Finding repetitive patterns in 3D human motion captured data

Abstract

Finding repetitive patterns is important to many applications such as bioinformatics, finance and speech processing, etc. Repetitive patterns can be either cyclic or acyclic such that the patterns are continuous and distributed respectively. In this paper, we are going to find repetitive patterns in a given motion signal without prior knowledge about the type of motion. It is relatively easier to find repetitive patterns in discrete signal that contains a limited number of states by dynamic programming. However, it is impractical to identify exactly matched states in a continuous signal such as captured human motion data. A point cloud similarity of the input motion signal itself is considered and the longest similar patterns are located by tracing and extending matched posture pairs. Through pattern alignment and autoclustering, cyclic and acyclic patterns are identified. Experiment results show that our approach can locate repetitive movements with small error rates. © 2008 ACM.