泛函回归代理及条件期望配准的机械摆动测量  

Mechanical swing measurement based on functional regression surrogate and conditional expectation registration

作　　者：郑思凡 陈平平[1,3] 苏凯雄[1,3] 吴永春 ZHENG Si-fan;CHEN Ping-ping;SU Kai-xiong;WU Yong-chun(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350116,China;Intelligent Manufacturing Institute,LiMing University,Quanzhou 362000,China;Digital TV Intelligent Technology National and Local Joint Engineering Research Center,Fuzhou 350116,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350116 [2]泉州黎明职业大学智能制造工程学院,福建泉州362000 [3]数字电视智能化技术国家地方联合工程研究中心,福建福州350116

出　　处：《光学精密工程》2021年第5期1154-1168,共15页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.U1505251,No.61871132);福建省高职院校智能制造协同创新中心项目资助(No.2016071);福建省政府招标采购交通执法信息系统运维服务项目资助(No.3500FRGK2017008-1);服务项目资助(No.3500FRGK2017008-1)。

摘　　要：在机械摆动的运动分割及视觉测量中,针对传统以块状轨迹群为单位的谱聚类运动分割因摆杆光流轨迹中断及线速度分布差异所导致的碎片化与过分割的局限性,提出一种以曲率为相似度度量的弧状轨迹群为单位的谱聚类分割算法,并结合点云配准完成转速图像测量。算法先用活动子集的稀疏高斯回归学习出弧状轨迹群的平均轨迹,将此平均轨迹作为稀疏子空间聚类的种子样本一次性完成运动分割,最后将非种子样本重新归入其被代理的种子样本聚类中以获得每帧最大稠密度的分割点云。在各帧点云基础上,通过条件期望点云配准算法求取帧间点云变换矩阵,并提取转动分量完成摆杆摆角测量。为证明有效性,结合客运车辆日次安全检测视觉自动化系统项目,以6种不同照度下5种车型的双摇杆刮水器总成为对象,比较了三种算法对摆角的测量精度。结果表明:本算法能完整学习出等长轨迹,且与人为标定角位移回归值误差均方值小于10%,同时运算量小于传统的交替方向乘子法(ADMM)单次迭代,可作为工业智能制造与自动控制系统中的机械视觉运动测量及机械视觉故障诊断方面应用。The process of motion segmentation and measurement of mechanical swing based on traditional block shape optical flow trajectory group clustering exhibits limitations in terms of over-segmentation and fragmentation due to the partial occlusion,interruption,and uneven velocity distribution of the optical flow trajectory.To overcome these limitations,we herein propose an arc-shaped trajectory clustering algorithm that uses curvature as a similarity metric and combines it with point cloud registration to perform mechanical swing measurement.The algorithm first performs sparse Gaussian regression of the active subset to learn the average trajectory of the arc-shaped trajectory group.Subsequently,the average trajectory is used as the seed sample of the sparse subspace clustering to complete the motion segmentation at one time.Finally,the non-seed sample is reclassified into its surrogate seed sample cluster to obtain the point set of each frame.Through conditional expectation point cloud registration,the rotation component is extracted to complete the swing angle measurement.The proposed algorithm is used for a vehicle windshield wiper under the four-link wiper assembly model and six different environment illuminances,as part of a visual automation system project targeting the daily safety inspection of passenger station vehicles,and compared with other algorithms.The experimental results show that the proposed algorithm can fully learn the blocked trajectory,and the mean square error with an artificially calibrated value is less than 10%.Furthermore,the computational complexity is only equivalent to that in the case of a single iteration of the alternating direction method of multipliers(ADMMs),Therefore,the proposed algorithm can be used for mechanical vision motion measurement in industrial intelligent manufacturing and automatic control systems.

关 键 词：变分光流 子空间聚类 稀疏高斯回归 RKHS空间 视觉里程计 Grashof双摇杆 李群流形 最大条件期望 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TH691.9[自动化与计算机技术—控制科学与工程]