Abstract
Latent structures exist in sequences ranging from biosequence, music, human motion to various real-world data. Such hidden information provides a more flexible way to model and use the original input data. In the thesis, we present a method to discover the latent structure and then to infer the hidden knowledge of the given skeletal animation (i.e. motion capture data). The proposed approach not only captures the local characteristics (i.e. the variation between two consecutive frames) of a given motion sequence, but also discovers its global structure (i.e. the sub-actions of an action). It is difficult and time-consuming to mine the high-dimensional data directly. Therefore, given a motion sequence, we first convert it into motion units called symbols through a cluster-based symbolization process. Consequently, the input sequence could be regarded as a motion string. Then our method discovers the latent structure based on an analysis of the repeating patterns of the motion string. Once the latent structure has been discovered, a grammar could be generated for obtaining a compact and hierarchical representation of the given sequence. Using the grammar that is produced by our method, we can provide both animation skimming and summary of the given animation. Therefore, an animation can be browsed by its highlights or summary sequence quickly; an animation can be illustrated with few selected keyframes; and we can understand and describe an animation via its meaningful sub-sequences.