基于聚类的无监督动作分割方法主要利用序列中相邻帧之间的结构相似性来提高动作分割的准确性.这在实现动作片段内部一致划分的同时给不同动作边界的准确分割带来隐患.为此提出了一种基于多图融合约束矩阵分解的动作分割方法.通过融合序列中的结构相似性和度量相似性信息构造多图融合约束项,融入到半非负矩阵分解中获得序列的低维表示,进而获得序列的k近邻图并利用图割的方法实现准确分割.在两类动作序列上的实验表明,所提方法在保持动作内部一致划分的同时能够准确划分动作边界,明显提升了分割准确性,时间效率也明显提升.
Most clustering-based action segmentation methods mainly exploit the structure similarity information between adjacent frames(points)in the sequence to improve the accuracy of action segmentation.These methods im-prove the consistency of segmentation inside each action but introduce potential issues for accurately segmenting action boundaries.Hence,this paper presents a novel action segmentation method based on multigraph fusion constraint semi-nonnegative matrix factorization(MGSeNMF).In this method,the structural and measurement similarity information is fused to build a multigraph fusion constraint term,which is fused to semi-NMF to obtain a low-dimensional representa-tion.A k-nearest neighbor graph is also generated for the action sequences,realizing accurate segmentation using the graph cut method.Experimental results on two kinds of real-action datasets show that MGSeNMF can accurately divide the boundary of actions while maintaining consistent segmentation inside each action.Thus,the proposed method im-proves the accuracy of segmentation and efficiency of running time significantly.