基于图像检测的三维激光点云聚类方法研究与应用  

Research and Application on 3D Laser Point Cloud Clustering Method Based on Image Detection

作　　者：林乐彬 周军[1,2,3] 皇攀凌[1,2,3] 李留昭 欧金顺 LIN Lebin;ZHOU Jun;HUANG Panling;LI Liuzhao;OU Jinshun(School of Mechanical Engineering,Shandong University,Jinan 250061,China;Key Laboratory of High Efficiency and Clean Mechanical Manufacture,Ministry of Education,Shandong University,Jinan 250061,China;National Experimental Teaching Demonstration Center of Mechanical Engineering,Shandong University,Jinan 250061,China)

机构地区：[1]山东大学机械工程学院,山东济南250061 [2]山东大学高效洁净机械制造教育部重点实验室,山东济南250061 [3]山东大学机械工程国家级实验教学示范中心,山东济南250061

出　　处：《控制工程》2024年第8期1478-1484,共7页Control Engineering of China

基　　金：山东省重点研发计划项目(2019JZZY010453)。

摘　　要：针对移动机器人对周围环境障碍物感知的需求,特别是针对激光雷达点云数据的的语义检测和尺寸识别问题,提出基于图像处理和点云数据处理相融合的方法。首先,对3D激光雷达和双目相机进行外参标定;然后,使用YOLOv4深度神经网络将检测出的二维图像实例通过双目相机恢复像素深度,将其检测目标定位到激光点云中;在对点云进行滤波、几何约束后,采取基于KD树的搜索方法对点云欧式聚类分割,最终将识别出的语义信息输出到点云聚类结果中。实验结果表明,设计的方法可以准确、快速地识别并分割聚类出点云,可以应用于移动机器人导航避障。To address the need for mobile robots to perceive obstacles in surrounding environment,especially the semantic detection and size recognition of LiDAR point cloud data,a method that integrates image processing and point cloud data processing is proposed.Firstly,the extrinsic parameters of the 3D LiDAR and stereo camera are calibrated.Then,the YOLOv4 deep neural network is used to detect 2D image instances and recover pixel depth through the stereo camera,mapping the detected targets to the LiDAR point cloud.After filtering and applying geometric constraints to the point cloud,a KD-tree-based search method is used for Euclidean clustering segmentation of the point cloud.Finally,the identified semantic information is output to the point cloud clustering results.Experimental results show that the designed method can accurately and quickly identify and segment point cloud clusters,making it applicable to mobile robot navigation and obstacle avoidance.

关 键 词：点云聚类 目标检测 双目视觉定位 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]