改进PointPillars的三维点云检测研究  

Research on Improving PointPillars 3D Point Cloud Detection

作　　者：阜远远 王嘉鑫 王建平 张太盛 周晋伟 方祥建 FU Yuanyuan;WANG Jiaxin;WANG Jianping;ZHANG Taisheng;ZHOU Jinwei;FANG Xiangjian(School of Mechanical and Automotive Engineering,Anhui Polytechnic University,Wuhu 241000,China;Engineering Department,CRRC Puzhen Alstom Transportation Systems Co.,Ltd.,Wuhu 241000,China)

机构地区：[1]安徽工程大学机械与汽车工程学院,安徽芜湖241000 [2]中车浦镇阿尔斯通运输系统有限公司工程部,安徽芜湖241000

出　　处：《安徽工程大学学报》2025年第1期31-37,共7页Journal of Anhui Polytechnic University

基　　金：芜湖市科技计划重点研发项目(2020yf53);安徽省科技重大专项(202103a05020033)。

摘　　要：针对目前三维点云存在漏检误检等问题,提出了一种基于改进PointPillars的车辆检测方法。在原有模型的基础上,引入注意力机制,改进激活函数,重新训练KITTI数据集。消融实验结果表明,改进算法模型比原算法模型的Map值提升了2%,不同帧点云检测的可视化结果显示,改进PointPillars模型增强了点云检测能力。改进模型可以在未来应用于智能车或机器人环境感知领域。In response to the current issues of missed and false detections in 3D point clouds,this paper proposes a vehicle detection method based on improved PointPillars.On the basis of the original model,attention mechanism is introduced,activation function is improved,and KITTI dataset is retrained.The results of the ablation experiment showed that the improved algorithm model increased the Map value by 2% compared to the original algorithm model.The visualization results of point cloud detection at different frames showed that the improved PointPillars model enhanced the ability of point cloud detection.The improved model can be applied in the field of intelligent vehicle or robot environment perception in the future.

关 键 词：PointPillars 点云检测 BEV伪图像 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]