Unsupervised motion capture data segmentation based on topic model

Abstract

AbstractIn this paper, we propose an unsupervised motion segmentation method based on topical model borrowed from Natural Language Processing. We apply hierarchical clustering on motion dataset to obtain a list of representative poses to constitute motion 'vocabulary'. By doing so, motion capture data can be viewed as text which comprises a sequence of motion words. We use sliding window to generate a sequence of motion documents (with overlap between consecutive motion documents). Then we use Sparse Topical Coding (STC) model to extract sparse topical codes of motion documents and conduct spectral clustering to get motion segmentations. Silhouette coefficient is used to determine the value of K (number of motion types). The results of experiments show that our method can segment motions with a very high accuracy. Our method has a strong generalization ability that also performs well on motion data which is captured by different subjects, with various motion types, even though they are from different motion dataset (HDM05 in our experiment).