融合注意力门控机制的大场景点云语义分割  

The semantic segmentation algorithm for large scene point cloud based on attention gating mechanism

作　　者：王蕾[1,2] 朱芬芬 李金萍[1] 刘华 WANG Lei;ZHU Fen-fen;LI Jin-Ping;LIU Hua(School of Information Engineering,East China University of Technology,Nanchang 330013,China;Jiangxi EngineeringLaboratory on Radioactive Geoscience and Big Data Technology,East China University of Technology,Nanchang 330013,China;School of Surveying and Mapping 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]江西省放射性地学大数据技术工程实验室,东华理工大学,江西南昌330013 [3]东华理工大学测绘工程学院,江西南昌330013

出　　处：《激光与红外》2023年第11期1785-1792,共8页Laser & Infrared

基　　金：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(No.JELRGBDT202202);江西省放射性地学大数据技术工程实验室开放基金项目(No.JELRGBDT202103);江西省自然科学基金项目(No.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项目(No.DLLJ202004);国家自然科学基金项目(No.42001411)资助。

摘　　要：室外大场景激光点云语义分割已成为3D场景理解、环境感知的关键性技术,在自动驾驶、智能机器人和增强现实(AR)等领域应用广泛。然而大场景的激光点云具有多目标、几何结构复杂,不同地物尺度变化大等特点,使得在稀疏的小目标点云(例如行人、摩托车等)上的分割性能较低。针对上述问题,本文提出一种融合注意力门控机制的室外点云语义分割算法,设计由注意力机制和多尺度上下文特征融合组成的注意力门控单元,提高对激光点云细粒度特征的表达,降低随机降采样过程中点云几何结构特征丢失程度,从而增强了网络对弱小目标的特征获取能力;同时设计基于共享MLP的平均池化单元,进一步简化自注意力局部特征聚合模块,有效地加速网络收敛,能高效地实现大场景点云的语义分割。本文方法在自动驾驶场景室外激光点云数据集SemanticKITTI上的实验表明,与文献RandLA-Net相比,收敛速度提升48.3%,平均交并比(mIoU)由53.9%提升至54.5%,提高0.6%,尤其是在小目标上交并比(IoU)均有明显提高,person类和motorcycle类的交并比分别提高0.8%和5.4%。Semantic segmentation for outdoor large-scale point cloud has become a key technology for 3D scene understanding and environmental awareness and is widely used in fields such as autonomic driving,intelligent robotic and augmented reality(AR).However,laser point clouds of large scenes are characterized by multi-targets,complex geometrical structures,and large variations in the scales of different features,making the segmentation performance on sparse point clouds of small targets(e.g.,pedestrians,motorcycles,etc.)low.To address the above problems,an outdoor point cloud semantic segmentation algorithm incorporating an attentive gating mechanism is proposed in this paper.An attentive Gating Unit based on attention mechanism and multi-scale feature fusion method is designed to improve the expression of fine-grained features of laser point clouds and significantly reduce the information loss during the random downsampling process,thus enhancing the feature extraction performance for weak targets.At the same time,anaverage pooling unit based on shared MLP is designed to further simplify the self-attention local feature aggregation module,which effectively accelerates the network convergence speed and can efficiently realize the semantic segmentation of point clouds in large scenes.The experiments on outdoor driving dataset semanticKITTI show that the convergence speed is increased by 48.3%,and the mean intersection-over-Union(mIoU)of all classes is improvedfrom 53.9%to 54.5%,an increase of 0.6%,compared with the literature RandLA-Net.Especially,the Intersection-over-Union(IoU)of small-scale class is significantlyimproved,for example,the IoU score of person and motorcycle are increased by 0.8%and 5.4%,respectively.

关 键 词：大场景激光点云 语义分割 随机降采样 平均池化单元 注意力门控单元 多尺度特征融合 

分 类 号：TN249[电子电信—物理电子学] TP391.41[自动化与计算机技术—计算机应用技术]