下文注意力的点云语义分割方法,该方法主要由基于外部注意力的全局上下文特征聚合和基于后向竞争性注意力的邻近尺度特征融合两部分组成。通过外部注意力学习不同局部结构之间的长距离依赖关系,从而获得丰富的全局上下文信息。为了进一步增强模型的上下文感知能力,设计基于后向竞争性注意力的邻近尺度特征融合模块,学习低级与高级语义特征之间的相似度,重新为中间特征通道分配权重。在S3DIS大规模室内点云数据集上对所提方法进行评估,结果表明,所提方法的平均交并比在Area5上达到了65.2%,相比于RandLA-Net提高了2.5%,在6折交叉验证上的平均交并比达到了71.4%,相比于RandLA-Net提高了1.4%。证明了所提方法能够有效提取全局上下文特征,提高了语义分割的精度。Point cloud semantic segmentation is the basis of 3D environment perception.The direct point⁃based semantic segmentation method prevents information loss brought on by structured processing of point clouds,but the majority of deep learning model studies primarily concentrate on extracting local geometric features without taking into account the contextual relationship between various local structures of point clouds and fail to take into account the semantic gap between low⁃level and high⁃level features,which restricts the capability of feature representation and affects the accuracy.A point cloud semantic segmentation based on global contextual attention is proposed.This method is mainly composed of two parts:global context feature aggregation based on external attention and neighboring scale feature fusion based on backward competitive attention.By learning the long⁃distance dependence between different local structures by means of the external attention,rich global context information can be obtained.In order to further enhance the context⁃aware ability of the model,a neighboring scale feature fusion module based on backward competitive attention is designed to learn the similarity between low⁃level and high⁃level semantic features,and re⁃assign the weights for the intermediate feature channels.The proposed method is evaluated on the S3DIS large⁃scale indoor point cloud dataset,and the results show that the proposed method achieves an average cross⁃merge ratio of 65.2%on Area5,which is 2.5%higher than RandLA⁃Net,and an mean intersection over union(mIoU)of 71.4%on 6⁃fold cross⁃validation,which is 1.4%higher than RandLA⁃Net.It is demonstrated that the proposed method can effectively extract global contextual features,and improve the accuracy of semantic segmentation.

关 键 词：点云语义分割 全局上下文特征 邻近尺度 外部注意力 后向竞争性注意力 平均交并比 

分 类 号：TN911.7-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程] TP183[自动化与计算机技术—计算机应用技术]