Motion Capture Data Segmentation Based on Spectral Clustering

A long motion capture data often contains several different motions or the same motion repeats several times. It is an important topic to segment a long motion capture data into different motions. In this paper, we propose a method to segment long motions into several different motions by using a spectral clustering algorithm. When computing similarity matrices, we first cut the original motion capture data into mocap clips and each clip contains k frames mocap data. We then apply PCA dimension reduction technique on each mocap clips and compute similarities between these clips. By doing so, our similarity measurement takes the motion continuity into account. Moreover, we also greatly improve the efficiency by avoiding frame by frame similarity computation which is much more time consuming. When applying spectral clustering on the similarity matrix directly, the resulting classification labels are with serious noises. To address this problem, we propose to use median filter to remove the noises, and get good segmentation points. The automatic segmentation results on 14 motion data demonstrate the effectiveness of the proposed method.