在基于机器视觉检测的客运车辆日趟故障安全例检自动化设计中，针对子空间聚类算法对汽车刮刷总成摆杆摆幅检测中因玻璃复杂背景导致光流轨迹过于稀疏的缺陷，提出了一种双粒度光流流形学习的汽车刮水器总成主副摆杆运动分割算法。首先将摆杆满幅等长 LDOF 变分光流轨迹作为粗粒度光流进行稀疏子空间聚类，获得可靠的种子样本; 然后通过构建稠密细粒度光流与粗粒度光流的轨迹时空相似度流形拓扑图，并在图上利用调和函数将种子轨迹样本邻接节点标签凸松弛为高斯随机场进行半监督标签扩散，从而获得稠密的雨刮运动区域，以便进一步做 RANSAC 直线拟合和摆角计算;最后，将该算法模块经过 ocx 插件封装后以回调函数体的形式嵌入客运站的车辆跟踪模块进行同步，并在客运站现场采集了 6 种不同照度下 4 种车型共 153 车次的进站安检视频，用于分析比较同步后的两种粒度流形学习算法对摆杆的运动分割后直线拟合误差与摆角误差。实验结果表明: 本算法对运动摆杆的拟合与其摆角计算的精确率均可以达到85%以上，具有进一步推广应用前景。

In the vehicle security inspection system design based on machine vision，a subspace clustering algo-rithm via two granularity optical flow manifold learning was proposed to detect automobile wiper swing angle ampli-tude in order to avoid the defect that the optical flow trajectory is too sparse caused by the complex background of glass． Firstly，the full-length wiper LDOF variational optical flow trajectory was used as coarse-granularity optical flow to perform sparse subspace clustering to obtain reliable seed sample． Then the temporal-spatial similarity to-pography graph of dense and fine granularity optical flow and coarse-granularity optical flow was constructed，and the harmonic function was used to relax the neighbor trajectory label of the seed's to the Gaussian random field for semi-supervised label diffusion to obtain a dense wiper motion region for further RANSAC line fitting and swing an-gle calculation． Finaly，the algorithm module was encapsulated by ocx plugin and embeded into the vehicle track-ing module in the form of callback function for synchronization． Security check video of 153 trains of 4 types of ve-
hicles under 6 different illuminances was collected． The videos was used to analyze and compare the accuracy of fitting and swing angle of two granularity optical flow manifold learning algorithms． The experiment results show that the accuracy of fitting and swing angle of the new algorithm can reach more than 85%，showing a broad appli-cation and promotion prospect．