运动秩1分解及其在运动检索中的应用

Motion Rank One Decomposition and its Application on Motion Retrieval

低秩分解可以有效地应用于运动检索中,然而目前有些方法是针对每个运动单独分解,在分解算法层次上忽略了不同运动之间的相关性.为此,提出一种在数据库上的低秩分解算法,在数据库中所有运动共享一组基,并加入稀疏约束得到运动数据的有效表示;提出一种合理的运动数据构成方式,得到优化目标方程,并给出相应的优化解法,证明了其收敛性.采用文中的分解算法,每个运动被低秩表示成一个基和一个时序向量,由于不同的运动共享一组基,因此该算法具有更好的聚类效果,即相似运动倾向于选择相同的基.实验结果表明,文中算法在运动检索应用上是有效的,并讨论了不同参数设置对检索结果的影响.

Recently, low rank decomposition has successfully been applied to human motion retrieval. However, the existing method works on the single motion sequence. Therefore, it ignores the motion correlation in the algorithm level. We propose a low rank decomposition method which could work on motion dataset and all motions share the same set of basis, so our method has clustering effect because similar motion tends to select the same basis. Furthermore, we add the sparse constraint and obtain the effective representation for motion data. In order to achieve this, we present a reasonable construction method for motion data and derive the objective function, based on which, we propose our optimal decomposition algorithm and demonstrate its convergence. We compare our method with other different human motion retrieval approaches and discuss how different parameters of our algorithm affect the results.