面向嵌入式平台的快速稳健地面点云分割方法  

A fast and robust ground point cloud segmentation method for embedded platforms

作　　者：刘珺玮 刘龙 苏云泉 谢光达 耿艳东 苗蕾 邵天翔 LIU Junwei;LIU Long;SU Yunquan;XIE Guangda;GENG Yandong;MIAO Lei;SHAO Tianxiang(Scientific Research Institution of Inner Mongolia First Machinery Group Co.,Ltd.,Baotou 014030,China)

机构地区：[1]内蒙古第一机械集团股份有限公司科研所,内蒙古包头014030

出　　处：《兵器装备工程学报》2024年第S2期217-222,共6页Journal of Ordnance Equipment Engineering

摘　　要：提出一种基于区域生长及自适应阈值的快速稳健地面点云分割方法。在基于栅格投影的地面分割方法基础上,根据地面点云特征对原始点云进行预处理,为后续算法提供良好初始状态;以栅格为单位进行区域生长,根据生长结果对需处理栅格进行点筛选,避免遍历所有点造成计算量增大;采用筛选阈值自适应方法,提高算法在城市及越野等不同地面场景的鲁棒性,实现地面分割。分别采用KITTI数据集和自采数据集对算法的城市场景地面和越野场景地面分割性能进行实验。实验结果表明,所提方法具备场景自适应能力,城市及越野场景下平均灵敏度指标达到95.663%,平均单帧处理耗时小于50 ms,满足嵌入式平台地面分割任务准确性和实时性要求,适用嵌入式平台的地面分割任务。To solve the problems of poor real-time performance and low accuracy of LiDAR point cloud ground segmentation algorithm on embedded platforms,a fast and robust ground point cloud segmentation method based on region growing and adaptive threshold is proposed.Based on the grid projection based ground segmentation method,the original point cloud is preprocessed according to its features,providing a good initial state for subsequent algorithms;Introducing the concept of region growth,region growth is carried out on a grid by grid basis,and point selection is performed on the grid to be processed based on the growth results,avoiding the problem of increased computational complexity caused by traversing all points;Adopting a threshold adaptive filtering method to improve the robustness performance of the algorithm in different ground scenarios such as urban and off-road environments,thereby achieving ground segmentation.Experimental validation was conducted on the ground segmentation performance of the algorithm in urban and off-road scenarios using both KITTI dataset and self collected dataset,and compared with existing algorithms.The experimental results show that the proposed method has scene adaptation capability,with an average sensitivity index of 95.663%in urban and off-road scenarios,and an average single frame processing time of less than 50 ms.It meets the accuracy and real-time requirements of embedded platform ground segmentation tasks and is suitable for ground segmentation tasks based on embedded platforms.

关 键 词：嵌入式平台 激光雷达 地面分割 区域生长 自适应阈值 实时性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]