激光点云中三维多目标在线跟踪方法  

3D multi-object online tracking method in laser point cloud

作　　者：陈伯云 陆安江 周骅 赵麒 黄际玮 CHEN Boyun;LU Anjiang;ZHOU Hua;ZHAO Qi;HUANG Jiwei(College of Big Data and Information Engineering,Guizhou University,Guizhou 550025,Guiyang,China;College of Mechanical Electronic and Engineering,Guizhou Mingzu University,Guizhou 550025,Guiyang,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵阳550025 [2]贵州民族大学机械电子学院,贵阳550025

出　　处：《激光杂志》2023年第4期68-73,共6页Laser Journal

基　　金：国家自然科学基金资助项目(No.62065002,61865002)。

摘　　要：为降低三维多目标跟踪系统计算复杂度,实现准确高效的多目标在线跟踪,提出一种激光点云中三维多目标在线跟踪方法。在目标检测阶段,通过3D深度相机获取五帧点云序列进行三维背景建模,将实时获取的序列与背景模型进行差分运算检测运动目标。在数据关联阶段,结合运动矢量一致性和最小欧式距离建立预测与检测之间的权值匹配矩阵,使用带权匈牙利算法完成匹配。实验结果表明,对于百万级点数的点云图本算法完成目标跟踪耗时每帧低于0.1 s,跟踪准确度高于95%,跟踪精度低于0.5 mm。能够实现毫米级工业小目标的在线跟踪,兼顾了精确性和实时性。In order to reduce the computational complexity of the three-dimensional multi-object tracking system and achieve accurate and efficient multi-object online tracking,a three-dimensional multi-target online tracking method in the laser point cloud is proposed.In the object detection phase,five-frames of point cloud sequences is obtained by the 3D depth camera for 3D background modeling,and the sequences obtained in real time are differentially calculated to detect moving targets.In the data association stage,the weight matching matrix between prediction and detection is established by combining the motion vector consistency and the minimum Euclidean distance,and the weighted Hungarian algorithm is used to complete the matching.The experimental results show that for the point cloud sequence of millions of points,the target tracking time of this algorithm is less than 0.1 s per frame,the tracking accuracy is higher than 95%,and the tracking accuracy is less than 0.5 mm.It can realize the online tracking of millimeter industrial small targets,taking into both accuracy and real-time.

关 键 词：多目标跟踪 三维背景差分法 深度相机 目标检测 数据关联 

分 类 号：TN249[电子电信—物理电子学]