基于光流拓扑稀疏加权的同步运动去混叠

Synchronous motion de-aliasing based on optical flow topological sparse weighting

针对欧式距离加权的稀疏子空间聚类在对多个同步运动刚体进行运动分割时不考虑刚体流形结构的局限性,提出了一种由光流轨迹流形拓扑结构加权的稀疏子空间聚类算法,在光流轨迹的时空相似度邻接矩阵里计算各轨迹相似度的流形距离并嵌入稀疏子空间字典表达的权值矩阵进行稀疏系数求解,使得流形距离较近的轨迹优先成为稀疏自表达字典,从而减少对欧式空间相距较小但不属于同一物体的同步运动刚体轨迹的聚类混叠,经过同步移动和同步摆动两种情况算法对比实验表明:本文提出算法可以将混叠降低到1%以下,最后,在双针床经编机贾卡针同步摆动的运动分割结果表明算法具有进一步的工业视觉应用前景。

This work aims to overcome the limitations of Euclidean distance weighted sparse subspace clustering that does not consider the manifold structure of a rigid body.Here,a sparse subspace clustering method weighted using optical flow trajectory manifold topology was proposed.In the proposed algorithm,the manifold distance of each trajectory in the space-time similarity adjacency matrix was embedded into the weight matrix to solve the sparse coefficient.This ensured that the trajectory with a relatively closed manifold distance became the sparse self-expression dictionary,thereby reducing the clustering aliasing error of synchronous motion.The comparison of experiments between synchronous motion and synchronous swing reveals that the proposed algorithm can reduce aliasing error down to1%.Finally,the motion segmentation results of Jacquard needle indicate that the algorithm can be potentially used for industrial applications.