基于全局注意力机制的Robust-PointPillars三维目标检测  

Robust-PointPillars 3-D Target Detection Based on Global Attention Machanism

作　　者：王盈丰 吴俭 宋佳 柯涛 付伟 WANG Ying-feng;WU Jian;SONG Jia;KE Tao;FU Wei(The 723 Institute of CSSC,Yangzhou 225101,China)

机构地区：[1]中国船舶集团有限公司第七二三研究所,江苏扬州225101

出　　处：《舰船电子对抗》2024年第2期86-92,共7页Shipboard Electronic Countermeasure

摘　　要：提出了一种基于全局注意力机制的Robust-PointPillars三维目标检测方法,在智能驾驶的应用中,提高了目标检测的精度和鲁棒性。PointPillars等神经网络通过使用点云柱表示点云,具有实现三维目标检测的潜力。首先介绍了空间和通道双重注意力模块,以增强有学习价值的点云特征,解决了PointPillars缺乏点云柱内部学习机制和特征提取不足的问题;挤压与激励网络(SENet)模块的引入,使PointPillars对特征信息的学习理解能力得到进一步提高;最终,对受到干扰或缺失的传感器信号进行抑制,并利用全局注意力算法来提高鲁棒性。基于KITTI数据集上的目标检测结果,本文算法具有良好的目标检测精度和鲁棒性。A Robust-PointPillars 3-D target detection method based on global attention machanism is proposed,which improves the accuracy and robustness of target detection in the application of intelligent driving.PointPillars and other neural networks have the potential to enable 3-D target detection by using pillars to represent the point cloud.This paper firstly introduces the dual attention module of space and channel for enhancing the feature of point cloud with learning value,which solves the lack of learning mechanism and feature extraction for Pointpillars.Then squeeze-and-excitation networks(SENet)module is introduced to further improve the learning and understanding ability of PointPillars to feature information.Finally,the disturbed or missing sensor signals are restrained and the global attention algorithm is used to improve the robustness.According to the target detection results based on KITTI data set,the algorithm proposed in this paper has excellent target detection accuracy and robustness.

关 键 词：三维目标检测 PointPillars 全局注意力机制 挤压与激励网络模块 

分 类 号：U462[机械工程—车辆工程] TN957.51[交通运输工程—载运工具运用工程]