基于视觉修正的激光雷达体积测量方法  

Volumetric measurement of Lidar based on visual correction

作　　者：郭隆强 何赟泽 杜旭 郭昱良 付玉轩 Guo Longqiang;He Yunze;Du Xu;Guo Yuliang;Fu Yuxuan(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;Shenzhen Research Institute,Hunan University,Shenzhen 518000,China;School of Automation,Central South University,Changsha 410083,China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082 [2]湖南大学深圳研究院,深圳518000 [3]中南大学自动化学院,长沙410083

出　　处：《仪器仪表学报》2023年第10期48-59,共12页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(62101184);湖南省科技创新领军人才(2023RC1039);湖南省自然科学基金重大项目(2021JC0004);广东省基础与应用基础研究基金海上风电联合基金(2022A1515240050);湖南省重点研发计划(2022GK2012)项目资助。

摘　　要：针对激光雷达点云数据稀疏、扰动、存在噪声和其他方法难以迁移,实时性差等难题,面向“L”型小尺寸目标研究了一种基于视觉修正的激光雷达体积测量方法。该方法首先通过联合标定和时间戳最近邻匹配实现相机与激光雷达数据的对齐;然后经过目标检测算法获取图像中目标的信息,与此同时对点云数据执行地面分割得到地面点云与非地面点云,利用视觉投影和点云聚类实现目标点云的分割,使用KDtree找到目标点云附近的地面点云;最后,设计了一种三维框的拟合算法初步完成点云目标三维框的粗拟合,并建立视觉修正模型对于目标三维框进行细修正,从而实现目标体积的计算。实验结果表明,对于武器箱道具、医疗箱和油桶等“L”型物体,提出的算法在一定范围内,体积测量的平均相对误差小于4.44%、最大误差小于6.12%、最大重复性小于5.61%,并且基于视觉的修正模型大幅提高了算法的精度和稳定性,在嵌入式平台的处理1帧用时55 ms,能够实现实时高精度的体积测量,具有良好的工程应用前景。To address challenges faced by most volume measurement methods,such as difficulties in transferring,poor real-time performance,and sparse,noisy,and perturbed Lidar point cloud data,this article presents a vision-based correction method for Lidar volume measurement for small‘L’-shaped objects.The proposed method first aligns camera and Lidar data through joint calibration and time-stamped nearest-neighbor matching.Subsequently,it leverages target detection algorithms to extract information from images while simultaneously performing ground segmentation on point cloud data to distinguish ground and non-ground points.By employing visual projection and point cloud clustering,the method segments target point clouds and utilizes KDtree to identify ground points in proximity to the target point cloud.Finally,a 3D box fitting algorithm is proposed to provide initial rough estimation of the point cloud target′s 3D box.A visual correction model is established to refine the target′s 3D box,and enable accurate volume calculation.Experimental results show that for‘L’-shaped objects like weapon crates,medical boxes,and barrels,the proposed algorithm achieves promising results within a certain range.The average relative error in volume measurement is less than 4.44%,with a maximum error below 6.12%and a maximum repeatability error of 5.61%.In addition,the integration of the visual correction model significantly enhances the algorithm′s accuracy and stability.The processing of frame on an embedded platform takes 55 ms,demonstrating the capability to achieve realtime,high-precision volume measurement.This method holds great promise for practical engineering applications.

关 键 词：体积测量 点云聚类 目标检测 视觉修正 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置] TH702[自动化与计算机技术—控制科学与工程]