Retrieval of motion capture data based on short-term feature extraction

This poster presents a motion retrieval algorithm, which searches the motions in the same category as a query's (known as logically similar motions) in a motion capture database. The challenge is that logically similar motions may not be numerically similar due to the motion variations [Müller et al. 2005]. In this poster, we propose a novel short-term feature that extracts both symbolized representation and continuous features from joint velocities in a motion clip, which is employed to effectively retrieve logically similar motions to the query. Although symbolized representation of human motion has been studied [Müller et al. 2005], our approach is different in that we consider temporal correlation instead of Müller's spatial relationship. Moreover, not only symbolized representation (dynamic pattern) but also continuous features (average speed) are extracted in our short-term feature. Furthermore, our method is more friendly to novices as it requires no prior knowledge to determine features. Our experiments demonstrate that our algorithm greatly improves the performance compared to two conventional methods.