Clustering in pursuit of temporal correlation for human motion segmentation.

Abstract

Temporal correlation is an important property of the video sequence. However, most methods only accomplish the clustering of frames via the measurement of similarity between frame pair, and the temporal correlation among frames is rarely taken into account. In this paper, a method for clustering in pursuit of temporal correlation is proposed to address human motion segmentation problem. Aiming at the video sequence, a one-hot indicator vector is extracted from a frame as a frame-level feature. The description of the relationship between the features is formulated as a minimization problem with respect to a similarity graph. A temporal constraint in the form of a trace is imposed on the similarity graph to capture the temporal correlation. On the premise of the non-negative similarity graph, an optimal solution to the graph augments the relationship between the selected features and their adjacent features, while suppressing its relevance to the features that are far away from it in terms of the time span. Normalized cut is implemented on the graph so as to give clustering results. The experiments on human motion segmentation demonstrate the superior performance of the proposed method in tackling the motion data.