Automated human motion segmentation via motion regularities

摘要: Analysis and reuse of human motion capture (mocap) data play an important role in animation, games medical rehabilitation. In various mocap-based animation techniques, segmentation is regarded as one the fundamental functions. Many proposed methods utilize little or no prior knowledge. However, has its own regularities, so reasonable assumptions on these regularities will lead to better performance. this paper, we focus learning intrinsic mocap based a small set training which only contain daily-life motions. By utilizing learnt can successfully segment long sequences containing types that not even include data. First, by assuming most motions be composed number typical poses, vocabulary (mo-vocabulary) obtained using key pose extraction clustering analysis, are low-level regularity. replacing each frame with similar mo-vocabulary, transformed into text-like documents. Second, use latent Dirichlet allocation patterns combinations frequently occur motions, namely topics (mo-topics), high-level regularities. representing target distribution over mo-topics, task naturally turned problem detecting notable changes distribution. Finally, propose local semantic coherence curve sequences. Since mo-topics semantically meaningful significantly increase abstraction-level representation, logically correct results obtained. The experiments demonstrate approach outperforms available CMU Bonn database.